diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_13.log b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_13.log new file mode 100644 index 0000000000000000000000000000000000000000..770be23736fed064d7e41e27864160c54cd53133 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_13.log @@ -0,0 +1,1994 @@ +2026-06-19 01:20:16.931 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-19 01:20:16.932 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-19 01:30:17.295 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 01:30:17.430 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:30:17.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:31:52.378 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:31:52.380 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:33:45.515 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 01:34:00.994 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 01:35:56.997 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.94s (1.93min) +2026-06-19 01:35:57.686 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 01:37:54.189 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 01:37:57.219 | INFO | hyvideo.utils.grpo_utils:train_one_step:3127 - Rank 13, sp_rank 0: init adaptive KL controller (init=1.000e-07, min=1.000e-07, max=1.000e-05) +2026-06-19 01:45:11.478 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 1, train_steps: 40 +2026-06-19 01:45:12.575 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 01:45:12.715 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:45:12.755 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:46:17.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:46:17.248 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:48:10.945 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 01:48:11.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 01:50:07.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.70s (1.93min) +2026-06-19 01:50:08.465 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 01:52:06.953 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 01:59:06.759 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 2, train_steps: 80 +2026-06-19 01:59:07.381 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 01:59:07.513 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:59:07.549 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:00:11.725 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:00:11.728 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:02:10.528 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 02:02:11.592 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:04:06.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.88s (1.91min) +2026-06-19 02:04:07.374 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:06:03.753 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:13:04.855 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 3, train_steps: 120 +2026-06-19 02:13:05.465 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 02:13:05.612 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:13:05.658 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:14:09.946 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:14:09.946 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:16:04.655 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 02:16:05.698 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:17:59.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.77s (1.90min) +2026-06-19 02:18:00.447 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:19:59.348 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:26:59.880 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 4, train_steps: 160 +2026-06-19 02:27:00.547 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 02:27:00.744 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:27:00.858 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:28:05.379 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:28:05.394 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:30:00.316 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 02:30:01.141 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:31:55.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.62s (1.91min) +2026-06-19 02:31:56.651 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:33:53.223 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:40:53.220 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 5, train_steps: 200 +2026-06-19 02:40:53.901 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 02:40:54.081 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:40:54.165 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:41:58.612 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:41:58.615 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:43:51.812 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 02:43:52.635 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:45:46.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.50s (1.89min) +2026-06-19 02:45:46.993 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:47:43.351 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:54:40.655 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 6, train_steps: 240 +2026-06-19 02:54:41.360 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 02:54:41.507 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:54:41.554 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:55:46.303 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:55:46.303 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:57:42.949 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 02:57:43.877 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:59:37.354 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.38s (1.89min) +2026-06-19 02:59:38.042 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:01:37.731 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:08:37.482 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 7, train_steps: 280 +2026-06-19 03:08:38.147 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 03:08:38.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:08:38.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:09:42.751 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:09:42.755 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:11:38.337 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 03:11:39.361 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:13:34.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.68s (1.91min) +2026-06-19 03:13:34.977 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:15:32.041 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:22:31.149 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 8, train_steps: 320 +2026-06-19 03:22:31.816 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 03:22:32.026 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:22:32.162 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:23:37.311 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:23:37.314 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:25:33.100 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 03:25:34.036 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:27:28.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-19 03:27:29.431 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:29:26.904 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:36:25.644 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 9, train_steps: 360 +2026-06-19 03:36:26.346 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 03:36:26.440 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:36:26.488 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:37:31.434 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:37:31.435 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:39:26.010 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 03:39:26.930 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:41:20.400 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.36s (1.89min) +2026-06-19 03:41:21.105 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:43:17.562 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:50:16.932 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 10, train_steps: 400 +2026-06-19 03:50:53.111 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=10 (missing=0, unexpected=0) +2026-06-19 03:50:56.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 03:50:56.344 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:50:56.448 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:52:02.690 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:52:02.693 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:53:57.828 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 03:53:58.763 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:55:52.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.65s (1.89min) +2026-06-19 03:55:53.208 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:57:49.762 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:04:48.377 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 11, train_steps: 440 +2026-06-19 04:04:49.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 04:04:49.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:04:49.315 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:05:53.859 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:05:53.862 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:07:48.853 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 04:07:49.679 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:09:43.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.25s (1.89min) +2026-06-19 04:09:43.796 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:11:40.267 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:18:37.288 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 12, train_steps: 480 +2026-06-19 04:18:37.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 04:18:38.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:18:38.281 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:19:42.785 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:19:42.785 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:21:37.180 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 04:21:38.116 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:23:33.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.19s (1.92min) +2026-06-19 04:23:34.070 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:25:31.243 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:32:29.817 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 13, train_steps: 520 +2026-06-19 04:32:30.512 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 04:32:30.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:32:30.686 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:33:35.457 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:33:35.463 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:35:30.313 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 04:35:31.228 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:37:24.102 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.77s (1.88min) +2026-06-19 04:37:24.791 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:39:21.076 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:46:18.588 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 14, train_steps: 560 +2026-06-19 04:46:19.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 04:46:19.420 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:46:19.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:47:24.485 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:47:24.485 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:49:17.445 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 04:49:18.369 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:51:12.963 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.49s (1.91min) +2026-06-19 04:51:13.649 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:53:12.995 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:00:11.834 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 15, train_steps: 600 +2026-06-19 05:00:12.525 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 05:00:12.667 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:00:12.701 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:01:17.454 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:01:17.454 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:03:11.349 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 05:03:12.576 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:05:07.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.64s (1.91min) +2026-06-19 05:05:08.060 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:07:05.978 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:14:04.890 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 16, train_steps: 640 +2026-06-19 05:14:05.573 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 05:14:05.713 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:14:05.756 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:15:10.239 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:15:10.242 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:17:06.988 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 05:17:07.896 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:19:03.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.02s (1.92min) +2026-06-19 05:19:03.741 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:21:01.961 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:28:03.321 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 17, train_steps: 680 +2026-06-19 05:28:04.006 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 05:28:04.109 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:28:04.155 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:29:09.012 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:29:09.015 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:31:03.926 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 05:31:04.961 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:32:58.689 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.48s (1.89min) +2026-06-19 05:32:59.386 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:34:56.266 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:41:58.105 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 18, train_steps: 720 +2026-06-19 05:41:58.777 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 05:41:58.942 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:41:59.029 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:43:03.562 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:43:03.565 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:45:00.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 05:45:01.126 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:46:55.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.88s (1.90min) +2026-06-19 05:46:55.797 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:48:52.330 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:55:50.497 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 19, train_steps: 760 +2026-06-19 05:55:51.178 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 05:55:51.308 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:55:51.340 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:56:55.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:56:55.960 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:58:49.068 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 05:58:49.902 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:00:43.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.10s (1.89min) +2026-06-19 06:00:43.876 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:02:40.341 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:09:39.753 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 20, train_steps: 800 +2026-06-19 06:10:15.836 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=20 (missing=0, unexpected=0) +2026-06-19 06:10:18.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 06:10:18.887 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:10:18.937 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:11:25.147 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:11:25.150 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:13:19.050 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 06:13:19.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:15:13.224 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.90s (1.88min) +2026-06-19 06:15:13.910 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:17:12.688 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:24:12.561 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 21, train_steps: 840 +2026-06-19 06:24:13.270 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 06:24:13.422 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:24:13.471 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:25:17.850 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:25:17.852 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:27:11.261 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 06:27:12.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:29:06.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.51s (1.91min) +2026-06-19 06:29:07.667 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:31:04.580 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:38:04.285 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 22, train_steps: 880 +2026-06-19 06:38:04.948 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 06:38:05.124 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:38:05.231 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:39:09.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:39:09.768 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:41:03.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 06:41:04.259 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:42:58.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.99s (1.90min) +2026-06-19 06:42:59.164 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:44:56.506 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:51:58.180 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 23, train_steps: 920 +2026-06-19 06:51:58.842 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 06:51:58.966 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:51:58.997 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:53:03.398 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:53:03.402 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:54:59.789 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 06:55:00.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:56:54.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.11s (1.90min) +2026-06-19 06:56:55.621 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:58:52.321 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:05:51.318 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 24, train_steps: 960 +2026-06-19 07:05:52.004 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 07:05:52.130 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:05:52.170 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:06:56.949 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:06:56.952 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:08:52.217 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 07:08:53.127 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:10:46.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-19 07:10:46.871 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:12:44.154 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:19:41.978 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 25, train_steps: 1000 +2026-06-19 07:20:52.104 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 07:20:52.196 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:20:52.247 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:21:56.846 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:21:56.856 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:23:52.666 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 07:23:53.605 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:25:46.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.81s (1.88min) +2026-06-19 07:25:47.205 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:27:44.380 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:34:43.692 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 26, train_steps: 1040 +2026-06-19 07:34:44.357 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 07:34:44.551 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:34:44.668 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:35:48.805 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:35:48.805 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:37:45.443 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 07:37:46.368 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:39:39.639 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.14s (1.89min) +2026-06-19 07:39:40.348 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:41:39.072 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:48:39.675 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 27, train_steps: 1080 +2026-06-19 07:48:40.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 07:48:40.502 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:48:40.567 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:49:45.198 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:49:45.198 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:51:43.285 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 07:51:44.204 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:53:38.298 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.99s (1.90min) +2026-06-19 07:53:39.007 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:55:38.422 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:02:38.166 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 28, train_steps: 1120 +2026-06-19 08:02:38.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 08:02:38.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:02:39.027 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:03:43.678 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:03:43.682 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:05:41.901 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 08:05:43.028 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:07:37.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.24s (1.90min) +2026-06-19 08:07:38.091 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:09:37.935 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:16:37.848 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 29, train_steps: 1160 +2026-06-19 08:16:38.692 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 08:16:38.889 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:16:39.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:17:43.423 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:17:43.424 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:19:42.881 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 08:19:43.808 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:21:37.264 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 08:21:37.960 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:23:37.407 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:30:39.546 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 30, train_steps: 1200 +2026-06-19 08:31:15.768 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=30 (missing=0, unexpected=0) +2026-06-19 08:31:18.259 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 08:31:18.453 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:31:18.567 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:32:24.955 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:32:24.958 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:34:19.980 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 08:34:20.759 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:36:14.510 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-19 08:36:15.226 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:38:14.171 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:45:14.282 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 31, train_steps: 1240 +2026-06-19 08:45:14.993 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 08:45:15.134 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:45:15.171 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:46:20.221 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:46:20.224 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:48:15.288 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 08:48:16.296 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:50:10.323 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.84s (1.90min) +2026-06-19 08:50:11.023 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:52:11.148 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:59:13.099 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 32, train_steps: 1280 +2026-06-19 08:59:13.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 08:59:13.996 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:59:14.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:00:18.531 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:00:18.534 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:02:13.436 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 09:02:14.493 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:04:08.713 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.91s (1.90min) +2026-06-19 09:04:09.420 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:06:10.408 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:13:13.279 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 33, train_steps: 1320 +2026-06-19 09:13:13.952 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 09:13:14.110 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:13:14.169 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:14:18.785 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:14:18.789 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:16:14.822 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 09:16:15.704 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:18:11.066 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.06s (1.92min) +2026-06-19 09:18:11.785 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:20:11.625 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:27:13.735 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 34, train_steps: 1360 +2026-06-19 09:27:14.445 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 09:27:14.612 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:27:14.686 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:28:19.810 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:28:19.810 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:30:13.695 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 09:30:15.077 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:32:08.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.81s (1.90min) +2026-06-19 09:32:09.591 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:34:09.778 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:41:12.392 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 35, train_steps: 1400 +2026-06-19 09:41:13.060 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 09:41:13.265 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:41:13.386 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:42:17.844 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:42:17.847 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:44:13.080 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 09:44:14.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:46:10.686 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 116.56s (1.94min) +2026-06-19 09:46:11.397 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:48:10.896 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:55:11.958 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 36, train_steps: 1440 +2026-06-19 09:55:12.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 09:55:12.823 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:55:12.924 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:56:17.757 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:56:17.757 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:58:12.646 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 09:58:13.723 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:00:08.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-19 10:00:09.671 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:02:10.006 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:09:11.706 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 37, train_steps: 1480 +2026-06-19 10:09:12.397 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 10:09:12.628 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:09:12.787 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:10:17.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:10:17.149 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:12:13.042 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 10:12:14.047 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:14:09.030 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.94s (1.92min) +2026-06-19 10:14:09.736 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:16:10.365 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:23:12.349 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 38, train_steps: 1520 +2026-06-19 10:23:13.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 10:23:13.210 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:23:13.269 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:24:17.742 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:24:17.758 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:26:12.667 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 10:26:13.732 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:28:09.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.16s (1.92min) +2026-06-19 10:28:09.954 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:30:10.741 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:37:11.947 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 39, train_steps: 1560 +2026-06-19 10:37:12.646 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 10:37:12.809 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:37:12.893 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:38:17.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:38:17.442 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:41:42.526 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 10:41:43.656 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:43:36.471 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.70s (1.88min) +2026-06-19 10:43:37.179 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:45:37.175 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:52:37.854 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 40, train_steps: 1600 +2026-06-19 10:53:14.086 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=40 (missing=0, unexpected=0) +2026-06-19 10:53:15.773 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 10:53:15.954 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:53:16.047 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:54:22.486 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:54:22.486 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:56:16.410 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 10:56:17.344 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:58:11.717 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.27s (1.90min) +2026-06-19 10:58:12.416 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:00:11.328 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:07:12.586 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 41, train_steps: 1640 +2026-06-19 11:07:13.291 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 11:07:13.402 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:07:13.438 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:08:18.233 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:08:18.233 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:10:12.668 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 11:10:14.065 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:12:08.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.28s (1.90min) +2026-06-19 11:12:09.086 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:14:08.757 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:21:10.891 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 42, train_steps: 1680 +2026-06-19 11:21:11.804 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 11:21:11.964 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:21:12.020 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:22:16.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:22:16.713 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:24:12.073 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 11:24:13.165 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:26:07.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.32s (1.91min) +2026-06-19 11:26:08.278 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:28:06.797 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:35:06.771 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 43, train_steps: 1720 +2026-06-19 11:35:07.458 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 11:35:07.562 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:35:07.606 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:36:12.069 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:36:12.081 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:38:10.299 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 11:38:11.250 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:40:05.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.38s (1.91min) +2026-06-19 11:40:06.548 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:42:03.525 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:49:04.431 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 44, train_steps: 1760 +2026-06-19 11:49:05.115 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 11:49:05.341 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:49:05.463 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:50:09.848 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:50:09.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:52:06.491 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 11:52:07.336 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:54:01.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.69s (1.89min) +2026-06-19 11:54:01.909 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:56:01.234 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:03:03.871 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 45, train_steps: 1800 +2026-06-19 12:03:04.557 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 12:03:04.703 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:03:04.769 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:04:09.459 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:04:09.463 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:06:05.347 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 12:06:06.186 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:08:01.417 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.05s (1.92min) +2026-06-19 12:08:02.100 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:10:01.380 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:17:03.594 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 46, train_steps: 1840 +2026-06-19 12:17:04.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 12:17:04.472 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:17:04.572 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:18:09.116 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:18:09.119 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:20:05.333 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 12:20:06.372 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:22:01.733 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.17s (1.92min) +2026-06-19 12:22:02.441 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:24:01.449 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:31:02.931 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 47, train_steps: 1880 +2026-06-19 12:31:03.608 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 12:31:03.802 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:31:03.931 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:32:08.278 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:32:08.281 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:34:03.182 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 12:34:04.198 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:35:58.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-19 12:35:58.981 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:37:55.889 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:44:57.675 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 48, train_steps: 1920 +2026-06-19 12:44:58.355 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 12:44:58.504 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:44:58.562 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:46:03.075 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:46:03.077 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:47:58.931 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 12:47:59.795 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:49:54.620 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-19 12:49:55.343 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:51:52.289 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:58:52.094 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 49, train_steps: 1960 +2026-06-19 12:58:52.782 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 12:58:52.919 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:58:52.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:59:57.538 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:59:57.541 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:01:54.369 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 13:01:55.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:03:48.446 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.04s (1.88min) +2026-06-19 13:03:49.148 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:05:45.466 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:12:45.270 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 50, train_steps: 2000 +2026-06-19 13:14:21.177 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=50 (missing=0, unexpected=0) +2026-06-19 13:14:23.203 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 13:14:23.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:14:23.401 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:15:29.848 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:15:29.852 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:17:25.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 13:17:26.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:19:20.081 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.84s (1.90min) +2026-06-19 13:19:20.806 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:21:17.523 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:28:18.470 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 51, train_steps: 2040 +2026-06-19 13:28:19.184 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 13:28:19.286 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:28:19.323 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:29:23.803 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:29:23.803 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:31:18.029 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 13:31:18.923 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:33:13.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.83s (1.90min) +2026-06-19 13:33:13.741 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:35:13.227 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:42:13.614 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 52, train_steps: 2080 +2026-06-19 13:42:14.277 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 13:42:14.450 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:42:14.521 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:43:19.866 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:43:19.866 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:45:15.287 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 13:45:16.238 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:47:10.481 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.03s (1.90min) +2026-06-19 13:47:11.196 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:49:11.265 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:56:12.603 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 53, train_steps: 2120 +2026-06-19 13:56:13.303 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 13:56:13.420 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:56:13.458 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:57:17.916 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:57:17.919 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:59:13.130 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 13:59:14.060 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:01:08.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.41s (1.91min) +2026-06-19 14:01:09.529 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:03:10.472 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:10:12.163 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 54, train_steps: 2160 +2026-06-19 14:10:12.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 14:10:13.051 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:10:13.132 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:11:17.653 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:11:17.656 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:13:12.866 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 14:13:14.012 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:15:08.493 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.42s (1.91min) +2026-06-19 14:15:09.218 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:17:09.305 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:24:11.907 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 55, train_steps: 2200 +2026-06-19 14:24:12.604 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 14:24:12.819 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:24:12.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:25:17.683 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:25:17.683 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:27:12.754 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 14:27:14.039 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:29:09.888 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.82s (1.93min) +2026-06-19 14:29:10.604 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:31:10.763 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:38:12.117 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 56, train_steps: 2240 +2026-06-19 14:38:12.811 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 14:38:12.974 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:38:13.053 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:39:17.755 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:39:17.756 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:41:13.092 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 14:41:14.042 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:43:08.781 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.66s (1.91min) +2026-06-19 14:43:09.474 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:45:09.717 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:52:11.844 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 57, train_steps: 2280 +2026-06-19 14:52:12.528 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 14:52:12.738 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:52:12.844 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:53:17.227 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:53:17.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:55:13.057 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 14:55:14.028 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:57:08.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.79s (1.91min) +2026-06-19 14:57:09.597 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:59:09.880 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:06:12.285 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 58, train_steps: 2320 +2026-06-19 15:06:12.985 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 15:06:13.174 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:06:13.277 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:07:18.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:07:18.272 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:09:13.062 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 15:09:14.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:11:08.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.71s (1.91min) +2026-06-19 15:11:09.497 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:13:09.811 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:20:12.051 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 59, train_steps: 2360 +2026-06-19 15:20:12.739 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 15:20:12.979 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:20:13.098 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:21:17.801 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:21:17.804 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:23:12.372 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 15:23:13.434 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:25:09.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.28s (1.92min) +2026-06-19 15:25:09.742 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:27:10.366 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:34:12.070 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 60, train_steps: 2400 +2026-06-19 15:34:48.378 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=60 (missing=0, unexpected=0) +2026-06-19 15:34:50.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 15:34:50.824 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:34:50.932 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:35:56.899 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:35:56.902 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:37:53.529 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 15:37:54.667 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:39:47.707 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-19 15:39:48.403 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:41:45.234 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:48:44.580 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 61, train_steps: 2440 +2026-06-19 15:48:45.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 15:48:45.394 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:48:45.429 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:49:49.849 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:49:49.850 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:51:47.609 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 15:51:48.560 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:53:41.681 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.04s (1.88min) +2026-06-19 15:53:42.399 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:55:39.327 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:02:39.016 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 62, train_steps: 2480 +2026-06-19 16:02:39.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 16:02:39.792 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:02:39.839 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:03:44.253 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:03:44.254 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:05:44.132 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 16:05:45.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:07:38.761 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.59s (1.89min) +2026-06-19 16:07:39.470 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:09:38.852 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:16:39.318 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 63, train_steps: 2520 +2026-06-19 16:16:40.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 16:16:40.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:16:40.324 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:17:44.856 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:17:44.859 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:19:45.197 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 16:19:46.043 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:21:39.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.36s (1.89min) +2026-06-19 16:21:40.302 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:23:38.879 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:30:40.936 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 64, train_steps: 2560 +2026-06-19 16:30:41.626 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 16:30:41.787 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:30:41.833 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:31:46.501 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:31:46.504 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:33:45.471 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 16:33:46.319 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:35:40.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.63s (1.89min) +2026-06-19 16:35:40.843 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:37:39.397 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:44:41.291 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 65, train_steps: 2600 +2026-06-19 16:44:41.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 16:44:42.109 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:44:42.153 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:45:46.811 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:45:46.814 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:47:47.482 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 16:47:48.417 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:49:42.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.72s (1.90min) +2026-06-19 16:49:42.958 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:51:40.036 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:58:41.280 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 66, train_steps: 2640 +2026-06-19 16:58:41.989 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 16:58:42.094 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:58:42.126 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:59:46.671 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:59:46.676 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:01:43.599 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 17:01:44.542 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:03:37.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.34s (1.89min) +2026-06-19 17:03:38.681 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:05:38.383 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:12:40.127 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 67, train_steps: 2680 +2026-06-19 17:12:40.908 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 17:12:41.228 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:12:41.308 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:13:46.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:13:46.062 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:15:44.529 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 17:15:45.420 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:17:38.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 17:17:39.678 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:19:39.111 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:26:39.552 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 68, train_steps: 2720 +2026-06-19 17:26:40.215 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 17:26:40.415 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:26:40.539 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:27:45.375 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:27:45.378 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:29:43.886 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 17:29:44.755 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:31:38.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.76s (1.90min) +2026-06-19 17:31:39.420 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:33:38.457 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:40:38.369 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 69, train_steps: 2760 +2026-06-19 17:40:39.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 17:40:39.169 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:40:39.203 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:41:43.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:41:43.592 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:43:41.069 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 17:43:41.996 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:45:35.854 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.55s (1.89min) +2026-06-19 17:45:36.569 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:47:37.116 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:54:38.576 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 70, train_steps: 2800 +2026-06-19 17:55:15.134 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=70 (missing=0, unexpected=0) +2026-06-19 17:55:16.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 17:55:17.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:55:17.100 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:56:23.508 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:56:23.512 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:58:18.228 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 17:58:19.309 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:00:13.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.58s (1.89min) +2026-06-19 18:00:13.837 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:02:13.850 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:09:13.895 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 71, train_steps: 2840 +2026-06-19 18:09:14.615 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 18:09:14.778 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:09:14.863 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:10:19.903 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:10:19.903 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:12:16.021 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 18:12:16.990 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:14:12.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.67s (1.93min) +2026-06-19 18:14:13.459 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:16:12.694 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:23:13.830 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 72, train_steps: 2880 +2026-06-19 18:23:14.506 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 18:23:14.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:23:14.739 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:24:20.242 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:24:20.243 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:26:14.711 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 18:26:15.972 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:28:10.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.45s (1.91min) +2026-06-19 18:28:11.145 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:30:10.989 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:37:12.783 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 73, train_steps: 2920 +2026-06-19 18:37:13.462 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 18:37:13.664 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:37:13.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:38:18.625 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:38:18.625 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:40:13.250 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 18:40:14.402 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:42:09.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.16s (1.92min) +2026-06-19 18:42:10.285 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:44:10.447 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:51:12.754 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 74, train_steps: 2960 +2026-06-19 18:51:13.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 18:51:13.592 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:51:13.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:52:18.734 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:52:18.734 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:54:13.982 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 18:54:15.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:56:10.312 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.21s (1.92min) +2026-06-19 18:56:11.030 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:58:11.442 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:05:14.014 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 75, train_steps: 3000 +2026-06-19 19:06:23.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 19:06:23.341 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:06:23.392 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:07:28.129 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:07:28.138 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:09:24.209 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 19:09:25.055 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:11:18.397 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.14s (1.89min) +2026-06-19 19:11:19.100 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:13:16.252 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:20:16.316 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 76, train_steps: 3040 +2026-06-19 19:20:16.989 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 19:20:17.148 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:20:17.204 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:21:22.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:21:22.126 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:23:18.455 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 19:23:19.586 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:25:13.441 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.66s (1.89min) +2026-06-19 19:25:14.145 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:27:12.733 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:34:13.268 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 77, train_steps: 3080 +2026-06-19 19:34:13.951 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 19:34:14.158 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:34:14.282 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:35:18.733 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:35:18.733 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:37:13.074 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 19:37:14.062 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:39:08.734 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.60s (1.91min) +2026-06-19 19:39:09.440 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:41:09.769 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:48:12.659 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 78, train_steps: 3120 +2026-06-19 19:48:13.372 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 19:48:13.571 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:48:13.690 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:49:18.385 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:49:18.386 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:51:13.323 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 19:51:14.406 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:53:09.267 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-19 19:53:09.975 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:55:09.984 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:02:13.175 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 79, train_steps: 3160 +2026-06-19 20:02:13.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 20:02:14.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:02:14.139 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:03:18.756 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:03:18.759 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:05:13.093 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 20:05:14.053 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:07:09.660 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.53s (1.93min) +2026-06-19 20:07:10.385 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:09:11.032 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:16:13.936 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 80, train_steps: 3200 +2026-06-19 20:16:51.366 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=80 (missing=0, unexpected=0) +2026-06-19 20:16:53.093 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 20:16:53.270 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:16:53.415 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:18:00.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:18:00.300 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:19:54.752 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 20:19:55.705 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:21:49.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-19 20:21:50.362 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:23:47.274 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:30:47.433 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 81, train_steps: 3240 +2026-06-19 20:30:48.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 20:30:48.354 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:30:48.497 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:31:53.169 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:31:53.172 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:33:49.273 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 20:33:50.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:35:44.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.70s (1.89min) +2026-06-19 20:35:44.752 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:37:41.676 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:44:40.005 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 82, train_steps: 3280 +2026-06-19 20:44:40.696 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 20:44:40.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:44:40.994 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:45:45.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:45:45.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:47:44.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 20:47:45.070 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:49:38.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.41s (1.89min) +2026-06-19 20:49:39.274 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:51:38.422 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:58:38.998 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 83, train_steps: 3320 +2026-06-19 20:58:39.739 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 20:58:39.834 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:58:39.974 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:59:44.392 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:59:44.394 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:01:44.304 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 21:01:45.209 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:03:38.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.28s (1.89min) +2026-06-19 21:03:39.339 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:05:38.816 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:12:38.924 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 84, train_steps: 3360 +2026-06-19 21:12:39.649 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 21:12:39.850 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:12:39.974 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:13:44.437 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:13:44.444 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:15:43.791 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 21:15:44.642 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:17:38.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.31s (1.89min) +2026-06-19 21:17:38.863 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:19:38.521 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:26:39.498 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 85, train_steps: 3400 +2026-06-19 21:26:40.168 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 21:26:40.377 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:26:40.518 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:27:45.299 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:27:45.302 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:29:43.697 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 21:29:44.555 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:31:38.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-19 21:31:38.987 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:33:38.754 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:40:38.650 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 86, train_steps: 3440 +2026-06-19 21:40:39.367 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 21:40:39.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:40:39.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:41:44.419 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:41:44.422 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:43:42.867 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 21:43:43.720 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:45:37.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 21:45:37.986 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:47:37.706 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:54:39.051 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 87, train_steps: 3480 +2026-06-19 21:54:39.743 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 21:54:39.938 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:54:40.042 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:55:44.298 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:55:44.301 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:57:43.169 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 21:57:44.014 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:59:37.604 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.39s (1.89min) +2026-06-19 21:59:38.317 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:01:38.552 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:08:39.985 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 88, train_steps: 3520 +2026-06-19 22:08:40.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 22:08:40.851 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:08:40.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:09:45.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:09:45.393 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:11:43.755 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 22:11:44.611 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:13:38.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.57s (1.89min) +2026-06-19 22:13:39.103 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:15:38.303 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:22:38.699 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 89, train_steps: 3560 +2026-06-19 22:22:39.425 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 22:22:39.606 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:22:39.730 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:23:44.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:23:44.114 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:25:43.683 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 22:25:44.644 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:27:38.162 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 22:27:38.886 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:29:38.492 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:36:39.424 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 90, train_steps: 3600 +2026-06-19 22:37:15.772 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=90 (missing=0, unexpected=0) +2026-06-19 22:37:17.307 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 22:37:17.433 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:37:17.480 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:38:23.328 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:38:23.331 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:40:19.241 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 22:40:20.126 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:42:14.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.70s (1.89min) +2026-06-19 22:42:14.863 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:44:13.686 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:51:14.855 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 91, train_steps: 3640 +2026-06-19 22:51:15.699 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 22:51:15.870 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:51:15.944 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:52:21.078 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:52:21.082 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:54:15.937 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 22:54:16.932 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:56:12.672 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.67s (1.93min) +2026-06-19 22:56:13.388 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:58:13.110 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:05:14.261 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 92, train_steps: 3680 +2026-06-19 23:05:15.118 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 23:05:15.318 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:05:15.442 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:06:20.614 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:06:20.615 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:08:16.058 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 23:08:17.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:10:11.751 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.62s (1.91min) +2026-06-19 23:10:12.470 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:12:11.281 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:19:13.483 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 93, train_steps: 3720 +2026-06-19 23:19:14.152 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 23:19:14.316 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:19:14.400 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:20:19.789 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:20:19.790 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:22:14.354 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 23:22:15.611 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:24:10.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.33s (1.92min) +2026-06-19 23:24:11.707 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:26:11.285 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:33:13.940 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 94, train_steps: 3760 +2026-06-19 23:33:14.640 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 23:33:14.783 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:33:14.816 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:34:20.272 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:34:20.273 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:36:16.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 23:36:16.986 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:38:12.062 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.89s (1.91min) +2026-06-19 23:38:12.783 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:40:11.319 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:47:11.876 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 95, train_steps: 3800 +2026-06-19 23:47:12.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-19 23:47:12.766 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:47:12.836 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:48:17.282 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:48:17.285 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:50:12.004 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-19 23:50:13.071 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:52:08.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.82s (1.91min) +2026-06-19 23:52:08.790 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:54:07.675 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:01:09.777 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 96, train_steps: 3840 +2026-06-20 00:01:10.475 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 00:01:10.638 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:01:10.796 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:02:15.433 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:02:15.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:04:11.783 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 00:04:12.788 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:06:07.928 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 00:06:08.644 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:08:07.686 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:15:08.509 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 97, train_steps: 3880 +2026-06-20 00:15:09.218 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 00:15:09.358 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:15:09.391 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:16:13.835 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:16:13.838 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:18:11.238 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 00:18:12.296 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:20:07.671 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.20s (1.92min) +2026-06-20 00:20:08.400 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:22:07.125 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:29:07.001 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 98, train_steps: 3920 +2026-06-20 00:29:07.692 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 00:29:07.866 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:29:07.951 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:30:12.461 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:30:12.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:32:10.419 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 00:32:11.377 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:34:06.389 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 00:34:07.117 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:36:03.850 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:43:05.342 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 99, train_steps: 3960 +2026-06-20 00:43:06.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 00:43:06.160 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:43:06.194 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:44:10.774 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:44:10.779 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:46:09.899 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 00:46:10.880 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:48:05.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.75s (1.91min) +2026-06-20 00:48:06.439 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:50:03.397 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:57:04.801 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 100, train_steps: 4000 +2026-06-20 00:58:47.634 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=100 (missing=0, unexpected=0) +2026-06-20 00:58:49.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 00:58:49.792 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:58:49.911 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:59:55.607 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:59:55.611 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:01:50.303 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 01:01:51.258 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:03:44.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-20 01:03:45.434 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:05:42.220 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:12:39.870 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 101, train_steps: 4040 +2026-06-20 01:12:40.567 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 01:12:40.723 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:12:40.781 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:13:45.875 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:13:45.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:15:43.499 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 01:15:44.460 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:17:38.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.62s (1.89min) +2026-06-20 01:17:38.931 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:19:38.324 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:26:37.815 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 102, train_steps: 4080 +2026-06-20 01:26:38.504 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 01:26:38.617 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:26:38.649 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:27:43.251 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:27:43.252 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:29:41.183 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 01:29:42.388 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:31:36.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-20 01:31:37.098 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:33:37.225 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:40:38.538 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 103, train_steps: 4120 +2026-06-20 01:40:39.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 01:40:39.423 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:40:39.534 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:41:44.139 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:41:44.142 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:43:43.334 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 01:43:44.332 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:45:37.767 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-20 01:45:38.487 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:47:38.265 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:54:37.794 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 104, train_steps: 4160 +2026-06-20 01:54:38.702 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 01:54:38.844 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:54:38.887 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:55:43.605 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:55:43.608 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:57:42.889 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 01:57:43.843 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:59:37.314 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-20 01:59:38.030 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:01:38.019 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:08:38.450 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 105, train_steps: 4200 +2026-06-20 02:08:39.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 02:08:39.285 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:08:39.323 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:09:43.873 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:09:43.876 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:11:43.945 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 02:11:44.890 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:13:38.217 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 02:13:38.925 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:15:38.267 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:22:37.848 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 106, train_steps: 4240 +2026-06-20 02:22:38.565 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 02:22:38.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:22:38.893 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:23:43.235 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:23:43.236 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:25:40.847 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 02:25:41.806 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:27:35.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.90s (1.90min) +2026-06-20 02:27:36.582 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:29:36.811 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:36:37.090 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 107, train_steps: 4280 +2026-06-20 02:36:37.802 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 02:36:38.008 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:36:38.143 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:37:42.380 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:37:42.383 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:39:39.229 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 02:39:40.358 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:41:35.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 02:41:36.071 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:43:35.144 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:50:35.008 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 108, train_steps: 4320 +2026-06-20 02:50:35.726 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 02:50:35.846 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:50:35.883 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:51:40.508 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:51:40.511 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:53:37.574 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 02:53:38.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:55:33.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.45s (1.91min) +2026-06-20 02:55:33.972 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:57:30.647 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:04:30.697 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 109, train_steps: 4360 +2026-06-20 03:04:31.405 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 03:04:31.542 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:04:31.580 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:05:35.960 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:05:35.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:07:33.407 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 03:07:34.367 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:09:29.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-20 03:09:30.014 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:11:26.574 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:18:25.331 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 110, train_steps: 4400 +2026-06-20 03:19:01.537 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=110 (missing=0, unexpected=0) +2026-06-20 03:19:03.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 03:19:03.418 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:19:03.446 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:20:10.103 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:20:10.108 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:22:07.524 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 03:22:08.466 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:24:03.966 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.41s (1.92min) +2026-06-20 03:24:04.689 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:26:02.472 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:33:05.391 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 111, train_steps: 4440 +2026-06-20 03:33:06.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 03:33:06.167 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:33:06.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:34:10.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:34:10.765 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:36:10.098 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 03:36:11.066 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:38:06.021 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 03:38:06.731 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:40:04.482 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:47:04.312 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 112, train_steps: 4480 +2026-06-20 03:47:05.043 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 03:47:05.161 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:47:05.197 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:48:09.568 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:48:09.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:50:04.983 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 03:50:06.214 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:52:00.943 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-20 03:52:01.657 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:54:00.963 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:01:04.006 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 113, train_steps: 4520 +2026-06-20 04:01:04.874 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 04:01:05.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:01:05.089 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:02:09.419 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:02:09.424 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:04:06.611 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 04:04:07.568 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:06:02.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.15s (1.92min) +2026-06-20 04:06:03.532 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:08:01.583 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:15:03.783 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 114, train_steps: 4560 +2026-06-20 04:15:04.473 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 04:15:04.650 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:15:04.729 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:16:09.065 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:16:09.068 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:18:04.152 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 04:18:05.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:20:00.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.72s (1.91min) +2026-06-20 04:20:00.894 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:22:00.098 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:29:01.474 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 115, train_steps: 4600 +2026-06-20 04:29:02.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 04:29:02.343 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:29:02.402 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:30:06.754 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:30:06.759 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:32:02.494 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 04:32:03.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:33:58.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.33s (1.91min) +2026-06-20 04:33:58.775 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:35:55.624 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:42:56.469 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 116, train_steps: 4640 +2026-06-20 04:42:57.159 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 04:42:57.333 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:42:57.419 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:44:01.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:44:01.977 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:45:58.381 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 04:45:59.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:47:53.698 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.25s (1.90min) +2026-06-20 04:47:54.407 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:49:50.887 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:56:50.234 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 117, train_steps: 4680 +2026-06-20 04:56:50.953 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 04:56:51.126 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:56:51.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:57:55.691 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:57:55.697 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:59:51.798 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 04:59:52.665 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:01:46.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.49s (1.89min) +2026-06-20 05:01:47.078 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:03:44.145 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:10:42.599 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 118, train_steps: 4720 +2026-06-20 05:10:43.295 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 05:10:43.415 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:10:43.447 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:11:47.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:11:47.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:13:46.957 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 05:13:47.824 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:15:41.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.26s (1.89min) +2026-06-20 05:15:42.000 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:17:38.995 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:24:37.904 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 119, train_steps: 4760 +2026-06-20 05:24:38.622 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 05:24:38.729 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:24:38.773 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:25:43.382 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:25:43.385 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:27:40.119 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 05:27:41.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:29:34.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-20 05:29:35.559 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:31:33.104 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:38:32.149 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 120, train_steps: 4800 +2026-06-20 05:39:09.085 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=120 (missing=0, unexpected=0) +2026-06-20 05:39:10.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 05:39:10.972 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:39:11.028 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:40:17.453 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:40:17.456 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:42:12.566 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 05:42:13.424 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:44:08.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.96s (1.92min) +2026-06-20 05:44:09.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:46:10.085 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:53:12.063 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 121, train_steps: 4840 +2026-06-20 05:53:12.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 05:53:12.929 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:53:12.981 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:54:17.927 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:54:17.927 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:56:13.156 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 05:56:14.364 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:58:09.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.97s (1.92min) +2026-06-20 05:58:10.075 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:00:10.576 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:07:11.954 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 122, train_steps: 4880 +2026-06-20 06:07:12.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 06:07:12.762 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:07:12.803 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:08:17.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:08:17.397 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:10:13.195 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 06:10:14.335 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:12:08.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.25s (1.90min) +2026-06-20 06:12:09.325 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:14:10.579 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:21:11.869 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 123, train_steps: 4920 +2026-06-20 06:21:12.561 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 06:21:12.697 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:21:12.738 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:22:17.229 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:22:17.229 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:24:12.071 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 06:24:13.220 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:26:08.788 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.47s (1.92min) +2026-06-20 06:26:09.511 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:28:09.833 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:35:11.853 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 124, train_steps: 4960 +2026-06-20 06:35:12.537 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 06:35:12.695 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:35:12.743 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:36:17.190 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:36:17.190 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:38:12.129 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 06:38:13.167 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:40:08.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.44s (1.92min) +2026-06-20 06:40:09.425 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:42:09.772 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:49:11.646 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 125, train_steps: 5000 +2026-06-20 06:50:15.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 06:50:15.385 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:50:15.401 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:51:20.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:51:20.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:53:15.926 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 06:53:16.785 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:55:11.654 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.68s (1.91min) +2026-06-20 06:55:12.371 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:57:11.761 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:04:11.932 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 126, train_steps: 5040 +2026-06-20 07:04:12.630 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 07:04:12.750 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:04:12.778 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:05:17.739 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:05:17.739 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:07:13.988 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 07:07:15.330 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:09:10.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.08s (1.92min) +2026-06-20 07:09:11.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:11:11.163 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:18:11.932 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 127, train_steps: 5080 +2026-06-20 07:18:12.629 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 07:18:12.761 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:18:12.794 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:19:17.260 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:19:17.260 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:21:13.050 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 07:21:14.046 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:23:09.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.50s (1.93min) +2026-06-20 07:23:10.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:25:10.157 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:32:11.175 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 128, train_steps: 5120 +2026-06-20 07:32:11.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 07:32:12.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:32:12.088 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:33:16.382 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:33:16.401 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:35:13.016 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 07:35:14.036 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:37:08.641 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.56s (1.91min) +2026-06-20 07:37:09.373 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:39:09.928 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:46:11.992 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 129, train_steps: 5160 +2026-06-20 07:46:12.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 07:46:12.823 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:46:12.865 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:47:17.791 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:47:17.791 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:49:13.070 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 07:49:14.025 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:51:09.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-20 07:51:09.758 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:53:10.302 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:00:11.800 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 130, train_steps: 5200 +2026-06-20 08:00:49.469 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=130 (missing=0, unexpected=0) +2026-06-20 08:00:51.244 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 08:00:51.339 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:00:51.379 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:01:58.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:01:58.720 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:03:52.509 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 08:03:53.367 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:05:47.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.72s (1.90min) +2026-06-20 08:05:48.018 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:07:44.843 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:14:45.323 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 131, train_steps: 5240 +2026-06-20 08:14:46.025 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 08:14:46.215 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:14:46.313 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:15:51.079 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:15:51.082 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:17:46.990 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 08:17:47.854 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:19:41.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.38s (1.89min) +2026-06-20 08:19:42.138 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:21:39.467 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:28:37.832 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 132, train_steps: 5280 +2026-06-20 08:28:38.520 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 08:28:38.665 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:28:38.697 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:29:43.483 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:29:43.486 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:31:40.849 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 08:31:41.725 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:33:35.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.44s (1.89min) +2026-06-20 08:33:36.118 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:35:35.261 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:42:34.334 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 133, train_steps: 5320 +2026-06-20 08:42:35.017 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 08:42:35.193 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:42:35.267 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:43:39.887 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:43:39.892 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:45:35.888 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 08:45:36.747 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:47:32.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.77s (1.93min) +2026-06-20 08:47:33.449 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:49:30.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:56:29.433 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 134, train_steps: 5360 +2026-06-20 08:56:30.114 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 08:56:30.263 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:56:30.320 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:57:34.910 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:57:34.912 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:59:32.041 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 08:59:32.903 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:01:26.319 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 09:01:27.048 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:03:23.807 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:10:22.986 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 135, train_steps: 5400 +2026-06-20 09:10:23.675 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 09:10:23.836 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:10:23.893 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:11:28.511 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:11:28.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:13:23.932 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 09:13:24.780 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:15:18.588 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.61s (1.89min) +2026-06-20 09:15:19.310 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:17:17.100 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:24:16.804 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 136, train_steps: 5440 +2026-06-20 09:24:17.532 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 09:24:17.658 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:24:17.698 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:25:22.658 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:25:22.659 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:27:18.656 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 09:27:19.642 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:29:12.959 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.15s (1.89min) +2026-06-20 09:29:13.686 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:31:13.177 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:38:13.280 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 137, train_steps: 5480 +2026-06-20 09:38:13.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 09:38:14.061 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:38:14.100 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:39:18.688 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:39:18.693 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:41:14.371 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 09:41:15.640 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:43:09.522 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-20 09:43:10.250 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:45:10.503 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:52:12.161 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 138, train_steps: 5520 +2026-06-20 09:52:12.865 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 09:52:12.980 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:52:13.016 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:53:17.739 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:53:17.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:55:13.383 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 09:55:14.447 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:57:10.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.46s (1.92min) +2026-06-20 09:57:10.921 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:59:11.148 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:06:12.682 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 139, train_steps: 5560 +2026-06-20 10:06:13.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 10:06:13.503 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:06:13.542 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:07:18.338 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:07:18.339 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:09:13.264 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 10:09:14.446 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:11:08.675 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.20s (1.90min) +2026-06-20 10:11:09.397 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:13:09.780 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:20:12.685 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 140, train_steps: 5600 +2026-06-20 10:20:49.959 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=140 (missing=0, unexpected=0) +2026-06-20 10:20:51.813 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 10:20:51.893 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:20:51.933 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:21:58.360 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:21:58.361 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:23:53.040 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 10:23:53.986 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:25:47.600 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-20 10:25:48.313 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:27:45.363 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:34:44.002 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 141, train_steps: 5640 +2026-06-20 10:34:44.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 10:34:44.783 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:34:44.830 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:35:49.515 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:35:49.518 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:37:46.846 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 10:37:47.769 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:39:41.415 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.51s (1.89min) +2026-06-20 10:39:42.148 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:41:39.460 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:48:38.464 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 142, train_steps: 5680 +2026-06-20 10:48:39.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 10:48:39.279 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:48:39.314 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:49:43.940 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:49:43.940 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:51:43.153 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 10:51:44.016 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:53:37.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-20 10:53:37.851 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:55:37.992 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:02:37.922 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 143, train_steps: 5720 +2026-06-20 11:02:38.631 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 11:02:38.751 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:02:38.787 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:03:43.565 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:03:43.600 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:05:42.169 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 11:05:43.196 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:07:36.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 11:07:37.312 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:09:36.842 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:16:37.102 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 144, train_steps: 5760 +2026-06-20 11:16:37.801 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 11:16:37.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:16:38.109 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:17:42.363 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:17:42.367 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:19:38.707 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 11:19:39.656 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:21:33.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.47s (1.89min) +2026-06-20 11:21:34.087 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:23:30.883 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:30:30.403 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 145, train_steps: 5800 +2026-06-20 11:30:31.096 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 11:30:31.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:30:31.249 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:31:35.743 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:31:35.746 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:33:33.208 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 11:33:34.084 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:35:28.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.89s (1.90min) +2026-06-20 11:35:28.900 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:37:25.685 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:44:23.915 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 146, train_steps: 5840 +2026-06-20 11:44:24.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 11:44:24.742 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:44:24.784 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:45:29.302 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:45:29.306 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:47:23.247 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 11:47:24.215 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:49:19.134 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-20 11:49:19.856 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:51:17.037 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:58:15.933 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 147, train_steps: 5880 +2026-06-20 11:58:16.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 11:58:16.743 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:58:16.782 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:59:22.265 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:59:22.266 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:01:16.232 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 12:01:17.212 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:03:12.586 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.29s (1.92min) +2026-06-20 12:03:13.293 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:05:12.905 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:12:12.697 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 148, train_steps: 5920 +2026-06-20 12:12:13.405 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 12:12:13.522 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:12:13.548 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:13:18.456 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:13:18.460 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:15:13.144 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 12:15:14.323 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:17:09.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.69s (1.91min) +2026-06-20 12:17:09.788 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:19:10.586 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:26:11.389 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 149, train_steps: 5960 +2026-06-20 12:26:12.136 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 12:26:12.298 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:26:12.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:27:16.773 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:27:16.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:29:13.029 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 12:29:14.037 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:31:08.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.72s (1.91min) +2026-06-20 12:31:09.533 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:33:09.830 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:40:10.986 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 150, train_steps: 6000 +2026-06-20 12:41:53.544 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=150 (missing=0, unexpected=0) +2026-06-20 12:41:56.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 12:41:56.235 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:41:56.327 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:43:02.463 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:43:02.468 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:44:59.744 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 12:45:00.635 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:46:55.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.70s (1.91min) +2026-06-20 12:46:56.236 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:48:52.987 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:55:51.609 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 151, train_steps: 6040 +2026-06-20 12:55:52.305 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 12:55:52.414 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:55:52.456 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:56:56.818 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:56:56.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:58:50.721 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 12:58:51.585 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:00:46.190 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.41s (1.91min) +2026-06-20 13:00:46.907 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:02:43.651 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:09:41.905 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 152, train_steps: 6080 +2026-06-20 13:09:42.596 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 13:09:42.719 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:09:42.755 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:10:47.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:10:47.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:12:43.910 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 13:12:44.894 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:14:38.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-20 13:14:39.491 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:16:38.550 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:23:37.220 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 153, train_steps: 6120 +2026-06-20 13:23:38.091 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 13:23:38.215 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:23:38.255 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:24:43.330 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:24:43.331 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:26:40.127 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 13:26:41.254 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:28:34.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.39s (1.89min) +2026-06-20 13:28:35.472 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:30:32.907 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:37:31.708 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 154, train_steps: 6160 +2026-06-20 13:37:32.392 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 13:37:32.512 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:37:32.544 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:38:37.083 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:38:37.087 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:40:33.788 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 13:40:34.693 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:42:29.329 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.47s (1.91min) +2026-06-20 13:42:30.061 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:44:27.160 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:51:24.979 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 155, train_steps: 6200 +2026-06-20 13:51:25.702 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 13:51:25.880 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:51:25.977 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:52:30.301 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:52:30.305 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:54:24.447 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 13:54:25.404 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:56:19.129 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.63s (1.89min) +2026-06-20 13:56:19.859 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:58:16.748 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:05:15.346 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 156, train_steps: 6240 +2026-06-20 14:05:16.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 14:05:16.136 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:05:16.181 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:06:21.995 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:06:21.996 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:08:15.931 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 14:08:16.789 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:10:11.722 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.74s (1.91min) +2026-06-20 14:10:12.456 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:12:11.102 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:19:12.250 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 157, train_steps: 6280 +2026-06-20 14:19:12.923 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 14:19:13.027 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:19:13.070 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:20:17.781 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:20:17.781 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:22:12.398 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 14:22:13.379 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:24:08.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.22s (1.92min) +2026-06-20 14:24:09.689 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:26:09.903 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:33:10.832 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 158, train_steps: 6320 +2026-06-20 14:33:11.520 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 14:33:11.737 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:33:11.870 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:34:16.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:34:16.036 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:36:10.935 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 14:36:11.990 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:38:07.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.88s (1.91min) +2026-06-20 14:38:07.784 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:40:05.256 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:47:04.569 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 159, train_steps: 6360 +2026-06-20 14:47:05.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 14:47:05.475 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:47:05.575 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:48:09.948 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:48:09.951 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:50:04.049 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 14:50:05.088 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:51:59.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.86s (1.90min) +2026-06-20 14:51:59.928 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:53:57.320 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:00:59.192 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 160, train_steps: 6400 +2026-06-20 15:01:37.815 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=160 (missing=0, unexpected=0) +2026-06-20 15:01:39.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 15:01:39.388 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:01:39.447 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:02:45.272 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:02:45.275 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:04:42.879 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 15:04:43.730 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:06:37.409 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.48s (1.89min) +2026-06-20 15:06:38.136 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:08:38.301 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:15:37.526 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 161, train_steps: 6440 +2026-06-20 15:15:38.206 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 15:15:38.330 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:15:38.367 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:16:42.925 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:16:42.928 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:18:40.375 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 15:18:41.422 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:20:35.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-20 15:20:36.332 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:22:36.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:29:36.899 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 162, train_steps: 6480 +2026-06-20 15:29:37.568 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 15:29:37.686 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:29:37.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:30:42.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:30:42.159 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:32:37.444 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 15:32:38.309 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:34:33.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.48s (1.91min) +2026-06-20 15:34:33.827 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:36:31.315 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:43:31.486 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 163, train_steps: 6520 +2026-06-20 15:43:32.171 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 15:43:32.327 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:43:32.382 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:44:36.952 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:44:36.955 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:46:33.676 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 15:46:34.538 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:48:29.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.47s (1.91min) +2026-06-20 15:48:29.934 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:50:26.537 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:57:25.506 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 164, train_steps: 6560 +2026-06-20 15:57:26.228 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 15:57:26.367 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:57:26.406 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:58:30.774 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:58:30.774 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:00:26.402 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 16:00:27.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:02:21.064 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.61s (1.89min) +2026-06-20 16:02:21.792 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:04:18.346 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:11:18.385 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 165, train_steps: 6600 +2026-06-20 16:11:19.085 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 16:11:19.181 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:11:19.226 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:12:24.029 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:12:24.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:14:19.106 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 16:14:20.120 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:16:14.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.75s (1.90min) +2026-06-20 16:16:14.869 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:18:14.124 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:25:13.927 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 166, train_steps: 6640 +2026-06-20 16:25:14.621 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 16:25:14.769 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:25:14.802 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:26:20.316 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:26:20.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:28:13.831 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 16:28:14.801 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:30:08.844 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.78s (1.90min) +2026-06-20 16:30:09.574 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:32:10.304 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:39:12.158 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 167, train_steps: 6680 +2026-06-20 16:39:12.848 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 16:39:13.007 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:39:13.071 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:40:17.867 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:40:17.869 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:42:12.095 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 16:42:13.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:44:08.745 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.36s (1.92min) +2026-06-20 16:44:09.476 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:46:10.018 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:53:11.379 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 168, train_steps: 6720 +2026-06-20 16:53:12.091 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 16:53:12.252 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:53:12.311 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:54:17.155 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:54:17.155 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:56:11.767 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 16:56:12.625 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:58:07.736 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-20 16:58:08.463 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:00:07.578 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:07:09.177 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 169, train_steps: 6760 +2026-06-20 17:07:09.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 17:07:10.115 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:07:10.275 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:08:14.804 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:08:14.806 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:10:11.904 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 17:10:12.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:12:08.255 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.20s (1.92min) +2026-06-20 17:12:08.968 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:14:09.318 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:21:11.050 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 170, train_steps: 6800 +2026-06-20 17:21:47.067 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=170 (missing=0, unexpected=0) +2026-06-20 17:21:49.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 17:21:49.367 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:21:49.411 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:22:54.873 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:22:54.874 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:24:49.925 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 17:24:50.769 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:26:44.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.05s (1.88min) +2026-06-20 17:26:44.743 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:28:41.424 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:35:39.743 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 13, sp_rank 0: update step: 171, train_steps: 6840 +2026-06-20 17:35:40.431 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 13, dp_rank 6: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=0 +2026-06-20 17:35:40.542 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:35:40.593 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:36:45.520 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 13 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:36:45.522 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 13 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:38:42.537 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 13: Using local metric names without global sync: ['avg'] +2026-06-20 17:38:43.690 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 13: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:40:36.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 13: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.89s (1.88min) +2026-06-20 17:40:37.377 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 13, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-29 01:28:16.742 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 01:28:16.745 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 01:51:23.452 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 01:51:23.453 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 02:21:21.521 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 02:21:21.524 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 02:46:50.878 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 02:46:50.879 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:08:56.783 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:08:56.784 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:32:26.725 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:32:26.727 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:59:01.817 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:59:01.817 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:27:01.100 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:27:01.102 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:52:08.389 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:52:08.391 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 05:16:07.790 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 05:16:07.790 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 05:39:45.776 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 05:39:45.777 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:03:36.211 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:03:36.211 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:25:05.454 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:25:05.454 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:49:46.613 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:49:46.615 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 07:12:17.188 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 07:12:17.190 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 07:35:01.531 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 07:35:01.533 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:05:09.219 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:05:09.221 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:34:41.413 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:34:41.413 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:59:04.349 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:59:04.351 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:22:49.793 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:22:49.793 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:47:48.502 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:47:48.505 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 10:11:33.785 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 10:11:33.786 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 10:36:01.007 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 10:36:01.008 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:03:38.764 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:03:38.767 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:27:27.953 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:27:27.953 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:55:35.249 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:55:35.252 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 12:18:29.788 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 12:18:29.789 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 12:46:33.143 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 12:46:33.144 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 13:12:55.375 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 13:12:55.377 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 13:37:16.178 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 13:37:16.180 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:04:10.124 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:04:10.124 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:27:06.376 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:27:06.377 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:49:26.142 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:49:26.143 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 15:21:05.714 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 15:21:05.717 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 15:46:20.640 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 15:46:20.643 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 16:16:21.799 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 16:16:21.800 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 16:42:12.838 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 16:42:12.839 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:08:24.336 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:08:24.336 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:32:10.531 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:32:10.534 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:53:40.877 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:53:40.880 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 18:16:14.493 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 18:16:14.494 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 18:39:04.910 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 18:39:04.911 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_14.log b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_14.log new file mode 100644 index 0000000000000000000000000000000000000000..e07c376863a0433fab3799f940d4ac5e3640ca51 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_14.log @@ -0,0 +1,3241 @@ +2026-06-19 01:22:18.214 | INFO | __main__:__init__:933 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-19 01:22:18.898 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-19 01:22:18.899 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-19 01:30:17.295 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 01:30:31.279 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 01:30:31.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:30:31.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:31:52.795 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_3/index_2841_seed_6716_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 01:31:55.857 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 01:32:14.639 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:32:14.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:32:14.649 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:33:22.277 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_3/index_2841_seed_6717_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 01:33:27.333 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 01:33:44.560 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:33:45.515 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 01:34:00.944 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 01:35:56.997 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.94s (1.93min) +2026-06-19 01:35:57.686 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 01:37:54.181 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 01:37:57.219 | INFO | hyvideo.utils.grpo_utils:train_one_step:3127 - Rank 14, sp_rank 0: init adaptive KL controller (init=1.000e-07, min=1.000e-07, max=1.000e-05) +2026-06-19 01:45:11.842 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 1, train_steps: 40 +2026-06-19 01:45:12.576 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 01:45:12.743 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 01:45:12.810 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:45:12.860 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:46:19.022 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000001/group_3/index_6359_seed_23301_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 01:46:23.829 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 01:46:41.405 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:46:41.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:46:41.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:47:47.191 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000001/group_3/index_6359_seed_23302_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 01:47:51.941 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 01:48:09.631 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:48:10.945 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 01:48:12.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 01:50:07.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.70s (1.93min) +2026-06-19 01:50:08.465 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 01:52:06.954 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 01:59:06.766 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 2, train_steps: 80 +2026-06-19 01:59:07.381 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 01:59:07.519 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 01:59:07.572 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:59:07.625 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:00:13.228 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000002/group_3/index_6174_seed_3012_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 02:00:17.719 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:00:34.901 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:00:34.909 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:00:34.909 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:01:46.279 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000002/group_3/index_6174_seed_3013_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 02:01:51.300 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:02:08.922 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:02:10.528 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:02:11.726 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:04:06.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.88s (1.91min) +2026-06-19 02:04:07.374 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:06:03.754 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:13:04.860 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 3, train_steps: 120 +2026-06-19 02:13:05.465 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 02:13:05.575 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:13:05.592 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:13:05.613 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:14:11.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_3/index_2516_seed_15651_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 02:14:16.794 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:14:34.111 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:14:34.119 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:14:34.124 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:15:40.435 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_3/index_2516_seed_15652_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 02:15:45.249 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:16:02.798 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:16:04.655 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:16:05.773 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:17:59.773 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.77s (1.90min) +2026-06-19 02:18:00.448 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:19:59.350 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:26:59.894 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 4, train_steps: 160 +2026-06-19 02:27:00.547 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 02:27:00.758 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:27:00.892 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:27:00.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:28:06.717 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000004/group_3/index_1721_seed_12485_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 02:28:10.510 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:28:28.143 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:28:28.152 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:28:28.152 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:29:35.138 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000004/group_3/index_1721_seed_12486_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 02:29:38.273 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:29:55.625 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:30:00.316 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:30:01.168 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:31:55.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.62s (1.91min) +2026-06-19 02:31:56.651 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:33:53.222 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:40:53.227 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 5, train_steps: 200 +2026-06-19 02:40:53.902 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 02:40:54.094 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:40:54.200 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:40:54.301 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:41:59.963 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_3/index_5453_seed_21629_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 02:42:05.029 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:42:22.690 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:42:22.699 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:42:22.701 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:43:28.485 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_3/index_5453_seed_21630_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 02:43:32.579 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:43:50.271 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:43:51.812 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:43:52.669 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:45:46.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.50s (1.89min) +2026-06-19 02:45:46.993 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:47:43.350 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:54:40.662 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 6, train_steps: 240 +2026-06-19 02:54:41.361 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 02:54:41.522 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:54:41.620 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:54:41.925 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:55:47.666 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_3/index_4380_seed_30902_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 02:55:52.645 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:56:10.198 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:56:10.207 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:56:10.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:57:19.364 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_3/index_4380_seed_30903_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 02:57:24.207 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:57:41.915 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:57:42.949 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:57:43.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:59:37.354 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.38s (1.89min) +2026-06-19 02:59:38.043 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:01:37.730 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:08:37.480 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 7, train_steps: 280 +2026-06-19 03:08:38.147 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 03:08:38.311 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:08:38.387 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:08:38.487 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:09:44.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_3/index_639_seed_11462_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 03:09:48.614 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:10:05.893 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:10:05.903 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:10:05.903 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:11:14.665 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_3/index_639_seed_11463_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 03:11:17.821 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:11:35.226 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:11:38.337 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:11:39.387 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:13:34.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.68s (1.91min) +2026-06-19 03:13:34.977 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:15:32.041 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:22:31.156 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 8, train_steps: 320 +2026-06-19 03:22:31.814 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 03:22:31.932 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:22:31.970 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:22:32.038 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:23:38.584 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000008/group_3/index_2533_seed_16469_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 03:23:43.666 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:24:01.372 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:24:01.380 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:24:01.380 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:25:09.047 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000008/group_3/index_2533_seed_16470_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 03:25:12.988 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:25:32.464 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:25:33.100 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:25:33.947 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:27:28.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-19 03:27:29.431 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:29:26.903 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:36:25.651 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 9, train_steps: 360 +2026-06-19 03:36:26.346 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 03:36:26.540 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:36:26.640 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:36:26.735 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:37:32.836 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_3/index_3137_seed_23252_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 03:37:37.834 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:37:55.428 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:37:55.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:37:55.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:39:02.083 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_3/index_3137_seed_23253_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 03:39:06.382 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:39:23.966 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:39:26.010 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:39:26.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:41:20.400 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.36s (1.89min) +2026-06-19 03:41:21.105 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:43:17.561 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:50:16.935 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 10, train_steps: 400 +2026-06-19 03:50:53.433 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=10 (missing=0, unexpected=0) +2026-06-19 03:50:56.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 03:50:56.329 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:50:56.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:50:56.511 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:52:03.814 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_3/index_2578_seed_17562_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 03:52:06.987 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:52:24.739 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:52:24.749 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:52:24.751 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:53:32.563 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_3/index_2578_seed_17563_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 03:53:37.607 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:53:55.376 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:53:57.828 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:53:58.848 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:55:52.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.65s (1.89min) +2026-06-19 03:55:53.208 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:57:49.763 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:04:48.379 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 11, train_steps: 440 +2026-06-19 04:04:49.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 04:04:49.199 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:04:49.234 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:04:49.299 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:05:55.319 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_3/index_4458_seed_1266_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 04:05:59.509 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:06:16.700 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:06:16.709 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:06:16.709 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:07:24.631 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_3/index_4458_seed_1267_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 04:07:27.832 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:07:45.485 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:07:48.853 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:07:49.707 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:09:43.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.25s (1.89min) +2026-06-19 04:09:43.797 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:11:40.267 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:18:37.292 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 12, train_steps: 480 +2026-06-19 04:18:37.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 04:18:38.128 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:18:38.186 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:18:38.298 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:19:44.336 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_3/index_4598_seed_26272_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 04:19:47.510 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:20:06.355 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:20:06.363 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:20:06.364 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:21:12.804 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_3/index_4598_seed_26273_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 04:21:17.918 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:21:37.169 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:21:37.180 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:21:38.190 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:23:33.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.19s (1.92min) +2026-06-19 04:23:34.070 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:25:31.243 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:32:29.824 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 13, train_steps: 520 +2026-06-19 04:32:30.512 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 04:32:30.609 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:32:30.655 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:32:30.703 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:33:36.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_3/index_8171_seed_30420_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 04:33:40.662 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:33:58.164 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:33:58.172 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:33:58.172 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:35:06.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_3/index_8171_seed_30421_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 04:35:09.306 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:35:26.980 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:35:30.313 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:35:31.216 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:37:24.102 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.77s (1.88min) +2026-06-19 04:37:24.791 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:39:21.076 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:46:18.588 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 14, train_steps: 560 +2026-06-19 04:46:19.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 04:46:19.360 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:46:19.401 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:46:19.466 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:47:25.950 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_3/index_9542_seed_27286_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 04:47:29.142 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:47:46.723 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:47:46.732 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:47:46.732 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:48:54.302 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_3/index_9542_seed_27287_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 04:48:57.525 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:49:16.053 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:49:17.445 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:49:18.289 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:51:12.963 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.49s (1.91min) +2026-06-19 04:51:13.650 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:53:12.994 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:00:11.833 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 15, train_steps: 600 +2026-06-19 05:00:12.526 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 05:00:12.742 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:00:12.859 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:00:12.920 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:01:18.949 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000015/group_3/index_9696_seed_6101_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 05:01:22.717 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:01:40.417 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:01:40.426 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:01:40.429 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:02:47.662 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000015/group_3/index_9696_seed_6102_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 05:02:50.898 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:03:08.703 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:03:11.349 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:03:12.510 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:05:07.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.64s (1.91min) +2026-06-19 05:05:08.060 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:07:05.978 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:14:04.897 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 16, train_steps: 640 +2026-06-19 05:14:05.573 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 05:14:05.673 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:14:05.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:14:05.771 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:15:11.650 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_3/index_8253_seed_26395_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 05:15:14.951 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:15:33.134 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:15:33.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:15:33.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:16:43.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_3/index_8253_seed_26396_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 05:16:48.339 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:17:06.374 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:17:06.988 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:17:07.984 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:19:03.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.02s (1.92min) +2026-06-19 05:19:03.741 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:21:01.961 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:28:03.329 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 17, train_steps: 680 +2026-06-19 05:28:04.007 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 05:28:04.182 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:28:04.257 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:28:04.378 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:29:10.440 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_3/index_9680_seed_8803_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 05:29:15.244 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:29:33.102 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:29:33.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:29:33.112 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:30:39.468 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_3/index_9680_seed_8804_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 05:30:43.818 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:31:01.275 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:31:03.926 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:31:04.933 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:32:58.689 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.48s (1.89min) +2026-06-19 05:32:59.386 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:34:56.264 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:41:58.100 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 18, train_steps: 720 +2026-06-19 05:41:58.777 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 05:41:58.986 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:41:59.127 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:41:59.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:43:04.949 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_3/index_8394_seed_13003_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 05:43:09.958 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:43:27.557 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:43:27.566 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:43:27.566 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:44:35.320 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_3/index_8394_seed_13004_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 05:44:40.233 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:44:58.032 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:45:00.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:45:01.214 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:46:55.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.88s (1.90min) +2026-06-19 05:46:55.797 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:48:52.331 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:55:50.494 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 19, train_steps: 760 +2026-06-19 05:55:51.178 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 05:55:51.338 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:55:51.414 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:55:51.513 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:56:57.454 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_3/index_6832_seed_19280_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 05:57:02.363 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:57:20.365 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:57:20.374 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:57:20.378 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:58:25.477 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_3/index_6832_seed_19281_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 05:58:30.601 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:58:48.415 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:58:49.068 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:58:50.068 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:00:43.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.10s (1.89min) +2026-06-19 06:00:43.876 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:02:40.341 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:09:39.761 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 20, train_steps: 800 +2026-06-19 06:10:16.262 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=20 (missing=0, unexpected=0) +2026-06-19 06:10:18.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 06:10:18.971 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:10:19.091 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:10:19.170 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:11:26.592 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_3/index_1957_seed_30415_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 06:11:31.375 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:11:48.758 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:11:48.768 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:11:48.768 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:12:54.611 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_3/index_1957_seed_30416_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 06:12:58.742 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:13:16.393 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:13:19.050 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:13:20.077 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:15:13.224 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.90s (1.88min) +2026-06-19 06:15:13.911 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:17:12.688 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:24:12.562 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 21, train_steps: 840 +2026-06-19 06:24:13.271 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 06:24:13.448 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:24:13.543 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:24:13.621 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:25:19.487 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_3/index_7672_seed_33100_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 06:25:24.624 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:25:42.288 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:25:42.297 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:25:42.297 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:26:47.522 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_3/index_7672_seed_33101_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 06:26:51.643 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:27:09.520 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:27:11.261 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:27:12.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:29:06.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.51s (1.91min) +2026-06-19 06:29:07.667 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:31:04.580 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:38:04.289 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 22, train_steps: 880 +2026-06-19 06:38:04.948 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 06:38:05.142 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:38:05.267 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:38:05.378 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:39:10.959 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_3/index_3184_seed_8662_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 06:39:16.154 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:39:33.838 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:39:33.846 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:39:33.847 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:40:39.781 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_3/index_3184_seed_8663_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 06:40:44.713 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:41:02.610 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:41:03.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:41:04.338 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:42:58.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.99s (1.90min) +2026-06-19 06:42:59.164 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:44:56.505 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:51:58.182 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 23, train_steps: 920 +2026-06-19 06:51:58.842 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 06:51:58.996 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:51:59.050 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:51:59.168 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:53:04.888 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000023/group_3/index_7777_seed_17403_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 06:53:09.168 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:53:26.553 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:53:26.562 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:53:26.562 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:54:35.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000023/group_3/index_7777_seed_17404_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 06:54:40.085 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:54:58.182 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:54:59.789 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:55:00.818 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:56:54.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.11s (1.90min) +2026-06-19 06:56:55.621 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:58:52.321 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:05:51.322 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 24, train_steps: 960 +2026-06-19 07:05:52.004 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 07:05:52.115 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:05:52.155 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:05:52.204 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:06:58.089 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_3/index_9614_seed_30112_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 07:07:02.314 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:07:19.803 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:07:19.811 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:07:19.811 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:08:27.513 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_3/index_9614_seed_30113_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 07:08:30.711 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:08:48.561 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:08:52.217 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:08:53.112 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:10:46.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-19 07:10:46.871 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:12:44.153 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:19:41.997 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 25, train_steps: 1000 +2026-06-19 07:20:52.105 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 07:20:52.251 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:20:52.318 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:20:52.443 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:21:58.377 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_3/index_8693_seed_5714_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 07:22:03.464 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:22:21.667 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:22:21.676 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:22:21.676 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:23:28.773 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_3/index_8693_seed_5715_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 07:23:32.856 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:23:51.097 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:23:52.666 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:23:53.597 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:25:46.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.81s (1.88min) +2026-06-19 07:25:47.205 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:27:44.379 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:34:43.699 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 26, train_steps: 1040 +2026-06-19 07:34:44.357 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 07:34:44.564 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:34:44.701 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:34:44.760 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:35:50.346 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_3/index_4280_seed_14354_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 07:35:53.495 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:36:11.031 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:36:11.041 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:36:11.041 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:37:21.053 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_3/index_4280_seed_14355_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 07:37:26.240 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:37:44.515 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:37:45.443 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:37:46.336 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:39:39.639 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.14s (1.89min) +2026-06-19 07:39:40.349 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:41:39.070 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:48:39.683 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 27, train_steps: 1080 +2026-06-19 07:48:40.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 07:48:40.550 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:48:40.676 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:48:40.768 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:49:46.745 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_3/index_3307_seed_19683_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 07:49:51.630 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:50:09.938 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:50:09.947 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:50:09.948 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:51:19.289 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_3/index_3307_seed_19684_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 07:51:24.247 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:51:42.740 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:51:43.285 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:51:44.294 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:53:38.298 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.99s (1.90min) +2026-06-19 07:53:39.007 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:55:38.422 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:02:38.171 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 28, train_steps: 1120 +2026-06-19 08:02:38.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 08:02:39.067 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:02:39.175 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:02:39.258 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:03:45.274 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_3/index_2611_seed_30869_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 08:03:50.151 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:04:08.281 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:04:08.291 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:04:08.291 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:05:17.530 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_3/index_2611_seed_30870_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 08:05:22.603 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:05:40.768 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:05:41.901 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:05:43.122 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:07:37.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.24s (1.90min) +2026-06-19 08:07:38.092 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:09:37.935 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:16:37.856 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 29, train_steps: 1160 +2026-06-19 08:16:38.691 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 08:16:38.837 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:16:38.885 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:16:39.137 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:17:44.965 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_3/index_3204_seed_19505_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 08:17:49.259 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:18:07.140 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:18:07.149 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:18:07.149 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:19:18.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_3/index_3204_seed_19506_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 08:19:21.252 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:19:39.427 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:19:42.881 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:19:43.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:21:37.264 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 08:21:37.961 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:23:37.406 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:30:39.548 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 30, train_steps: 1200 +2026-06-19 08:31:15.785 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=30 (missing=0, unexpected=0) +2026-06-19 08:31:18.258 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 08:31:18.422 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:31:18.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:31:18.615 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:32:26.561 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000030/group_3/index_9721_seed_27081_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 08:32:29.808 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:32:47.829 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:32:47.839 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:32:47.839 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:33:55.681 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000030/group_3/index_9721_seed_27082_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 08:34:00.974 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:34:19.565 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:34:19.980 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:34:21.137 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:36:14.510 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-19 08:36:15.227 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:38:14.173 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:45:14.287 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 31, train_steps: 1240 +2026-06-19 08:45:14.993 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 08:45:15.178 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:45:15.290 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:45:15.374 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:46:21.837 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_3/index_6154_seed_18000_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 08:46:26.770 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:46:45.017 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:46:45.026 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:46:45.026 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:47:50.619 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_3/index_6154_seed_18001_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 08:47:55.608 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:48:15.240 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:48:15.288 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:48:16.453 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:50:10.323 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.84s (1.90min) +2026-06-19 08:50:11.023 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:52:11.149 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:59:13.095 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 32, train_steps: 1280 +2026-06-19 08:59:13.775 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 08:59:13.883 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:59:13.927 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:59:13.974 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:00:20.084 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000032/group_3/index_2794_seed_5439_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 09:00:25.086 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:00:43.379 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:00:43.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:00:43.385 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:01:48.625 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000032/group_3/index_2794_seed_5440_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 09:01:52.828 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:02:11.229 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:02:13.436 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:02:14.553 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:04:08.713 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.91s (1.90min) +2026-06-19 09:04:09.421 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:06:10.408 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:13:13.282 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 33, train_steps: 1320 +2026-06-19 09:13:13.952 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 09:13:14.093 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:13:14.132 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:13:14.226 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:14:20.854 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000033/group_3/index_3201_seed_7523_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 09:14:25.998 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:14:44.470 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:14:44.480 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:14:44.480 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:15:50.088 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000033/group_3/index_3201_seed_7524_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 09:15:54.226 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:16:12.447 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:16:14.822 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:16:15.746 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:18:11.066 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.06s (1.92min) +2026-06-19 09:18:11.785 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:20:11.626 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:27:13.738 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 34, train_steps: 1360 +2026-06-19 09:27:14.444 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 09:27:14.598 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:27:14.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:27:14.754 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:28:21.346 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_3/index_3278_seed_22246_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 09:28:24.510 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:28:42.612 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:28:42.621 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:28:42.621 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:29:49.927 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_3/index_3278_seed_22247_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 09:29:53.144 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:30:11.308 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:30:13.695 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:30:15.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:32:08.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.81s (1.90min) +2026-06-19 09:32:09.592 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:34:09.779 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:41:12.401 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 35, train_steps: 1400 +2026-06-19 09:41:13.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 09:41:13.248 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:41:13.339 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:41:13.434 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:42:19.159 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000035/group_3/index_9481_seed_3105_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 09:42:24.330 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:42:42.852 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:42:42.862 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:42:42.862 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:43:48.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000035/group_3/index_9481_seed_3106_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 09:43:53.043 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:44:11.339 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:44:13.080 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:44:14.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:46:10.686 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 116.56s (1.94min) +2026-06-19 09:46:11.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:48:10.895 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:55:11.964 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 36, train_steps: 1440 +2026-06-19 09:55:12.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 09:55:12.834 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:55:12.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:55:13.047 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:56:19.328 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_3/index_4208_seed_15458_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 09:56:24.335 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:56:42.458 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:56:42.468 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:56:42.468 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:57:48.081 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_3/index_4208_seed_15459_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 09:57:52.438 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:58:10.329 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:58:12.646 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:58:13.739 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:00:08.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-19 10:00:09.671 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:02:10.006 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:09:11.706 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 37, train_steps: 1480 +2026-06-19 10:09:12.395 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 10:09:12.525 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:09:12.573 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:09:12.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:10:18.686 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_3/index_1630_seed_12739_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 10:10:21.917 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:10:39.876 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:10:39.885 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:10:39.885 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:11:47.572 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_3/index_1630_seed_12740_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 10:11:50.784 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:12:08.792 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:12:13.042 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:12:14.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:14:09.030 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.94s (1.92min) +2026-06-19 10:14:09.737 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:16:10.365 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:23:12.352 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 38, train_steps: 1520 +2026-06-19 10:23:13.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 10:23:13.222 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:23:13.308 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:23:13.410 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:24:19.486 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000038/group_3/index_4969_seed_12786_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 10:24:24.637 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:24:42.948 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:24:42.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:24:42.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:25:48.559 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000038/group_3/index_4969_seed_12787_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 10:25:53.451 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:26:11.964 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:26:12.667 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:26:14.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:28:09.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.16s (1.92min) +2026-06-19 10:28:09.954 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:30:10.743 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:37:11.951 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 39, train_steps: 1560 +2026-06-19 10:37:12.646 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 10:37:12.771 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:37:12.807 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:37:12.874 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:38:18.990 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000039/group_3/index_2742_seed_4808_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 10:38:23.135 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:38:41.040 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:38:41.047 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:38:41.051 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:41:17.590 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000039/group_3/index_2742_seed_4809_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 10:41:22.727 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:41:41.432 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:41:42.526 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:41:43.728 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:43:36.471 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.70s (1.88min) +2026-06-19 10:43:37.180 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:45:37.176 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:52:37.857 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 40, train_steps: 1600 +2026-06-19 10:53:14.046 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=40 (missing=0, unexpected=0) +2026-06-19 10:53:15.773 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 10:53:15.973 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:53:16.079 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:53:16.171 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:54:24.041 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_3/index_5614_seed_9025_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 10:54:29.016 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:54:47.408 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:54:47.417 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:54:47.417 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:55:52.543 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_3/index_5614_seed_9026_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 10:55:57.589 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:56:16.051 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:56:16.410 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:56:17.330 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:58:11.717 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.27s (1.90min) +2026-06-19 10:58:12.416 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:00:11.327 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:07:12.583 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 41, train_steps: 1640 +2026-06-19 11:07:13.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 11:07:13.479 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:07:13.596 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:07:13.703 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:08:19.682 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_3/index_4638_seed_7833_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 11:08:24.812 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:08:43.095 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:08:43.104 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:08:43.105 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:09:48.407 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_3/index_4638_seed_7834_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 11:09:53.584 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:10:12.029 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:10:12.668 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:10:14.057 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:12:08.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.28s (1.90min) +2026-06-19 11:12:09.087 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:14:08.758 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:21:10.888 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 42, train_steps: 1680 +2026-06-19 11:21:11.804 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 11:21:11.950 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:21:11.985 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:21:12.084 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:22:18.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000042/group_3/index_7389_seed_23067_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 11:22:22.553 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:22:40.470 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:22:40.479 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:22:40.479 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:23:47.524 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000042/group_3/index_7389_seed_23068_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 11:23:52.791 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:24:11.463 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:24:12.073 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:24:13.253 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:26:07.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.32s (1.91min) +2026-06-19 11:26:08.279 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:28:06.796 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:35:06.775 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 43, train_steps: 1720 +2026-06-19 11:35:07.464 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 11:35:07.669 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:35:07.790 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:35:07.886 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:36:13.588 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_3/index_9142_seed_17990_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 11:36:18.753 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:36:36.946 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:36:36.955 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:36:36.955 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:37:46.153 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_3/index_9142_seed_17991_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 11:37:51.199 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:38:09.758 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:38:10.299 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:38:11.341 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:40:05.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.38s (1.91min) +2026-06-19 11:40:06.548 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:42:03.525 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:49:04.435 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 44, train_steps: 1760 +2026-06-19 11:49:05.113 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 11:49:05.242 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:49:05.274 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:49:05.350 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:50:11.402 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000044/group_3/index_4033_seed_20686_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 11:50:16.510 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:50:34.578 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:50:34.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:50:34.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:51:41.583 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000044/group_3/index_4033_seed_20687_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 11:51:46.494 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:52:05.029 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:52:06.491 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:52:07.513 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:54:01.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.69s (1.89min) +2026-06-19 11:54:01.910 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:56:01.236 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:03:03.872 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 45, train_steps: 1800 +2026-06-19 12:03:04.557 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 12:03:04.806 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:03:04.927 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:03:05.010 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:04:10.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000045/group_3/index_7516_seed_8287_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 12:04:15.699 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:04:33.695 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:04:33.701 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:04:33.701 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:05:40.357 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000045/group_3/index_7516_seed_8288_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 12:05:43.604 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:06:01.757 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:06:05.347 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:06:06.267 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:08:01.417 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.05s (1.92min) +2026-06-19 12:08:02.101 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:10:01.381 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:17:03.592 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 46, train_steps: 1840 +2026-06-19 12:17:04.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 12:17:04.409 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:17:04.453 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:17:04.518 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:18:10.757 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_3/index_8539_seed_27548_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 12:18:13.976 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:18:33.288 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:18:33.297 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:18:33.298 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:19:40.636 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_3/index_8539_seed_27549_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 12:19:43.903 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:20:02.112 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:20:05.333 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:20:06.395 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:22:01.733 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.17s (1.92min) +2026-06-19 12:22:02.441 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:24:01.449 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:31:02.932 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 47, train_steps: 1880 +2026-06-19 12:31:03.606 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 12:31:03.698 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:31:03.747 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:31:03.795 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:32:09.784 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000047/group_3/index_6364_seed_15594_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 12:32:12.982 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:32:33.105 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:32:33.114 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:32:33.114 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:33:39.369 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000047/group_3/index_6364_seed_15595_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 12:33:44.528 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:34:02.793 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:34:03.182 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:34:04.357 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:35:58.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-19 12:35:58.982 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:37:55.891 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:44:57.681 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 48, train_steps: 1920 +2026-06-19 12:44:58.355 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 12:44:58.449 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:44:58.495 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:44:58.545 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:46:04.520 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_3/index_1205_seed_10006_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 12:46:08.837 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:46:26.550 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:46:26.559 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:46:26.561 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:47:34.651 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_3/index_1205_seed_10007_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 12:47:37.917 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:47:55.944 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:47:58.931 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:47:59.871 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:49:54.620 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-19 12:49:55.343 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:51:52.290 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:58:52.092 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 49, train_steps: 1960 +2026-06-19 12:58:52.783 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 12:58:52.954 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:58:53.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:58:53.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:59:59.177 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_3/index_2821_seed_24116_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 13:00:04.294 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:00:22.308 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:00:22.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:00:22.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:01:30.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_3/index_2821_seed_24117_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 13:01:34.609 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:01:52.762 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:01:54.370 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:01:55.296 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:03:48.446 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.04s (1.88min) +2026-06-19 13:03:49.148 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:05:45.465 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:12:45.277 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 50, train_steps: 2000 +2026-06-19 13:14:21.014 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=50 (missing=0, unexpected=0) +2026-06-19 13:14:23.203 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 13:14:23.390 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:14:23.476 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:14:23.572 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:15:31.592 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_3/index_569_seed_29349_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 13:15:36.736 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:15:54.750 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:15:54.757 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:15:54.758 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:17:00.995 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_3/index_569_seed_29350_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 13:17:05.931 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:17:24.204 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:17:25.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:17:26.218 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:19:20.081 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.84s (1.90min) +2026-06-19 13:19:20.807 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:21:17.524 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:28:18.470 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 51, train_steps: 2040 +2026-06-19 13:28:19.185 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 13:28:19.368 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:28:19.463 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:28:19.567 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:29:25.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000051/group_3/index_3935_seed_28027_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 13:29:29.563 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:29:47.631 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:29:47.636 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:29:47.637 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:30:54.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000051/group_3/index_3935_seed_28028_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 13:30:59.263 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:31:17.637 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:31:18.029 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:31:19.104 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:33:13.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.83s (1.90min) +2026-06-19 13:33:13.741 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:35:13.229 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:42:13.677 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 52, train_steps: 2080 +2026-06-19 13:42:14.277 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 13:42:14.482 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:42:14.593 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:42:14.686 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:43:21.426 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_3/index_3134_seed_22063_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 13:43:25.932 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:43:44.723 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:43:44.729 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:43:44.729 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:44:50.114 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_3/index_3134_seed_22064_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 13:44:55.237 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:45:14.377 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:45:15.287 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:45:16.415 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:47:10.481 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.03s (1.90min) +2026-06-19 13:47:11.196 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:49:11.264 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:56:12.604 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 53, train_steps: 2120 +2026-06-19 13:56:13.303 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 13:56:13.439 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:56:13.474 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:56:13.567 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:57:19.518 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_3/index_5672_seed_5224_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 13:57:24.442 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:57:42.619 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:57:42.628 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:57:42.628 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:58:48.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_3/index_5672_seed_5225_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 13:58:53.809 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:59:12.576 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:59:13.130 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:59:14.376 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:01:08.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.41s (1.91min) +2026-06-19 14:01:09.529 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:03:10.473 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:10:12.170 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 54, train_steps: 2160 +2026-06-19 14:10:12.881 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 14:10:13.099 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:10:13.227 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:10:13.321 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:11:19.260 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_3/index_5965_seed_21414_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 14:11:24.214 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:11:42.160 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:11:42.170 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:11:42.170 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:12:47.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_3/index_5965_seed_21415_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 14:12:52.828 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:13:10.962 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:13:12.866 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:13:14.068 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:15:08.493 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.42s (1.91min) +2026-06-19 14:15:09.218 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:17:09.305 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:24:11.905 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 55, train_steps: 2200 +2026-06-19 14:24:12.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 14:24:12.808 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:24:12.902 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:24:12.998 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:25:19.367 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_3/index_7301_seed_8751_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 14:25:24.481 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:25:42.998 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:25:43.006 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:25:43.006 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:26:48.288 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_3/index_7301_seed_8752_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 14:26:53.325 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:27:11.881 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:27:12.754 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:27:14.064 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:29:09.888 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.82s (1.93min) +2026-06-19 14:29:10.605 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:31:10.764 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:38:12.119 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 56, train_steps: 2240 +2026-06-19 14:38:12.812 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 14:38:13.021 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:38:13.158 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:38:13.238 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:39:19.501 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_3/index_5644_seed_26611_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 14:39:23.737 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:39:41.748 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:39:41.757 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:39:41.757 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:40:48.899 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_3/index_5644_seed_26612_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 14:40:54.154 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:41:12.879 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:41:13.092 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:41:14.034 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:43:08.781 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.66s (1.91min) +2026-06-19 14:43:09.475 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:45:09.717 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:52:11.847 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 57, train_steps: 2280 +2026-06-19 14:52:12.528 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 14:52:12.758 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:52:12.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:52:12.971 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:53:18.828 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_3/index_9871_seed_13578_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 14:53:22.067 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:53:40.266 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:53:40.275 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:53:40.275 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:54:48.427 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_3/index_9871_seed_13579_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 14:54:51.672 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:55:10.503 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:55:13.057 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:55:14.088 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:57:08.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.79s (1.91min) +2026-06-19 14:57:09.597 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:59:09.880 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:06:12.285 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 58, train_steps: 2320 +2026-06-19 15:06:12.985 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 15:06:13.144 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:06:13.198 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:06:13.312 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:07:19.979 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_3/index_9291_seed_12720_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 15:07:24.224 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:07:41.881 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:07:41.889 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:07:41.892 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:08:48.517 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_3/index_9291_seed_12721_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 15:08:51.732 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:09:09.763 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:09:13.062 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:09:14.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:11:08.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.71s (1.91min) +2026-06-19 15:11:09.498 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:13:09.811 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:20:12.057 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 59, train_steps: 2360 +2026-06-19 15:20:12.738 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 15:20:12.881 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:20:12.915 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:20:13.004 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:21:19.194 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_3/index_4620_seed_26159_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 15:21:24.468 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:21:42.685 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:21:42.697 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:21:42.698 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:22:47.973 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_3/index_4620_seed_26160_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 15:22:53.060 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:23:11.476 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:23:12.372 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:23:13.755 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:25:09.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.28s (1.92min) +2026-06-19 15:25:09.742 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:27:10.366 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:34:12.069 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 60, train_steps: 2400 +2026-06-19 15:34:48.333 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=60 (missing=0, unexpected=0) +2026-06-19 15:34:50.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 15:34:50.774 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:34:50.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:34:50.948 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:35:58.604 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_3/index_6239_seed_28840_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 15:36:03.658 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:36:20.460 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:36:20.466 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:36:20.466 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:37:28.403 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_3/index_6239_seed_28841_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 15:37:32.336 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:37:49.506 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:37:53.529 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:37:54.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:39:47.707 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-19 15:39:48.403 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:41:45.232 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:48:44.587 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 61, train_steps: 2440 +2026-06-19 15:48:45.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 15:48:45.502 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:48:45.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:48:45.707 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:49:51.412 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_3/index_8678_seed_2847_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 15:49:54.511 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:50:12.715 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:50:12.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:50:12.727 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:51:23.561 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_3/index_8678_seed_2848_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 15:51:26.855 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:51:45.963 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:51:47.609 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:51:48.476 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:53:41.681 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.04s (1.88min) +2026-06-19 15:53:42.400 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:55:39.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:02:39.022 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 62, train_steps: 2480 +2026-06-19 16:02:39.688 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 16:02:39.891 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:02:40.007 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:02:40.103 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:03:45.674 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000062/group_3/index_9032_seed_19736_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 16:03:50.871 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:04:09.280 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:04:09.289 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:04:09.289 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:05:18.833 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000062/group_3/index_9032_seed_19737_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 16:05:23.820 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:05:42.532 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:05:44.132 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:05:45.163 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:07:38.761 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.59s (1.89min) +2026-06-19 16:07:39.471 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:09:38.853 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:16:39.322 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 63, train_steps: 2520 +2026-06-19 16:16:40.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 16:16:40.143 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:16:40.187 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:16:40.233 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:17:46.423 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_3/index_2747_seed_23181_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 16:17:51.605 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:18:10.274 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:18:10.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:18:10.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:19:19.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_3/index_2747_seed_23182_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 16:19:25.034 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:19:43.677 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:19:45.197 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:19:46.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:21:39.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.36s (1.89min) +2026-06-19 16:21:40.303 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:23:38.881 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:30:40.928 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 64, train_steps: 2560 +2026-06-19 16:30:41.626 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 16:30:41.804 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:30:41.883 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:30:41.982 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:31:48.097 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_3/index_1465_seed_10760_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 16:31:53.257 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:32:11.238 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:32:11.248 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:32:11.248 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:33:21.266 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_3/index_1465_seed_10761_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 16:33:26.499 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:33:45.100 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:33:45.471 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:33:46.489 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:35:40.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.63s (1.89min) +2026-06-19 16:35:40.844 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:37:39.399 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:44:41.294 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 65, train_steps: 2600 +2026-06-19 16:44:41.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 16:44:42.125 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:44:42.195 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:44:42.322 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:45:48.392 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_3/index_2873_seed_22352_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 16:45:53.361 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:46:11.380 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:46:11.388 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:46:11.388 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:47:23.034 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_3/index_2873_seed_22353_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 16:47:26.253 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:47:44.429 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:47:47.482 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:47:48.401 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:49:42.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.72s (1.90min) +2026-06-19 16:49:42.958 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:51:40.034 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:58:41.279 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 66, train_steps: 2640 +2026-06-19 16:58:41.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 16:58:42.184 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:58:42.287 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:58:42.386 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:59:48.095 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_3/index_3117_seed_22816_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 16:59:51.351 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:00:09.525 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:00:09.534 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:00:09.534 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:01:19.405 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_3/index_3117_seed_22817_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 17:01:22.682 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:01:40.591 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:01:43.599 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:01:44.464 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:03:37.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.34s (1.89min) +2026-06-19 17:03:38.681 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:05:38.381 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:12:40.130 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 67, train_steps: 2680 +2026-06-19 17:12:40.909 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 17:12:41.308 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:12:41.486 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:12:41.586 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:13:47.567 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_3/index_182_seed_31397_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 17:13:50.799 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:14:08.618 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:14:08.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:14:08.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:15:20.691 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_3/index_182_seed_31398_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 17:15:24.271 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:15:42.516 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:15:44.529 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:15:45.443 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:17:38.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 17:17:39.679 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:19:39.112 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:26:39.568 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 68, train_steps: 2720 +2026-06-19 17:26:40.215 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 17:26:40.360 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:26:40.429 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:26:40.556 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:27:47.113 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_3/index_2416_seed_7593_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 17:27:50.387 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:28:08.551 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:28:08.562 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:28:08.565 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:29:20.124 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_3/index_2416_seed_7594_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 17:29:25.375 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:29:43.879 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:29:43.886 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:29:44.932 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:31:38.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.76s (1.90min) +2026-06-19 17:31:39.420 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:33:38.458 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:40:38.366 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 69, train_steps: 2760 +2026-06-19 17:40:39.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 17:40:39.198 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:40:39.279 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:40:39.408 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:41:45.142 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000069/group_3/index_7115_seed_31474_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 17:41:49.074 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:42:07.294 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:42:07.301 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:42:07.301 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:43:17.147 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000069/group_3/index_7115_seed_31475_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 17:43:22.446 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:43:41.059 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:43:41.069 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:43:42.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:45:35.854 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.55s (1.89min) +2026-06-19 17:45:36.569 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:47:37.119 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:54:38.581 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 70, train_steps: 2800 +2026-06-19 17:55:14.576 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=70 (missing=0, unexpected=0) +2026-06-19 17:55:16.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 17:55:17.131 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:55:17.255 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:55:17.336 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:56:25.327 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_3/index_4941_seed_28066_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 17:56:30.346 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:56:48.599 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:56:48.608 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:56:48.609 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:57:54.604 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_3/index_4941_seed_28067_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 17:57:58.805 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:58:17.084 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:58:18.228 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:58:19.385 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:00:13.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.58s (1.89min) +2026-06-19 18:00:13.837 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:02:13.850 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:09:13.897 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 71, train_steps: 2840 +2026-06-19 18:09:14.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 18:09:14.793 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:09:14.906 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:09:15.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:10:21.423 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_3/index_7642_seed_19504_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 18:10:24.732 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:10:42.845 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:10:42.855 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:10:42.855 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:11:51.593 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_3/index_7642_seed_19505_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 18:11:54.913 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:12:14.707 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:12:16.021 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:12:16.908 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:14:12.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.67s (1.93min) +2026-06-19 18:14:13.459 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:16:12.693 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:23:13.830 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 72, train_steps: 2880 +2026-06-19 18:23:14.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 18:23:14.635 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:23:14.683 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:23:14.761 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:24:21.936 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_3/index_9208_seed_19642_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 18:24:26.941 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:24:45.198 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:24:45.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:24:45.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:25:51.091 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_3/index_9208_seed_19643_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 18:25:55.573 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:26:14.700 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:26:14.711 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:26:15.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:28:10.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.45s (1.91min) +2026-06-19 18:28:11.145 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:30:10.988 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:37:12.791 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 73, train_steps: 2920 +2026-06-19 18:37:13.457 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 18:37:13.656 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:37:13.774 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:37:13.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:38:20.076 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_3/index_3844_seed_22674_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 18:38:23.308 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:38:41.171 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:38:41.181 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:38:41.181 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:39:50.057 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_3/index_3844_seed_22675_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 18:39:53.448 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:40:12.125 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:40:13.250 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:40:14.410 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:42:09.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.16s (1.92min) +2026-06-19 18:42:10.285 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:44:10.448 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:51:12.749 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 74, train_steps: 2960 +2026-06-19 18:51:13.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 18:51:13.610 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:51:13.658 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:51:13.775 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:52:20.289 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_3/index_8575_seed_30406_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 18:52:25.354 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:52:43.837 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:52:43.846 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:52:43.847 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:53:49.757 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_3/index_8575_seed_30407_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 18:53:54.356 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:54:12.477 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:54:13.982 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:54:15.057 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:56:10.312 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.21s (1.92min) +2026-06-19 18:56:11.031 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:58:11.441 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:05:14.019 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 75, train_steps: 3000 +2026-06-19 19:06:23.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 19:06:23.392 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:06:23.491 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:06:23.620 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:07:29.655 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000075/group_3/index_5118_seed_6889_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 19:07:32.869 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:07:50.592 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:07:50.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:07:50.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:08:58.852 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000075/group_3/index_5118_seed_6890_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 19:09:02.080 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:09:20.143 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:09:24.209 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:09:25.138 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:11:18.397 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.14s (1.89min) +2026-06-19 19:11:19.100 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:13:16.252 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:20:16.316 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 76, train_steps: 3040 +2026-06-19 19:20:16.989 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 19:20:17.131 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:20:17.171 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:20:17.260 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:21:23.951 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000076/group_3/index_7649_seed_5176_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 19:21:29.323 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:21:48.121 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:21:48.127 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:21:48.127 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:22:53.703 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000076/group_3/index_7649_seed_5177_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 19:22:57.906 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:23:16.113 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:23:18.455 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:23:19.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:25:13.441 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.66s (1.89min) +2026-06-19 19:25:14.145 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:27:12.733 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:34:13.275 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 77, train_steps: 3080 +2026-06-19 19:34:13.950 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 19:34:14.141 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:34:14.243 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:34:14.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:35:20.195 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_3/index_2635_seed_5906_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 19:35:23.421 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:35:41.288 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:35:41.302 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:35:41.303 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:36:49.136 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_3/index_2635_seed_5907_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 19:36:52.430 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:37:10.745 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:37:13.074 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:37:14.036 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:39:08.734 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.60s (1.91min) +2026-06-19 19:39:09.441 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:41:09.767 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:48:12.654 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 78, train_steps: 3120 +2026-06-19 19:48:13.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 19:48:13.549 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:48:13.637 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:48:13.755 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:49:20.006 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_3/index_4841_seed_21634_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 19:49:23.245 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:49:40.977 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:49:40.987 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:49:40.987 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:50:49.932 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_3/index_4841_seed_21635_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 19:50:53.221 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:51:11.517 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:51:13.323 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:51:14.373 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:53:09.267 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-19 19:53:09.976 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:55:09.985 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:02:13.178 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 79, train_steps: 3160 +2026-06-19 20:02:13.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 20:02:14.008 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:02:14.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:02:14.118 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:03:20.299 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000079/group_3/index_5452_seed_23553_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 20:03:25.515 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:03:43.480 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:03:43.489 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:03:43.489 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:04:49.622 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000079/group_3/index_5452_seed_23554_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 20:04:53.856 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:05:11.724 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:05:13.093 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:05:14.030 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:07:09.660 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.53s (1.93min) +2026-06-19 20:07:10.385 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:09:11.031 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:16:13.942 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 80, train_steps: 3200 +2026-06-19 20:16:51.626 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=80 (missing=0, unexpected=0) +2026-06-19 20:16:53.093 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 20:16:53.201 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:16:53.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:16:53.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:18:02.086 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_3/index_6428_seed_27396_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 20:18:07.385 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:18:25.642 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:18:25.651 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:18:25.651 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:19:30.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_3/index_6428_seed_27397_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 20:19:35.526 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:19:53.295 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:19:54.752 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:19:55.681 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:21:49.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-19 20:21:50.362 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:23:47.273 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:30:47.431 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 81, train_steps: 3240 +2026-06-19 20:30:48.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 20:30:48.273 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:30:48.316 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:30:48.394 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:31:54.612 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_3/index_3026_seed_22239_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 20:31:57.802 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:32:16.046 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:32:16.057 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:32:16.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:33:25.014 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_3/index_3026_seed_22240_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 20:33:30.310 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:33:49.151 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:33:49.273 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:33:50.330 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:35:44.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.70s (1.89min) +2026-06-19 20:35:44.752 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:37:41.677 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:44:40.009 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 82, train_steps: 3280 +2026-06-19 20:44:40.696 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 20:44:40.820 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:44:40.858 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:44:40.932 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:45:47.253 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_3/index_9647_seed_14010_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 20:45:52.183 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:46:10.360 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:46:10.374 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:46:10.377 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:47:20.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_3/index_9647_seed_14011_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 20:47:25.037 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:47:43.318 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:47:44.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:47:45.153 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:49:38.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.41s (1.89min) +2026-06-19 20:49:39.275 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:51:38.423 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:58:39.010 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 83, train_steps: 3320 +2026-06-19 20:58:39.739 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 20:58:39.900 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:58:39.990 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:58:40.060 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:59:45.924 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_3/index_3892_seed_25943_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 20:59:50.943 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:00:09.161 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:00:09.171 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:00:09.172 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:01:19.665 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_3/index_3892_seed_25944_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 21:01:24.639 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:01:42.939 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:01:44.304 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:01:45.335 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:03:38.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.28s (1.89min) +2026-06-19 21:03:39.339 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:05:38.818 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:12:38.926 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 84, train_steps: 3360 +2026-06-19 21:12:39.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 21:12:39.754 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:12:39.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:12:39.855 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:13:45.983 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_3/index_7963_seed_6345_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 21:13:51.003 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:14:09.439 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:14:09.449 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:14:09.449 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:15:19.653 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_3/index_7963_seed_6346_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 21:15:24.652 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:15:43.288 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:15:43.791 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:15:44.823 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:17:38.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.31s (1.89min) +2026-06-19 21:17:38.864 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:19:38.523 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:26:39.498 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 85, train_steps: 3400 +2026-06-19 21:26:40.167 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 21:26:40.306 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:26:40.341 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:26:40.433 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:27:46.945 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_3/index_3775_seed_15608_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 21:27:51.274 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:28:09.005 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:28:09.014 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:28:09.014 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:29:20.139 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_3/index_3775_seed_15609_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 21:29:25.341 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:29:43.396 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:29:43.697 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:29:44.634 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:31:38.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-19 21:31:38.988 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:33:38.754 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:40:38.649 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 86, train_steps: 3440 +2026-06-19 21:40:39.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 21:40:39.529 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:40:39.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:40:39.749 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:41:46.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_3/index_9760_seed_12813_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 21:41:51.207 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:42:09.489 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:42:09.499 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:42:09.499 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:43:18.822 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_3/index_9760_seed_12814_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 21:43:23.143 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:43:41.255 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:43:42.867 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:43:43.759 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:45:37.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 21:45:37.987 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:47:37.707 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:54:39.053 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 87, train_steps: 3480 +2026-06-19 21:54:39.743 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 21:54:39.913 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:54:39.985 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:54:40.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:55:45.842 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000087/group_3/index_9520_seed_17645_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 21:55:49.062 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:56:07.199 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:56:07.209 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:56:07.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:57:19.022 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000087/group_3/index_9520_seed_17646_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 21:57:24.277 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:57:42.956 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:57:43.169 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:57:44.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:59:37.604 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.39s (1.89min) +2026-06-19 21:59:38.317 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:01:38.554 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:08:39.992 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 88, train_steps: 3520 +2026-06-19 22:08:40.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 22:08:40.792 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:08:40.820 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:08:40.868 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:09:46.912 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_3/index_3176_seed_13267_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 22:09:51.150 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:10:09.131 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:10:09.141 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:10:09.141 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:11:19.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_3/index_3176_seed_13268_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 22:11:24.826 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:11:43.744 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:11:43.755 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:11:44.807 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:13:38.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.57s (1.89min) +2026-06-19 22:13:39.104 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:15:38.304 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:22:38.700 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 89, train_steps: 3560 +2026-06-19 22:22:39.425 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 22:22:39.627 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:22:39.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:22:39.895 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:23:45.713 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_3/index_3312_seed_7891_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 22:23:50.328 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:24:08.061 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:24:08.070 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:24:08.070 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:25:19.682 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_3/index_3312_seed_7892_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 22:25:22.979 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:25:41.266 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:25:43.683 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:25:44.679 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:27:38.162 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 22:27:38.886 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:29:38.492 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:36:39.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 90, train_steps: 3600 +2026-06-19 22:37:15.617 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=90 (missing=0, unexpected=0) +2026-06-19 22:37:17.307 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 22:37:17.456 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:37:17.497 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:37:17.612 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:38:24.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_3/index_2205_seed_14671_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 22:38:30.233 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:38:48.680 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:38:48.686 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:38:48.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:39:55.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_3/index_2205_seed_14672_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 22:39:59.381 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:40:17.914 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:40:19.241 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:40:20.160 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:42:14.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.70s (1.89min) +2026-06-19 22:42:14.864 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:44:13.686 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:51:14.859 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 91, train_steps: 3640 +2026-06-19 22:51:15.700 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 22:51:15.898 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:51:16.000 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:51:16.094 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:52:23.116 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000091/group_3/index_2187_seed_3444_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 22:52:26.404 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:52:44.339 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:52:44.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:52:44.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:53:52.036 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000091/group_3/index_2187_seed_3445_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 22:53:55.391 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:54:14.388 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:54:15.937 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:54:16.829 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:56:12.672 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.67s (1.93min) +2026-06-19 22:56:13.388 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:58:13.109 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:05:14.271 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 92, train_steps: 3680 +2026-06-19 23:05:15.117 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 23:05:15.284 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:05:15.380 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:05:15.501 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:06:22.148 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000092/group_3/index_123_seed_21724_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 23:06:25.391 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:06:43.712 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:06:43.722 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:06:43.722 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:07:51.247 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000092/group_3/index_123_seed_21725_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 23:07:56.493 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:08:16.051 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:08:16.058 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:08:17.114 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:10:11.751 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.62s (1.91min) +2026-06-19 23:10:12.471 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:12:11.281 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:19:13.488 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 93, train_steps: 3720 +2026-06-19 23:19:14.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 23:19:14.243 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:19:14.286 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:19:14.330 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:20:21.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_3/index_8935_seed_26027_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 23:20:26.035 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:20:45.022 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:20:45.028 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:20:45.028 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:21:50.068 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_3/index_8935_seed_26028_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 23:21:54.961 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:22:13.150 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:22:14.354 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:22:15.604 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:24:10.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.33s (1.92min) +2026-06-19 23:24:11.707 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:26:11.285 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:33:13.941 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 94, train_steps: 3760 +2026-06-19 23:33:14.641 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 23:33:14.824 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:33:14.932 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:33:15.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:34:21.929 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_3/index_7167_seed_32249_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 23:34:27.230 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:34:46.427 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:34:46.434 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:34:46.434 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:35:51.538 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_3/index_7167_seed_32250_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 23:35:56.440 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:36:16.128 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:36:16.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:36:17.148 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:38:12.062 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.89s (1.91min) +2026-06-19 23:38:12.784 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:40:11.321 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:47:11.877 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 95, train_steps: 3800 +2026-06-19 23:47:12.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-19 23:47:12.751 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:47:12.819 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:47:12.918 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:48:18.726 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_3/index_6580_seed_12489_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-19 23:48:23.551 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:48:42.171 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:48:42.181 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:48:42.184 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:49:48.034 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_3/index_6580_seed_12490_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-19 23:49:52.513 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:50:10.552 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:50:12.004 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:50:13.169 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:52:08.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.82s (1.91min) +2026-06-19 23:52:08.791 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:54:07.675 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:01:09.779 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 96, train_steps: 3840 +2026-06-20 00:01:10.476 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 00:01:10.697 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:01:10.940 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:01:11.019 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:02:17.043 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_3/index_4199_seed_1328_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 00:02:20.286 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:02:38.512 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:02:38.520 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:02:38.522 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:03:47.445 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_3/index_4199_seed_1329_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 00:03:52.797 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:04:11.509 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:04:11.783 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:04:12.972 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:06:07.928 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 00:06:08.644 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:08:07.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:15:08.506 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 97, train_steps: 3880 +2026-06-20 00:15:09.218 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 00:15:09.318 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:15:09.361 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:15:09.408 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:16:15.288 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_3/index_6427_seed_24448_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 00:16:19.594 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:16:37.713 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:16:37.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:16:37.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:17:47.345 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_3/index_6427_seed_24449_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 00:17:52.579 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:18:11.181 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:18:11.238 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:18:12.367 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:20:07.671 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.20s (1.92min) +2026-06-20 00:20:08.400 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:22:07.125 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:29:07.006 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 98, train_steps: 3920 +2026-06-20 00:29:07.693 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 00:29:07.916 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:29:08.042 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:29:08.137 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:30:13.881 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000098/group_3/index_1814_seed_20611_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 00:30:18.976 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:30:37.831 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:30:37.839 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:30:37.842 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:31:46.191 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000098/group_3/index_1814_seed_20612_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 00:31:50.281 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:32:08.662 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:32:10.419 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:32:11.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:34:06.389 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 00:34:07.118 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:36:03.849 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:43:05.348 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 99, train_steps: 3960 +2026-06-20 00:43:06.025 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 00:43:06.215 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:43:06.344 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:43:06.426 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:44:12.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_3/index_9236_seed_1410_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 00:44:18.026 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:44:37.135 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:44:37.141 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:44:37.142 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:45:46.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_3/index_9236_seed_1411_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 00:45:51.282 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:46:09.887 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:46:09.899 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:46:10.955 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:48:05.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.75s (1.91min) +2026-06-20 00:48:06.440 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:50:03.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:57:04.804 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 100, train_steps: 4000 +2026-06-20 00:58:47.819 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=100 (missing=0, unexpected=0) +2026-06-20 00:58:49.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 00:58:49.818 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:58:49.927 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:58:50.021 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:59:57.261 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_3/index_7068_seed_29951_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 01:00:01.150 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:00:18.049 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:00:18.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:00:18.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:01:26.933 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_3/index_7068_seed_29952_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 01:01:32.182 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:01:49.328 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:01:50.303 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:01:51.338 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:03:44.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-20 01:03:45.434 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:05:42.221 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:12:39.874 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 101, train_steps: 4040 +2026-06-20 01:12:40.569 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 01:12:40.783 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:12:40.892 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:12:40.950 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:13:47.250 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_3/index_6062_seed_24673_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 01:13:52.005 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:14:09.534 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:14:09.544 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:14:09.546 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:15:19.961 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_3/index_6062_seed_24674_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 01:15:23.736 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:15:40.926 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:15:43.499 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:15:44.448 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:17:38.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.62s (1.89min) +2026-06-20 01:17:38.931 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:19:38.322 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:26:37.818 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 102, train_steps: 4080 +2026-06-20 01:26:38.504 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 01:26:38.665 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:26:38.738 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:26:38.847 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:27:44.845 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_3/index_6737_seed_18480_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 01:27:49.902 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:28:08.594 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:28:08.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:28:08.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:29:17.393 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_3/index_6737_seed_18481_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 01:29:22.582 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:29:39.860 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:29:41.183 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:29:42.526 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:31:36.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-20 01:31:37.098 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:33:37.226 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:40:38.545 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 103, train_steps: 4120 +2026-06-20 01:40:39.244 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 01:40:39.415 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:40:39.496 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:40:39.614 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:41:45.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000103/group_3/index_5503_seed_7081_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 01:41:49.554 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:42:07.161 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:42:07.170 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:42:07.172 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:43:19.186 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000103/group_3/index_5503_seed_7082_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 01:43:22.367 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:43:39.352 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:43:43.334 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:43:44.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:45:37.767 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-20 01:45:38.487 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:47:38.263 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:54:37.791 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 104, train_steps: 4160 +2026-06-20 01:54:38.703 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 01:54:38.894 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:54:39.020 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:54:39.103 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:55:45.174 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_3/index_2934_seed_6322_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 01:55:49.860 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:56:07.810 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:56:07.820 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:56:07.820 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:57:18.841 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_3/index_2934_seed_6323_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 01:57:21.966 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:57:39.299 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:57:42.889 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:57:43.784 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:59:37.314 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-20 01:59:38.031 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:01:38.018 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:08:38.450 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 105, train_steps: 4200 +2026-06-20 02:08:39.147 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 02:08:39.347 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:08:39.448 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:08:39.542 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:09:45.357 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000105/group_3/index_9448_seed_15080_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 02:09:50.378 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:10:08.331 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:10:08.338 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:10:08.338 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:11:18.808 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000105/group_3/index_9448_seed_15081_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 02:11:22.638 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:11:39.946 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:11:43.945 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:11:44.875 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:13:38.217 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 02:13:38.925 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:15:38.265 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:22:37.846 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 106, train_steps: 4240 +2026-06-20 02:22:38.564 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 02:22:38.758 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:22:38.877 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:22:38.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:23:44.824 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000106/group_3/index_5728_seed_22200_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 02:23:47.942 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:24:05.019 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:24:05.029 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:24:05.029 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:25:16.825 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000106/group_3/index_5728_seed_22201_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 02:25:21.863 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:25:39.437 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:25:40.847 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:25:41.888 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:27:35.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.90s (1.90min) +2026-06-20 02:27:36.582 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:29:36.811 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:36:37.096 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 107, train_steps: 4280 +2026-06-20 02:36:37.801 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 02:36:37.968 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:36:38.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:36:38.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:37:44.061 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_3/index_6281_seed_12532_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 02:37:48.881 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:38:06.639 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:38:06.646 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:38:06.646 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:39:15.221 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_3/index_6281_seed_12533_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 02:39:19.307 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:39:37.383 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:39:39.229 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:39:40.448 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:41:35.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 02:41:36.071 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:43:35.145 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:50:34.999 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 108, train_steps: 4320 +2026-06-20 02:50:35.726 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 02:50:35.866 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:50:35.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:50:35.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:51:42.066 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_3/index_6371_seed_15796_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 02:51:45.766 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:52:03.155 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:52:03.164 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:52:03.164 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:53:13.363 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_3/index_6371_seed_15797_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 02:53:16.676 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:53:34.237 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:53:37.574 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:53:38.450 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:55:33.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.45s (1.91min) +2026-06-20 02:55:33.972 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:57:30.647 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:04:30.699 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 109, train_steps: 4360 +2026-06-20 03:04:31.406 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 03:04:31.605 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:04:31.728 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:04:31.805 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:05:37.599 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000109/group_3/index_459_seed_30017_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 03:05:42.774 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:06:00.610 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:06:00.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:06:00.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:07:09.623 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000109/group_3/index_459_seed_30018_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 03:07:14.715 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:07:33.160 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:07:33.407 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:07:34.441 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:09:29.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-20 03:09:30.014 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:11:26.575 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:18:25.335 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 110, train_steps: 4400 +2026-06-20 03:19:01.481 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=110 (missing=0, unexpected=0) +2026-06-20 03:19:03.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 03:19:03.496 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:19:03.583 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:19:03.662 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:20:11.496 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_3/index_8430_seed_13687_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 03:20:15.715 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:20:32.515 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:20:32.524 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:20:32.524 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:21:43.286 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_3/index_8430_seed_13688_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 03:21:48.253 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:22:05.391 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:22:07.524 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:22:08.550 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:24:03.966 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.41s (1.92min) +2026-06-20 03:24:04.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:26:02.473 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:33:05.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 111, train_steps: 4440 +2026-06-20 03:33:06.074 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 03:33:06.269 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:33:06.369 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:33:06.451 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:34:12.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_3/index_2012_seed_3245_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 03:34:17.378 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:34:33.961 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:34:33.970 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:34:33.970 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:35:46.394 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_3/index_2012_seed_3246_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 03:35:50.287 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:36:07.032 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:36:10.098 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:36:11.047 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:38:06.021 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 03:38:06.732 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:40:04.481 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:47:04.313 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 112, train_steps: 4480 +2026-06-20 03:47:05.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 03:47:05.251 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:47:05.385 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:47:05.458 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:48:11.226 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_3/index_449_seed_5215_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 03:48:14.655 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:48:32.128 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:48:32.139 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:48:32.139 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:49:40.829 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_3/index_449_seed_5216_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 03:49:44.450 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:50:00.620 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:50:04.983 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:50:06.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:52:00.943 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-20 03:52:01.657 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:54:00.962 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:01:04.009 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 113, train_steps: 4520 +2026-06-20 04:01:04.875 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 04:01:05.042 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:01:05.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:01:05.243 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:02:11.034 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000113/group_3/index_2943_seed_3552_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 04:02:14.334 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:02:31.823 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:02:31.832 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:02:31.833 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:03:42.127 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000113/group_3/index_2943_seed_3553_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 04:03:45.317 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:04:02.002 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:04:06.611 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:04:07.490 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:06:02.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.15s (1.92min) +2026-06-20 04:06:03.532 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:08:01.582 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:15:03.784 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 114, train_steps: 4560 +2026-06-20 04:15:04.473 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 04:15:04.601 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:15:04.650 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:15:04.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:16:10.596 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000114/group_3/index_680_seed_32733_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 04:16:13.704 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:16:31.894 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:16:31.904 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:16:31.904 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:17:40.528 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000114/group_3/index_680_seed_32734_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 04:17:45.490 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:18:02.502 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:18:04.152 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:18:05.359 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:20:00.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.72s (1.91min) +2026-06-20 04:20:00.894 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:22:00.100 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:29:01.469 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 115, train_steps: 4600 +2026-06-20 04:29:02.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 04:29:02.329 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:29:02.368 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:29:02.461 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:30:08.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_3/index_9597_seed_6277_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 04:30:13.056 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:30:31.727 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:30:31.736 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:30:31.736 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:31:38.927 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_3/index_9597_seed_6278_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 04:31:43.712 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:32:00.590 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:32:02.494 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:32:03.524 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:33:58.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.33s (1.91min) +2026-06-20 04:33:58.775 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:35:55.626 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:42:56.470 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 116, train_steps: 4640 +2026-06-20 04:42:57.158 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 04:42:57.258 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:42:57.291 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:42:57.323 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:44:03.518 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_3/index_1064_seed_6600_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 04:44:07.250 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:44:23.784 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:44:23.794 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:44:23.794 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:45:34.130 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_3/index_1064_seed_6601_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 04:45:39.120 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:45:55.838 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:45:58.381 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:45:59.445 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:47:53.698 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.25s (1.90min) +2026-06-20 04:47:54.407 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:49:50.888 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:56:50.232 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 117, train_steps: 4680 +2026-06-20 04:56:50.953 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 04:56:51.112 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:56:51.169 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:56:51.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:57:57.287 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_3/index_975_seed_31509_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 04:58:02.013 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:58:18.404 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:58:18.410 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:58:18.410 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:59:28.069 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_3/index_975_seed_31510_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 04:59:32.018 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:59:49.461 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:59:51.798 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:59:52.747 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:01:46.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.49s (1.89min) +2026-06-20 05:01:47.078 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:03:44.145 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:10:42.594 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 118, train_steps: 4720 +2026-06-20 05:10:43.298 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 05:10:43.502 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:10:43.631 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:10:43.727 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:11:49.678 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000118/group_3/index_5403_seed_26840_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 05:11:52.920 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:12:09.323 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:12:09.332 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:12:09.333 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:13:22.985 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000118/group_3/index_5403_seed_26841_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 05:13:26.429 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:13:43.243 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:13:46.957 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:13:47.845 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:15:41.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.26s (1.89min) +2026-06-20 05:15:42.000 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:17:38.993 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:24:37.904 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 119, train_steps: 4760 +2026-06-20 05:24:38.623 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 05:24:38.796 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:24:38.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:24:38.982 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:25:45.029 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_3/index_706_seed_28512_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 05:25:48.189 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:26:04.386 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:26:04.395 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:26:04.395 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:27:15.139 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_3/index_706_seed_28513_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 05:27:18.383 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:27:34.788 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:27:40.119 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:27:41.164 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:29:34.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-20 05:29:35.559 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:31:33.104 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:38:32.152 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 120, train_steps: 4800 +2026-06-20 05:39:08.936 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=120 (missing=0, unexpected=0) +2026-06-20 05:39:10.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 05:39:11.074 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:39:11.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:39:11.265 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:40:18.874 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000120/group_3/index_9348_seed_19065_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 05:40:22.023 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:40:38.704 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:40:38.714 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:40:38.714 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:41:48.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000120/group_3/index_9348_seed_19066_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 05:41:53.636 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:42:10.537 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:42:12.566 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:42:13.753 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:44:08.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.96s (1.92min) +2026-06-20 05:44:09.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:46:10.087 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:53:12.068 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 121, train_steps: 4840 +2026-06-20 05:53:12.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 05:53:12.914 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:53:12.947 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:53:13.035 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:54:19.497 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_3/index_2407_seed_1105_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 05:54:23.472 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:54:39.642 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:54:39.651 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:54:39.652 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:55:48.426 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_3/index_2407_seed_1106_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 05:55:53.412 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:56:10.327 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:56:13.156 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:56:14.380 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:58:09.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.97s (1.92min) +2026-06-20 05:58:10.076 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:00:10.576 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:07:11.956 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 122, train_steps: 4880 +2026-06-20 06:07:12.650 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 06:07:12.863 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:07:12.972 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:07:13.035 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:08:18.970 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_3/index_1748_seed_3607_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 06:08:23.715 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:08:40.342 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:08:40.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:08:40.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:09:48.680 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_3/index_1748_seed_3608_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 06:09:52.543 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:10:09.476 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:10:13.195 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:10:14.335 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:12:08.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.25s (1.90min) +2026-06-20 06:12:09.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:14:10.578 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:21:11.867 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 123, train_steps: 4920 +2026-06-20 06:21:12.561 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 06:21:12.680 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:21:12.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:21:12.787 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:22:18.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_3/index_2636_seed_5398_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 06:22:21.935 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:22:38.214 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:22:38.223 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:22:38.223 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:23:48.228 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_3/index_2636_seed_5399_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 06:23:51.474 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:24:07.875 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:24:12.071 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:24:13.155 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:26:08.788 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.47s (1.92min) +2026-06-20 06:26:09.511 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:28:09.833 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:35:11.852 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 124, train_steps: 4960 +2026-06-20 06:35:12.539 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 06:35:12.725 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:35:12.812 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:35:12.912 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:36:18.910 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_3/index_575_seed_19064_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 06:36:22.101 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:36:38.604 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:36:38.614 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:36:38.614 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:37:48.097 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_3/index_575_seed_19065_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 06:37:53.024 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:38:10.184 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:38:12.129 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:38:13.265 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:40:08.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.44s (1.92min) +2026-06-20 06:40:09.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:42:09.772 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:49:11.648 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 125, train_steps: 5000 +2026-06-20 06:50:15.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 06:50:15.355 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:50:15.440 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:50:15.564 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:51:22.178 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_3/index_2031_seed_13665_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 06:51:25.955 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:51:42.548 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:51:42.557 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:51:42.557 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:52:51.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_3/index_2031_seed_13666_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 06:52:56.288 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:53:13.477 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:53:15.926 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:53:16.959 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:55:11.654 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.68s (1.91min) +2026-06-20 06:55:12.372 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:57:11.763 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:04:11.940 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 126, train_steps: 5040 +2026-06-20 07:04:12.632 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 07:04:12.811 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:04:12.892 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:04:12.971 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:05:19.235 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_3/index_2740_seed_7264_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 07:05:23.272 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:05:39.883 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:05:39.893 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:05:39.893 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:06:49.307 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_3/index_2740_seed_7265_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 07:06:52.624 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:07:09.166 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:07:13.988 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:07:15.339 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:09:10.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.08s (1.92min) +2026-06-20 07:09:11.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:11:11.163 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:18:11.933 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 127, train_steps: 5080 +2026-06-20 07:18:12.629 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 07:18:12.777 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:18:12.811 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:18:12.899 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:19:18.840 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_3/index_2923_seed_21221_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 07:19:21.967 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:19:38.426 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:19:38.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:19:38.437 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:20:48.855 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_3/index_2923_seed_21222_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 07:20:53.799 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:21:10.695 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:21:13.050 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:21:14.093 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:23:09.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.50s (1.93min) +2026-06-20 07:23:10.327 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:25:10.158 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:32:11.175 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 128, train_steps: 5120 +2026-06-20 07:32:11.883 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 07:32:12.045 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:32:12.124 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:32:12.239 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:33:17.951 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_3/index_8940_seed_1797_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 07:33:21.713 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:33:38.094 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:33:38.103 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:33:38.104 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:34:48.713 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_3/index_8940_seed_1798_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 07:34:53.546 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:35:10.599 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:35:13.016 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:35:14.081 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:37:08.641 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.56s (1.91min) +2026-06-20 07:37:09.373 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:39:09.930 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:46:11.997 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 129, train_steps: 5160 +2026-06-20 07:46:12.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 07:46:12.858 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:46:12.902 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:46:12.999 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:47:19.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_3/index_337_seed_28234_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 07:47:22.811 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:47:39.060 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:47:39.071 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:47:39.071 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:48:48.892 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_3/index_337_seed_28235_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 07:48:51.891 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:49:08.415 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:49:13.070 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:49:14.051 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:51:09.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-20 07:51:09.758 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:53:10.302 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:00:11.800 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 130, train_steps: 5200 +2026-06-20 08:00:49.231 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=130 (missing=0, unexpected=0) +2026-06-20 08:00:51.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 08:00:51.396 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:00:51.479 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:00:51.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:02:00.282 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_3/index_6033_seed_23007_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 08:02:04.038 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:02:21.141 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:02:21.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:02:21.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:03:28.901 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_3/index_6033_seed_23008_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 08:03:31.997 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:03:48.866 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:03:52.509 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:03:53.393 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:05:47.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.72s (1.90min) +2026-06-20 08:05:48.018 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:07:44.842 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:14:45.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 131, train_steps: 5240 +2026-06-20 08:14:46.025 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 08:14:46.164 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:14:46.204 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:14:46.297 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:15:52.594 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_3/index_3024_seed_438_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 08:15:57.364 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:16:15.306 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:16:15.316 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:16:15.316 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:17:23.409 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_3/index_3024_seed_439_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 08:17:27.154 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:17:43.748 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:17:46.990 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:17:47.928 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:19:41.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.38s (1.89min) +2026-06-20 08:19:42.139 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:21:39.467 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:28:37.833 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 132, train_steps: 5280 +2026-06-20 08:28:38.521 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 08:28:38.708 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:28:38.803 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:28:38.896 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:29:45.164 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000132/group_3/index_2125_seed_13198_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 08:29:48.158 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:30:04.588 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:30:04.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:30:04.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:31:15.556 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000132/group_3/index_2125_seed_13199_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 08:31:18.670 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:31:35.177 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:31:40.849 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:31:41.739 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:33:35.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.44s (1.89min) +2026-06-20 08:33:36.119 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:35:35.261 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:42:34.338 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 133, train_steps: 5320 +2026-06-20 08:42:35.020 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 08:42:35.234 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:42:35.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:42:35.447 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:43:41.378 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_3/index_3667_seed_25635_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 08:43:44.471 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:44:00.715 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:44:00.725 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:44:00.725 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:45:11.727 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_3/index_3667_seed_25636_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 08:45:16.676 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:45:33.744 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:45:35.888 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:45:36.928 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:47:32.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.77s (1.93min) +2026-06-20 08:47:33.449 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:49:30.691 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:56:29.438 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 134, train_steps: 5360 +2026-06-20 08:56:30.114 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 08:56:30.282 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:56:30.359 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:56:30.463 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:57:36.494 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000134/group_3/index_3608_seed_17477_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 08:57:41.182 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:57:57.970 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:57:57.981 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:57:57.984 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:59:07.488 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000134/group_3/index_3608_seed_17478_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 08:59:11.237 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:59:27.702 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:59:32.041 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:59:32.985 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:01:26.319 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 09:01:27.048 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:03:23.808 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:10:22.991 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 135, train_steps: 5400 +2026-06-20 09:10:23.674 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 09:10:23.800 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:10:23.839 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:10:23.913 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:11:30.025 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000135/group_3/index_125_seed_19839_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 09:11:34.821 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:11:51.961 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:11:51.969 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:11:51.969 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:13:00.297 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000135/group_3/index_125_seed_19840_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 09:13:04.943 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:13:21.885 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:13:23.933 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:13:24.964 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:15:18.588 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.61s (1.89min) +2026-06-20 09:15:19.310 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:17:17.101 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:24:16.802 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 136, train_steps: 5440 +2026-06-20 09:24:17.533 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 09:24:17.693 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:24:17.751 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:24:17.855 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:25:24.159 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_3/index_3187_seed_3088_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 09:25:28.187 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:25:45.090 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:25:45.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:25:45.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:26:54.049 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_3/index_3187_seed_3089_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 09:26:57.186 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:27:14.699 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:27:18.656 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:27:19.539 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:29:12.959 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.15s (1.89min) +2026-06-20 09:29:13.686 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:31:13.175 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:38:13.219 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 137, train_steps: 5480 +2026-06-20 09:38:13.938 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 09:38:14.125 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:38:14.220 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:38:14.316 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:39:20.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000137/group_3/index_6818_seed_9736_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 09:39:24.992 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:39:41.853 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:39:41.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:39:41.864 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:40:50.180 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000137/group_3/index_6818_seed_9737_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 09:40:54.181 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:41:11.156 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:41:14.371 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:41:15.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:43:09.522 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-20 09:43:10.250 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:45:10.502 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:52:12.164 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 138, train_steps: 5520 +2026-06-20 09:52:12.868 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 09:52:13.073 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:52:13.177 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:52:13.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:53:19.193 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_3/index_5685_seed_10522_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 09:53:24.014 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:53:40.969 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:53:40.978 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:53:40.978 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:54:49.411 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_3/index_5685_seed_10523_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 09:54:53.206 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:55:10.247 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:55:13.383 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:55:14.700 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:57:10.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.46s (1.92min) +2026-06-20 09:57:10.921 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:59:11.149 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:06:12.680 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 139, train_steps: 5560 +2026-06-20 10:06:13.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 10:06:13.523 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:06:13.576 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:06:13.678 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:07:19.759 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_3/index_7883_seed_29079_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 10:07:24.597 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:07:41.594 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:07:41.605 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:07:41.605 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:08:50.055 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_3/index_7883_seed_29080_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 10:08:53.862 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:09:10.648 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:09:13.265 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:09:14.209 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:11:08.675 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.20s (1.90min) +2026-06-20 10:11:09.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:13:09.779 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:20:12.689 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 140, train_steps: 5600 +2026-06-20 10:20:49.511 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=140 (missing=0, unexpected=0) +2026-06-20 10:20:51.812 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 10:20:51.891 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:20:51.904 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:20:51.992 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:21:59.867 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000140/group_3/index_5122_seed_19228_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 10:22:02.947 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:22:20.027 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:22:20.037 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:22:20.037 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:23:28.948 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000140/group_3/index_5122_seed_19229_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 10:23:33.819 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:23:51.069 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:23:53.040 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:23:54.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:25:47.600 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-20 10:25:48.314 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:27:45.364 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:34:44.003 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 141, train_steps: 5640 +2026-06-20 10:34:44.689 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 10:34:44.857 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:34:44.933 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:34:45.041 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:35:51.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_3/index_7692_seed_790_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 10:35:55.300 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:36:13.035 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:36:13.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:36:13.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:37:23.049 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_3/index_7692_seed_791_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 10:37:26.089 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:37:44.277 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:37:46.846 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:37:47.750 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:39:41.415 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.51s (1.89min) +2026-06-20 10:39:42.148 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:41:39.460 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:48:38.468 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 142, train_steps: 5680 +2026-06-20 10:48:39.159 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 10:48:39.340 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:48:39.434 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:48:39.514 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:49:45.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000142/group_3/index_7823_seed_18305_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 10:49:50.246 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:50:06.975 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:50:06.984 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:50:06.985 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:51:19.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000142/group_3/index_7823_seed_18306_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 10:51:22.960 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:51:39.743 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:51:43.153 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:51:44.047 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:53:37.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-20 10:53:37.851 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:55:37.992 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:02:37.922 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 143, train_steps: 5720 +2026-06-20 11:02:38.632 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 11:02:38.797 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:02:38.875 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:02:38.971 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:03:45.141 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000143/group_3/index_4465_seed_33422_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 11:03:48.252 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:04:04.474 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:04:04.483 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:04:04.483 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:05:18.241 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000143/group_3/index_4465_seed_33423_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 11:05:23.178 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:05:40.209 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:05:42.169 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:05:43.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:07:36.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 11:07:37.312 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:09:36.843 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:16:37.102 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 144, train_steps: 5760 +2026-06-20 11:16:37.800 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 11:16:37.915 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:16:37.951 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:16:38.020 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:17:43.920 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_3/index_1459_seed_28584_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 11:17:47.800 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:18:04.420 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:18:04.430 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:18:04.430 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:19:13.157 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_3/index_1459_seed_28585_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 11:19:16.455 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:19:33.038 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:19:38.707 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:19:39.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:21:33.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.47s (1.89min) +2026-06-20 11:21:34.087 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:23:30.885 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:30:30.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 145, train_steps: 5800 +2026-06-20 11:30:31.096 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 11:30:31.226 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:30:31.265 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:30:31.339 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:31:37.229 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000145/group_3/index_8036_seed_20043_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 11:31:42.072 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:31:58.720 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:31:58.727 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:31:58.727 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:33:08.942 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000145/group_3/index_8036_seed_20044_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 11:33:13.809 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:33:32.100 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:33:33.208 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:33:34.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:35:28.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.89s (1.90min) +2026-06-20 11:35:28.901 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:37:25.685 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:44:23.924 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 146, train_steps: 5840 +2026-06-20 11:44:24.629 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 11:44:24.806 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:44:24.910 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:44:25.003 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:45:30.926 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_3/index_8288_seed_18646_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 11:45:35.037 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:45:51.714 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:45:51.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:45:51.727 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:46:59.109 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_3/index_8288_seed_18647_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 11:47:04.061 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:47:20.782 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:47:23.247 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:47:24.135 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:49:19.134 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-20 11:49:19.856 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:51:17.038 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:58:15.937 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 147, train_steps: 5880 +2026-06-20 11:58:16.619 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 11:58:16.819 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:58:16.947 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:58:17.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:59:23.805 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_3/index_4655_seed_14679_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 11:59:28.656 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:59:45.434 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:59:45.445 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:59:45.445 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:00:52.813 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_3/index_4655_seed_14680_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 12:00:56.680 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:01:14.697 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:01:16.232 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:01:17.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:03:12.586 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.29s (1.92min) +2026-06-20 12:03:13.293 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:05:12.904 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:12:12.707 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 148, train_steps: 5920 +2026-06-20 12:12:13.408 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 12:12:13.625 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:12:13.711 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:12:13.783 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:13:20.123 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000148/group_3/index_6711_seed_24370_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 12:13:23.198 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:13:41.069 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:13:41.078 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:13:41.080 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:14:49.609 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000148/group_3/index_6711_seed_24371_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 12:14:54.474 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:15:12.824 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:15:13.144 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:15:14.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:17:09.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.69s (1.91min) +2026-06-20 12:17:09.788 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:19:10.588 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:26:11.396 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 149, train_steps: 5960 +2026-06-20 12:26:12.135 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 12:26:12.282 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:26:12.332 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:26:12.417 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:27:18.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000149/group_3/index_8283_seed_24359_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 12:27:22.925 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:27:39.651 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:27:39.662 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:27:39.662 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:28:48.443 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000149/group_3/index_8283_seed_24360_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 12:28:52.281 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:29:08.898 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:29:13.029 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:29:14.045 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:31:08.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.71s (1.91min) +2026-06-20 12:31:09.533 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:33:09.830 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:40:10.988 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 150, train_steps: 6000 +2026-06-20 12:41:53.431 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=150 (missing=0, unexpected=0) +2026-06-20 12:41:56.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 12:41:56.262 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:41:56.385 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:41:56.484 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:43:03.939 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_3/index_2286_seed_27840_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 12:43:06.979 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:43:23.474 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:43:23.484 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:43:23.485 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:44:32.540 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_3/index_2286_seed_27841_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 12:44:37.494 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:44:54.728 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:44:59.744 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:45:00.810 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:46:55.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.70s (1.91min) +2026-06-20 12:46:56.236 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:48:52.989 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:55:51.609 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 151, train_steps: 6040 +2026-06-20 12:55:52.308 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 12:55:52.533 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:55:52.643 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:55:52.735 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:56:58.572 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_3/index_2264_seed_18583_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 12:57:02.336 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:57:19.085 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:57:19.096 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:57:19.096 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:58:27.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_3/index_2264_seed_18584_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 12:58:32.127 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:58:49.782 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:58:50.721 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:58:51.766 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:00:46.190 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.41s (1.91min) +2026-06-20 13:00:46.908 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:02:43.653 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:09:41.910 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 152, train_steps: 6080 +2026-06-20 13:09:42.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 13:09:42.770 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:09:42.844 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:09:42.943 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:10:49.234 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_3/index_283_seed_20671_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 13:10:52.954 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:11:09.339 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:11:09.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:11:09.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:12:20.614 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_3/index_283_seed_20672_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 13:12:23.840 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:12:40.478 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:12:43.910 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:12:44.864 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:14:38.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-20 13:14:39.491 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:16:38.549 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:23:37.228 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 153, train_steps: 6120 +2026-06-20 13:23:38.091 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 13:23:38.231 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:23:38.272 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:23:38.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:24:44.808 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_3/index_1935_seed_8802_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 13:24:49.740 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:25:07.196 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:25:07.206 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:25:07.206 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:26:15.174 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_3/index_1935_seed_8803_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 13:26:19.946 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:26:37.199 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:26:40.127 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:26:41.339 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:28:34.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.39s (1.89min) +2026-06-20 13:28:35.473 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:30:32.908 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:37:31.707 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 154, train_steps: 6160 +2026-06-20 13:37:32.393 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 13:37:32.575 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:37:32.671 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:37:32.751 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:38:38.754 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_3/index_5276_seed_18332_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 13:38:43.431 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:39:00.062 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:39:00.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:39:00.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:40:09.395 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_3/index_5276_seed_18333_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 13:40:14.393 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:40:32.522 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:40:33.788 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:40:34.837 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:42:29.329 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.47s (1.91min) +2026-06-20 13:42:30.061 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:44:27.162 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:51:24.973 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 155, train_steps: 6200 +2026-06-20 13:51:25.702 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 13:51:25.919 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:51:26.043 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:51:26.121 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:52:32.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_3/index_3699_seed_2794_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 13:52:36.727 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:52:53.376 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:52:53.386 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:52:53.388 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:54:01.074 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_3/index_3699_seed_2795_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 13:54:05.058 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:54:21.941 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:54:24.447 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:54:25.359 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:56:19.129 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.63s (1.89min) +2026-06-20 13:56:19.859 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:58:16.748 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:05:15.350 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 156, train_steps: 6240 +2026-06-20 14:05:16.035 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 14:05:16.237 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:05:16.347 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:05:16.451 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:06:23.535 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_3/index_3603_seed_31476_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 14:06:26.705 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:06:43.121 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:06:43.131 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:06:43.131 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:07:51.684 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_3/index_3603_seed_31477_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 14:07:56.538 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:08:13.398 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:08:15.931 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:08:16.977 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:10:11.722 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.74s (1.91min) +2026-06-20 14:10:12.457 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:12:11.105 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:19:12.249 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 157, train_steps: 6280 +2026-06-20 14:19:12.925 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 14:19:13.113 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:19:13.212 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:19:13.305 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:20:19.408 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000157/group_3/index_4415_seed_32589_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 14:20:24.230 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:20:41.206 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:20:41.215 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:20:41.216 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:21:48.038 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000157/group_3/index_4415_seed_32590_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 14:21:52.838 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:22:09.749 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:22:12.398 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:22:13.731 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:24:08.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.22s (1.92min) +2026-06-20 14:24:09.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:26:09.904 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:33:10.835 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 158, train_steps: 6320 +2026-06-20 14:33:11.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 14:33:11.708 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:33:11.806 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:33:11.902 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:34:17.625 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_3/index_3963_seed_31397_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 14:34:22.462 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:34:39.852 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:34:39.862 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:34:39.862 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:35:47.078 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_3/index_3963_seed_31398_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 14:35:51.780 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:36:09.102 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:36:10.935 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:36:12.076 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:38:07.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.88s (1.91min) +2026-06-20 14:38:07.784 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:40:05.257 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:47:04.565 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 159, train_steps: 6360 +2026-06-20 14:47:05.291 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 14:47:05.396 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:47:05.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:47:05.483 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:48:11.584 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_3/index_6395_seed_12890_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 14:48:15.788 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:48:32.284 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:48:32.294 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:48:32.294 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:49:40.101 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_3/index_6395_seed_12891_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 14:49:44.947 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:50:01.872 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:50:04.049 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:50:05.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:51:59.193 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.86s (1.90min) +2026-06-20 14:51:59.929 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:53:57.320 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:00:59.191 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 160, train_steps: 6400 +2026-06-20 15:01:37.532 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=160 (missing=0, unexpected=0) +2026-06-20 15:01:39.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 15:01:39.382 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:01:39.397 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:01:39.535 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:02:46.919 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_3/index_1684_seed_23436_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 15:02:51.522 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:03:07.713 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:03:07.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:03:07.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:04:18.797 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_3/index_1684_seed_23437_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 15:04:23.448 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:04:39.925 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:04:42.879 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:04:43.916 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:06:37.409 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.48s (1.89min) +2026-06-20 15:06:38.136 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:08:38.304 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:15:37.528 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 161, train_steps: 6440 +2026-06-20 15:15:38.206 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 15:15:38.347 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:15:38.385 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:15:38.478 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:16:44.371 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_3/index_7012_seed_24124_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 15:16:47.785 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:17:03.390 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:17:03.400 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:17:03.400 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:18:16.494 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_3/index_7012_seed_24125_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 15:18:21.293 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:18:37.570 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:18:40.375 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:18:41.608 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:20:35.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-20 15:20:36.332 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:22:36.148 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:29:36.894 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 162, train_steps: 6480 +2026-06-20 15:29:37.570 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 15:29:37.698 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:29:37.731 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:29:37.794 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:30:43.891 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_3/index_1838_seed_19274_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 15:30:47.290 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:31:03.514 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:31:03.525 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:31:03.525 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:32:11.924 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_3/index_1838_seed_19275_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 15:32:15.291 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:32:31.580 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:32:37.444 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:32:38.379 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:34:33.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.48s (1.91min) +2026-06-20 15:34:33.827 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:36:31.317 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:43:31.482 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 163, train_steps: 6520 +2026-06-20 15:43:32.171 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 15:43:32.309 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:43:32.347 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:43:32.440 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:44:38.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000163/group_3/index_7721_seed_23608_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 15:44:43.528 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:44:59.693 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:44:59.704 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:44:59.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:46:09.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000163/group_3/index_7721_seed_23609_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 15:46:14.159 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:46:31.967 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:46:33.676 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:46:34.722 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:48:29.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.47s (1.91min) +2026-06-20 15:48:29.934 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:50:26.537 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:57:25.505 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 164, train_steps: 6560 +2026-06-20 15:57:26.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 15:57:26.430 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:57:26.553 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:57:26.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:58:32.383 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_3/index_1965_seed_13486_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 15:58:37.105 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:58:53.423 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:58:53.430 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:58:53.431 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:00:02.604 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_3/index_1965_seed_13487_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 16:00:06.250 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:00:22.440 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:00:26.402 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:00:27.358 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:02:21.064 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.61s (1.89min) +2026-06-20 16:02:21.793 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:04:18.345 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:11:18.388 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 165, train_steps: 6600 +2026-06-20 16:11:19.088 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 16:11:19.251 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:11:19.326 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:11:19.451 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:12:25.606 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_3/index_8720_seed_10993_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 16:12:28.572 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:12:45.207 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:12:45.217 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:12:45.217 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:13:54.767 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_3/index_8720_seed_10994_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 16:13:59.590 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:14:16.021 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:14:19.106 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:14:20.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:16:14.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.75s (1.90min) +2026-06-20 16:16:14.870 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:18:14.123 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:25:13.935 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 166, train_steps: 6640 +2026-06-20 16:25:14.621 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 16:25:14.822 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:25:14.951 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:25:15.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:26:21.933 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_3/index_3093_seed_30439_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 16:26:26.589 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:26:42.695 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:26:42.704 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:26:42.705 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:27:49.965 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_3/index_3093_seed_30440_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 16:27:54.606 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:28:10.988 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:28:13.831 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:28:15.063 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:30:08.844 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.78s (1.90min) +2026-06-20 16:30:09.574 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:32:10.306 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:39:12.153 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 167, train_steps: 6680 +2026-06-20 16:39:12.851 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 16:39:13.067 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:39:13.152 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:39:13.243 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:40:19.571 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_3/index_7219_seed_24697_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 16:40:23.793 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:40:39.375 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:40:39.385 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:40:39.385 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:41:47.958 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_3/index_7219_seed_24698_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 16:41:51.047 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:42:06.987 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:42:12.095 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:42:13.267 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:44:08.745 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.36s (1.92min) +2026-06-20 16:44:09.477 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:46:10.018 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:53:11.383 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 168, train_steps: 6720 +2026-06-20 16:53:12.092 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 16:53:12.314 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:53:12.444 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:53:12.511 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:54:18.860 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_3/index_5004_seed_32914_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 16:54:23.618 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:54:39.581 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:54:39.591 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:54:39.591 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:55:47.408 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_3/index_5004_seed_32915_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 16:55:51.977 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:56:08.190 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:56:11.768 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:56:12.803 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:58:07.736 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-20 16:58:08.464 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:00:07.578 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:07:09.188 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 169, train_steps: 6760 +2026-06-20 17:07:09.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 17:07:10.098 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 17:07:10.238 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:07:10.342 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:08:16.416 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_3/index_32_seed_31154_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 17:08:20.982 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 17:08:38.579 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:08:38.590 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:08:38.590 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:09:47.827 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_3/index_32_seed_31155_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 17:09:52.544 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 17:10:10.170 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:10:11.904 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 17:10:12.949 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:12:08.255 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.20s (1.92min) +2026-06-20 17:12:08.969 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:14:09.318 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:21:11.051 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 170, train_steps: 6800 +2026-06-20 17:21:47.292 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=170 (missing=0, unexpected=0) +2026-06-20 17:21:49.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 17:21:49.431 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 17:21:49.546 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:21:49.632 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:22:56.514 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_3/index_413_seed_25252_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 17:23:00.247 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 17:23:15.960 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:23:15.970 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:23:15.970 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:24:25.364 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_3/index_413_seed_25253_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 17:24:28.427 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 17:24:44.789 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:24:49.925 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 17:24:50.956 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:26:44.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.05s (1.88min) +2026-06-20 17:26:44.743 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:28:41.425 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:35:39.735 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 14, sp_rank 0: update step: 171, train_steps: 6840 +2026-06-20 17:35:40.432 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 14, dp_rank 7: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=3, rank_in_group=1 +2026-06-20 17:35:40.593 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 17:35:40.780 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:35:40.933 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:36:47.120 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000171/group_3/index_5440_seed_31138_dp_7_rank_14_sample_2_slot_0.mp4 +2026-06-20 17:36:50.387 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 17:37:06.220 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:37:06.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 14 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:37:06.231 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 14 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:38:19.223 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 14: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000171/group_3/index_5440_seed_31139_dp_7_rank_14_sample_3_slot_1.mp4 +2026-06-20 17:38:24.015 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 17:38:40.351 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:38:42.537 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 14: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 17:38:43.758 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 14: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:40:36.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 14: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.89s (1.88min) +2026-06-20 17:40:37.378 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 14, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-29 01:30:24.943 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 01:30:25.606 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 01:30:25.607 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 01:53:14.558 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 01:53:15.248 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 01:53:15.249 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 02:23:31.673 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 02:23:32.349 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 02:23:32.350 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 02:48:39.387 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 02:48:40.057 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 02:48:40.058 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:11:16.690 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 03:11:17.346 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:11:17.347 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:34:31.149 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 03:34:31.830 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:34:31.831 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:00:52.006 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 04:00:52.669 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:00:52.672 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:28:55.639 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 04:28:56.323 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:28:56.323 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:54:00.341 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 04:54:01.022 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:54:01.024 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 05:18:09.640 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 05:18:10.299 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 05:18:10.302 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 05:42:02.804 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 05:42:03.484 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 05:42:03.485 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:05:56.595 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 06:05:57.268 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:05:57.269 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:26:59.948 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 06:27:00.608 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:27:00.610 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:51:38.655 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 06:51:39.315 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:51:39.318 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 07:14:09.509 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 07:14:10.186 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 07:14:10.189 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 07:37:06.458 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 07:37:07.123 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 07:37:07.123 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:07:16.031 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 08:07:16.701 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:07:16.702 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:37:01.667 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 08:37:02.342 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:37:02.342 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:01:09.938 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 09:01:10.622 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:01:10.623 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:24:43.042 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 09:24:43.713 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:24:43.714 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:49:42.166 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 09:49:42.827 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:49:42.831 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 10:13:57.682 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 10:13:58.358 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 10:13:58.359 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 10:38:10.273 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 10:38:10.932 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 10:38:10.935 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:05:35.472 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 11:05:36.147 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:05:36.147 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:29:16.966 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 11:29:17.645 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:29:17.648 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:57:21.353 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 11:57:22.009 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:57:22.010 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 12:20:18.543 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 12:20:19.204 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 12:20:19.206 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 12:48:19.759 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 12:48:20.429 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 12:48:20.429 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 13:14:44.293 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 13:14:44.967 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 13:14:44.968 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 13:39:09.189 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 13:39:09.845 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 13:39:09.852 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:06:07.866 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 14:06:08.531 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:06:08.532 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:28:55.269 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 14:28:55.933 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:28:55.935 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:51:29.444 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 14:51:30.101 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:51:30.104 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 15:22:59.801 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 15:23:00.457 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 15:23:00.460 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 15:48:17.137 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 15:48:17.789 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 15:48:17.792 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 16:18:25.295 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 16:18:25.979 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 16:18:25.980 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 16:44:48.609 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 16:44:49.264 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 16:44:49.264 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:10:30.031 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 17:10:30.720 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:10:30.721 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:34:03.152 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 17:34:03.836 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:34:03.836 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:55:28.113 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 17:55:28.765 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:55:28.768 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 18:18:05.300 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 18:18:05.981 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 18:18:05.982 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 18:40:44.340 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 18:40:45.012 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 18:40:45.012 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_21.log b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_21.log new file mode 100644 index 0000000000000000000000000000000000000000..d4402b708441c1333b17874539d8ea16b86eec9b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_21.log @@ -0,0 +1,1994 @@ +2026-06-19 01:20:15.025 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-19 01:20:15.026 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-19 01:30:17.294 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 01:30:17.405 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:30:17.411 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:31:52.326 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:31:52.328 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:33:45.514 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 01:34:00.996 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 01:35:56.996 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.94s (1.93min) +2026-06-19 01:35:57.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 01:37:54.189 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 01:37:57.219 | INFO | hyvideo.utils.grpo_utils:train_one_step:3127 - Rank 21, sp_rank 0: init adaptive KL controller (init=1.000e-07, min=1.000e-07, max=1.000e-05) +2026-06-19 01:45:11.468 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 1, train_steps: 40 +2026-06-19 01:45:12.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 01:45:12.696 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:45:12.725 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:46:17.238 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:46:17.241 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:48:10.944 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 01:48:11.987 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 01:50:07.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.70s (1.93min) +2026-06-19 01:50:08.465 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 01:52:06.953 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 01:59:06.754 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 2, train_steps: 80 +2026-06-19 01:59:07.380 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 01:59:07.484 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:59:07.499 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:00:12.405 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:00:12.406 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:02:10.528 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 02:02:11.674 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:04:06.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.88s (1.91min) +2026-06-19 02:04:07.373 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:06:03.752 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:13:04.855 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 3, train_steps: 120 +2026-06-19 02:13:05.465 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 02:13:05.582 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:13:05.593 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:14:10.693 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:14:10.694 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:16:04.655 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 02:16:05.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:17:59.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.77s (1.90min) +2026-06-19 02:18:00.447 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:19:59.348 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:26:59.883 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 4, train_steps: 160 +2026-06-19 02:27:00.546 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 02:27:00.656 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:27:00.702 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:28:05.194 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:28:05.330 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:30:00.316 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 02:30:01.107 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:31:55.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.62s (1.91min) +2026-06-19 02:31:56.650 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:33:53.222 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:40:53.301 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 5, train_steps: 200 +2026-06-19 02:40:53.901 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 02:40:54.075 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:40:54.141 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:41:58.702 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:41:58.705 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:43:51.811 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 02:43:52.597 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:45:46.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.50s (1.89min) +2026-06-19 02:45:46.993 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:47:43.350 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:54:40.652 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 6, train_steps: 240 +2026-06-19 02:54:41.360 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 02:54:41.493 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:54:41.534 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:55:46.481 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:55:46.484 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:57:42.949 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 02:57:43.899 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:59:37.353 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.38s (1.89min) +2026-06-19 02:59:38.042 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:01:37.729 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:08:37.480 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 7, train_steps: 280 +2026-06-19 03:08:38.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 03:08:38.240 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:08:38.279 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:09:42.818 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:09:42.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:11:38.337 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 03:11:39.434 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:13:34.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.68s (1.91min) +2026-06-19 03:13:34.976 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:15:32.042 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:22:31.149 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 8, train_steps: 320 +2026-06-19 03:22:31.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 03:22:32.016 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:22:32.152 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:23:37.303 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:23:37.306 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:25:33.100 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 03:25:34.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:27:28.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-19 03:27:29.431 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:29:26.903 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:36:25.644 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 9, train_steps: 360 +2026-06-19 03:36:26.346 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 03:36:26.460 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:36:26.501 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:37:31.095 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:37:31.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:39:26.010 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 03:39:26.948 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:41:20.400 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.36s (1.89min) +2026-06-19 03:41:21.105 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:43:17.560 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:50:16.929 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 10, train_steps: 400 +2026-06-19 03:50:53.207 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=10 (missing=0, unexpected=0) +2026-06-19 03:50:56.181 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 03:50:56.313 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:50:56.368 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:52:02.286 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:52:02.289 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:53:57.828 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 03:53:58.697 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:55:52.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.65s (1.89min) +2026-06-19 03:55:53.208 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:57:49.762 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:04:48.377 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 11, train_steps: 440 +2026-06-19 04:04:49.074 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 04:04:49.270 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:04:49.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:05:53.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:05:53.970 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:07:48.852 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 04:07:49.649 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:09:43.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.25s (1.89min) +2026-06-19 04:09:43.796 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:11:40.268 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:18:37.289 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 12, train_steps: 480 +2026-06-19 04:18:37.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 04:18:38.064 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:18:38.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:19:42.606 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:19:42.606 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:21:37.180 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 04:21:37.983 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:23:33.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.19s (1.92min) +2026-06-19 04:23:34.069 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:25:31.242 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:32:29.817 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 13, train_steps: 520 +2026-06-19 04:32:30.515 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 04:32:30.726 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:32:30.843 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:33:35.457 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:33:35.460 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:35:30.313 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 04:35:31.110 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:37:24.102 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.77s (1.88min) +2026-06-19 04:37:24.791 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:39:21.075 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:46:18.591 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 14, train_steps: 560 +2026-06-19 04:46:19.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 04:46:19.387 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:46:19.423 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:47:24.285 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:47:24.285 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:49:17.445 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 04:49:18.392 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:51:12.963 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.49s (1.91min) +2026-06-19 04:51:13.649 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:53:12.993 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:00:11.824 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 15, train_steps: 600 +2026-06-19 05:00:12.526 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 05:00:12.709 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:00:12.807 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:01:17.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:01:17.673 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:03:11.349 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 05:03:12.602 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:05:07.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.64s (1.91min) +2026-06-19 05:05:08.060 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:07:05.977 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:14:04.889 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 16, train_steps: 640 +2026-06-19 05:14:05.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 05:14:05.711 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:14:05.770 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:15:10.715 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:15:10.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:17:06.988 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 05:17:07.836 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:19:03.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.02s (1.92min) +2026-06-19 05:19:03.742 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:21:01.959 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:28:03.321 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 17, train_steps: 680 +2026-06-19 05:28:04.007 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 05:28:04.217 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:28:04.352 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:29:08.922 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:29:08.922 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:31:03.926 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 05:31:04.843 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:32:58.689 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.48s (1.89min) +2026-06-19 05:32:59.386 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:34:56.265 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:41:58.103 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 18, train_steps: 720 +2026-06-19 05:41:58.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 05:41:58.866 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:41:58.913 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:43:03.638 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:43:03.640 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:45:00.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 05:45:01.022 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:46:55.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.88s (1.90min) +2026-06-19 05:46:55.796 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:48:52.329 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:55:50.495 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 19, train_steps: 760 +2026-06-19 05:55:51.177 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 05:55:51.310 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:55:51.346 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:56:56.020 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:56:56.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:58:49.068 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 05:58:49.871 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:00:43.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.10s (1.89min) +2026-06-19 06:00:43.875 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:02:40.342 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:09:39.824 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 20, train_steps: 800 +2026-06-19 06:10:15.860 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=20 (missing=0, unexpected=0) +2026-06-19 06:10:18.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 06:10:18.871 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:10:18.924 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:11:24.997 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:11:24.999 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:13:19.050 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 06:13:19.963 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:15:13.224 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.90s (1.88min) +2026-06-19 06:15:13.910 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:17:12.688 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:24:12.560 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 21, train_steps: 840 +2026-06-19 06:24:13.271 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 06:24:13.429 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:24:13.504 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:25:17.886 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:25:17.890 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:27:11.260 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 06:27:12.136 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:29:06.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.51s (1.91min) +2026-06-19 06:29:07.667 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:31:04.579 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:38:04.295 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 22, train_steps: 880 +2026-06-19 06:38:04.947 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 06:38:05.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:38:05.094 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:39:10.462 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:39:10.466 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:41:03.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 06:41:04.158 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:42:58.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.99s (1.90min) +2026-06-19 06:42:59.164 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:44:56.506 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:51:58.166 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 23, train_steps: 920 +2026-06-19 06:51:58.842 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 06:51:59.001 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:51:59.084 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:53:03.404 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:53:03.407 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:54:59.789 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 06:55:00.756 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:56:54.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.11s (1.90min) +2026-06-19 06:56:55.621 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:58:52.317 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:05:51.316 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 24, train_steps: 960 +2026-06-19 07:05:52.005 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 07:05:52.184 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:05:52.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:06:56.942 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:06:56.946 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:08:52.217 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 07:08:53.163 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:10:46.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-19 07:10:46.870 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:12:44.152 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:19:42.041 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 25, train_steps: 1000 +2026-06-19 07:20:52.105 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 07:20:52.307 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:20:52.406 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:21:56.908 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:21:56.911 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:23:52.666 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 07:23:53.660 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:25:46.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.81s (1.88min) +2026-06-19 07:25:47.204 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:27:44.378 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:34:43.683 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 26, train_steps: 1040 +2026-06-19 07:34:44.357 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 07:34:44.561 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:34:44.678 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:35:49.007 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:35:49.011 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:37:45.443 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 07:37:46.412 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:39:39.638 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.14s (1.89min) +2026-06-19 07:39:40.347 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:41:39.070 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:48:39.674 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 27, train_steps: 1080 +2026-06-19 07:48:40.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 07:48:40.511 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:48:40.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:49:45.010 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:49:45.010 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:51:43.285 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 07:51:44.241 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:53:38.298 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.99s (1.90min) +2026-06-19 07:53:39.006 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:55:38.420 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:02:38.168 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 28, train_steps: 1120 +2026-06-19 08:02:38.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 08:02:39.074 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:02:39.181 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:03:43.507 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:03:43.512 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:05:41.901 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 08:05:43.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:07:37.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.24s (1.90min) +2026-06-19 08:07:38.091 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:09:37.932 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:16:37.854 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 29, train_steps: 1160 +2026-06-19 08:16:38.692 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 08:16:38.874 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:16:39.135 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:17:43.212 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:17:43.216 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:19:42.881 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 08:19:43.820 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:21:37.264 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 08:21:37.960 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:23:37.404 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:30:39.540 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 30, train_steps: 1200 +2026-06-19 08:31:15.682 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=30 (missing=0, unexpected=0) +2026-06-19 08:31:18.258 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 08:31:18.397 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:31:18.444 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:32:25.115 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:32:25.118 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:34:19.980 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 08:34:21.063 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:36:14.510 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-19 08:36:15.226 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:38:14.171 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:45:14.280 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 31, train_steps: 1240 +2026-06-19 08:45:14.993 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 08:45:15.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:45:15.165 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:46:19.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:46:19.582 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:48:15.288 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 08:48:16.416 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:50:10.323 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.84s (1.90min) +2026-06-19 08:50:11.022 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:52:11.147 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:59:13.099 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 32, train_steps: 1280 +2026-06-19 08:59:13.775 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 08:59:13.875 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:59:13.919 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:00:18.427 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:00:18.430 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:02:13.436 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 09:02:14.602 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:04:08.713 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.91s (1.90min) +2026-06-19 09:04:09.420 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:06:10.407 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:13:13.286 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 33, train_steps: 1320 +2026-06-19 09:13:13.952 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 09:13:14.132 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:13:14.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:14:18.864 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:14:18.868 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:16:14.822 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 09:16:15.845 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:18:11.066 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.06s (1.92min) +2026-06-19 09:18:11.785 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:20:11.626 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:27:13.735 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 34, train_steps: 1360 +2026-06-19 09:27:14.444 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 09:27:14.569 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:27:14.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:28:19.108 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:28:19.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:30:13.695 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 09:30:14.797 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:32:08.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.81s (1.90min) +2026-06-19 09:32:09.591 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:34:09.780 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:41:12.401 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 35, train_steps: 1400 +2026-06-19 09:41:13.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 09:41:13.227 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:41:13.298 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:42:17.650 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:42:17.653 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:44:13.080 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 09:44:14.049 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:46:10.686 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 116.56s (1.94min) +2026-06-19 09:46:11.397 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:48:10.895 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:55:11.961 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 36, train_steps: 1440 +2026-06-19 09:55:12.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 09:55:12.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:55:12.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:56:17.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:56:17.521 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:58:12.646 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 09:58:13.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:00:08.966 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-19 10:00:09.671 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:02:10.006 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:09:11.705 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 37, train_steps: 1480 +2026-06-19 10:09:12.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 10:09:12.544 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:09:12.587 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:10:17.010 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:10:17.013 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:12:13.042 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 10:12:13.851 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:14:09.030 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.94s (1.92min) +2026-06-19 10:14:09.737 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:16:10.366 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:23:12.347 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 38, train_steps: 1520 +2026-06-19 10:23:13.053 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 10:23:13.266 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:23:13.392 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:24:17.863 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:24:17.864 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:26:12.667 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 10:26:13.690 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:28:09.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.16s (1.92min) +2026-06-19 10:28:09.954 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:30:10.743 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:37:11.955 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 39, train_steps: 1560 +2026-06-19 10:37:12.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 10:37:12.845 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:37:12.969 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:38:17.493 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:38:17.498 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:41:42.526 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 10:41:43.525 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:43:36.471 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.70s (1.88min) +2026-06-19 10:43:37.179 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:45:37.174 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:52:37.859 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 40, train_steps: 1600 +2026-06-19 10:53:13.570 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=40 (missing=0, unexpected=0) +2026-06-19 10:53:15.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 10:53:15.888 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:53:15.924 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:54:22.092 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:54:22.095 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:56:16.410 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 10:56:17.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:58:11.717 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.27s (1.90min) +2026-06-19 10:58:12.415 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:00:11.327 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:07:12.584 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 41, train_steps: 1640 +2026-06-19 11:07:13.291 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 11:07:13.398 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:07:13.435 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:08:17.948 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:08:17.948 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:10:12.668 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 11:10:13.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:12:08.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.28s (1.90min) +2026-06-19 11:12:09.086 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:14:08.755 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:21:10.889 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 42, train_steps: 1680 +2026-06-19 11:21:11.804 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 11:21:11.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:21:12.094 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:22:16.556 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:22:16.559 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:24:12.073 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 11:24:13.093 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:26:07.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.32s (1.91min) +2026-06-19 11:26:08.278 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:28:06.796 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:35:06.766 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 43, train_steps: 1720 +2026-06-19 11:35:07.458 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 11:35:07.645 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:35:07.754 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:36:12.609 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:36:12.613 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:38:10.299 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 11:38:11.273 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:40:05.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.38s (1.91min) +2026-06-19 11:40:06.548 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:42:03.522 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:49:04.433 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 44, train_steps: 1760 +2026-06-19 11:49:05.115 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 11:49:05.323 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:49:05.472 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:50:10.667 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:50:10.667 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:52:06.491 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 11:52:07.360 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:54:01.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.69s (1.89min) +2026-06-19 11:54:01.909 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:56:01.233 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:03:03.873 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 45, train_steps: 1800 +2026-06-19 12:03:04.557 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 12:03:04.742 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:03:04.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:04:10.061 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:04:10.062 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:06:05.347 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 12:06:06.309 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:08:01.417 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.05s (1.92min) +2026-06-19 12:08:02.100 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:10:01.381 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:17:03.593 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 46, train_steps: 1840 +2026-06-19 12:17:04.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 12:17:04.425 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:17:04.466 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:18:09.986 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:18:09.987 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:20:05.333 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 12:20:06.493 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:22:01.733 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.17s (1.92min) +2026-06-19 12:22:02.441 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:24:01.449 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:31:02.928 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 47, train_steps: 1880 +2026-06-19 12:31:03.607 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 12:31:03.774 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:31:03.856 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:32:08.456 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:32:08.456 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:34:03.182 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 12:34:04.452 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:35:58.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-19 12:35:58.981 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:37:55.889 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:44:57.677 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 48, train_steps: 1920 +2026-06-19 12:44:58.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 12:44:58.557 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:44:58.703 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:46:03.081 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:46:03.084 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:47:58.930 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 12:47:59.813 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:49:54.620 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-19 12:49:55.342 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:51:52.289 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:58:52.091 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 49, train_steps: 1960 +2026-06-19 12:58:52.782 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 12:58:52.905 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:58:52.943 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:59:57.521 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:59:57.521 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:01:54.369 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 13:01:55.178 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:03:48.446 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.04s (1.88min) +2026-06-19 13:03:49.148 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:05:45.465 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:12:45.277 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 50, train_steps: 2000 +2026-06-19 13:14:21.459 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=50 (missing=0, unexpected=0) +2026-06-19 13:14:23.203 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 13:14:23.333 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:14:23.378 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:15:30.038 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:15:30.041 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:17:25.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 13:17:26.085 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:19:20.081 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.84s (1.90min) +2026-06-19 13:19:20.806 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:21:17.522 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:28:18.472 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 51, train_steps: 2040 +2026-06-19 13:28:19.185 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 13:28:19.357 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:28:19.455 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:29:23.998 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:29:24.001 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:31:18.028 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 13:31:19.043 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:33:13.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.83s (1.90min) +2026-06-19 13:33:13.741 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:35:13.226 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:42:13.609 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 52, train_steps: 2080 +2026-06-19 13:42:14.276 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 13:42:14.434 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:42:14.483 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:43:19.451 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:43:19.453 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:45:15.287 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 13:45:16.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:47:10.481 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.03s (1.90min) +2026-06-19 13:47:11.196 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:49:11.262 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:56:12.612 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 53, train_steps: 2120 +2026-06-19 13:56:13.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 13:56:13.498 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:56:13.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:57:18.175 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:57:18.178 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:59:13.130 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 13:59:14.003 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:01:08.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.41s (1.91min) +2026-06-19 14:01:09.528 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:03:10.473 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:10:12.170 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 54, train_steps: 2160 +2026-06-19 14:10:12.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 14:10:13.011 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:10:13.047 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:11:17.599 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:11:17.602 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:13:12.866 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 14:13:13.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:15:08.493 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.42s (1.91min) +2026-06-19 14:15:09.217 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:17:09.303 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:24:11.907 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 55, train_steps: 2200 +2026-06-19 14:24:12.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 14:24:12.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:24:12.911 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:25:17.548 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:25:17.549 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:27:12.754 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 14:27:13.738 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:29:09.888 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.82s (1.93min) +2026-06-19 14:29:10.604 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:31:10.762 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:38:12.115 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 56, train_steps: 2240 +2026-06-19 14:38:12.811 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 14:38:12.998 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:38:13.101 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:39:17.651 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:39:17.652 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:41:13.092 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 14:41:13.915 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:43:08.781 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.66s (1.91min) +2026-06-19 14:43:09.474 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:45:09.715 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:52:11.847 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 57, train_steps: 2280 +2026-06-19 14:52:12.526 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 14:52:12.654 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:52:12.696 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:53:17.218 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:53:17.220 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:55:13.057 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 14:55:13.892 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:57:08.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.79s (1.91min) +2026-06-19 14:57:09.596 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:59:09.878 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:06:12.289 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 58, train_steps: 2320 +2026-06-19 15:06:12.986 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 15:06:13.175 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:06:13.297 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:07:17.812 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:07:17.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:09:13.062 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 15:09:13.889 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:11:08.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.71s (1.91min) +2026-06-19 15:11:09.497 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:13:09.810 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:20:12.054 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 59, train_steps: 2360 +2026-06-19 15:20:12.738 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 15:20:12.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:20:12.911 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:21:17.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:21:17.582 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:23:12.372 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 15:23:13.350 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:25:09.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.28s (1.92min) +2026-06-19 15:25:09.742 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:27:10.364 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:34:12.072 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 60, train_steps: 2400 +2026-06-19 15:34:48.114 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=60 (missing=0, unexpected=0) +2026-06-19 15:34:50.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 15:34:50.801 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:34:50.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:35:56.953 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:35:56.956 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:37:53.529 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 15:37:54.563 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:39:47.707 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-19 15:39:48.403 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:41:45.233 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:48:44.584 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 61, train_steps: 2440 +2026-06-19 15:48:45.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 15:48:45.454 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:48:45.540 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:49:50.126 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:49:50.129 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:51:47.609 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 15:51:48.427 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:53:41.681 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.04s (1.88min) +2026-06-19 15:53:42.399 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:55:39.325 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:02:39.008 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 62, train_steps: 2480 +2026-06-19 16:02:39.694 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 16:02:39.892 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:02:40.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:03:44.403 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:03:44.406 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:05:44.132 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 16:05:44.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:07:38.761 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.59s (1.89min) +2026-06-19 16:07:39.470 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:09:38.853 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:16:39.334 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 63, train_steps: 2520 +2026-06-19 16:16:40.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 16:16:40.127 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:16:40.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:17:44.912 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:17:44.915 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:19:45.197 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 16:19:46.038 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:21:39.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.36s (1.89min) +2026-06-19 16:21:40.302 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:23:38.880 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:30:40.937 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 64, train_steps: 2560 +2026-06-19 16:30:41.625 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 16:30:41.751 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:30:41.786 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:31:46.620 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:31:46.620 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:33:45.471 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 16:33:46.295 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:35:40.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.63s (1.89min) +2026-06-19 16:35:40.843 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:37:39.397 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:44:41.296 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 65, train_steps: 2600 +2026-06-19 16:44:41.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 16:44:42.117 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:44:42.171 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:45:46.804 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:45:46.807 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:47:47.482 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 16:47:48.298 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:49:42.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.72s (1.90min) +2026-06-19 16:49:42.958 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:51:40.034 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:58:41.281 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 66, train_steps: 2640 +2026-06-19 16:58:41.989 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 16:58:42.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:58:42.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:59:46.752 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:59:46.753 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:01:43.599 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 17:01:44.425 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:03:37.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.34s (1.89min) +2026-06-19 17:03:38.681 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:05:38.381 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:12:40.119 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 67, train_steps: 2680 +2026-06-19 17:12:40.909 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 17:12:41.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:12:41.481 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:13:45.918 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:13:45.922 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:15:44.529 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 17:15:45.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:17:38.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 17:17:39.678 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:19:39.110 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:26:39.547 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 68, train_steps: 2720 +2026-06-19 17:26:40.215 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 17:26:40.362 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:26:40.417 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:27:45.181 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:27:45.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:29:43.886 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 17:29:44.865 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:31:38.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.76s (1.90min) +2026-06-19 17:31:39.420 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:33:38.459 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:40:38.366 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 69, train_steps: 2760 +2026-06-19 17:40:39.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 17:40:39.142 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:40:39.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:41:43.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:41:43.604 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:43:41.069 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 17:43:42.019 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:45:35.854 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.55s (1.89min) +2026-06-19 17:45:36.569 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:47:37.113 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:54:38.578 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 70, train_steps: 2800 +2026-06-19 17:55:14.745 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=70 (missing=0, unexpected=0) +2026-06-19 17:55:16.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 17:55:17.105 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:55:17.136 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:56:23.802 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:56:23.805 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:58:18.228 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 17:58:19.274 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:00:13.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.58s (1.89min) +2026-06-19 18:00:13.837 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:02:13.850 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:09:13.911 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 71, train_steps: 2840 +2026-06-19 18:09:14.615 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 18:09:14.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:09:14.831 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:10:19.420 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:10:19.423 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:12:16.021 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 18:12:16.863 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:14:12.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.67s (1.93min) +2026-06-19 18:14:13.459 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:16:12.692 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:23:13.836 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 72, train_steps: 2880 +2026-06-19 18:23:14.506 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 18:23:14.664 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:23:14.732 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:24:19.524 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:24:19.525 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:26:14.711 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 18:26:15.691 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:28:10.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.45s (1.91min) +2026-06-19 18:28:11.145 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:30:10.988 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:37:12.793 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 73, train_steps: 2920 +2026-06-19 18:37:13.455 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 18:37:13.580 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:37:13.618 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:38:18.455 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:38:18.455 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:40:13.249 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 18:40:14.080 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:42:09.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.16s (1.92min) +2026-06-19 18:42:10.284 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:44:10.447 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:51:12.754 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 74, train_steps: 2960 +2026-06-19 18:51:13.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 18:51:13.646 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:51:13.749 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:52:18.535 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:52:18.536 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:54:13.982 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 18:54:14.801 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:56:10.312 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.21s (1.92min) +2026-06-19 18:56:11.030 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:58:11.441 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:05:14.041 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 75, train_steps: 3000 +2026-06-19 19:06:23.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 19:06:23.407 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:06:23.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:07:28.239 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:07:28.242 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:09:24.209 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 19:09:25.046 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:11:18.397 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.14s (1.89min) +2026-06-19 19:11:19.100 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:13:16.252 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:20:16.319 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 76, train_steps: 3040 +2026-06-19 19:20:16.989 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 19:20:17.152 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:20:17.232 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:21:21.542 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:21:21.542 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:23:18.455 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 19:23:19.460 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:25:13.441 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.66s (1.89min) +2026-06-19 19:25:14.145 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:27:12.732 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:34:13.271 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 77, train_steps: 3080 +2026-06-19 19:34:13.950 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 19:34:14.131 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:34:14.234 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:35:18.562 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:35:18.567 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:37:13.073 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 19:37:13.902 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:39:08.734 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.60s (1.91min) +2026-06-19 19:39:09.440 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:41:09.768 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:48:12.652 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 78, train_steps: 3120 +2026-06-19 19:48:13.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 19:48:13.554 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:48:13.680 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:49:18.270 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:49:18.273 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:51:13.323 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 19:51:14.171 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:53:09.267 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-19 19:53:09.975 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:55:09.984 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:02:13.175 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 79, train_steps: 3160 +2026-06-19 20:02:13.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 20:02:14.036 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:02:14.091 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:03:18.731 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:03:18.731 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:05:13.093 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 20:05:13.909 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:07:09.660 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.53s (1.93min) +2026-06-19 20:07:10.385 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:09:11.031 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:16:13.934 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 80, train_steps: 3200 +2026-06-19 20:16:51.456 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=80 (missing=0, unexpected=0) +2026-06-19 20:16:53.092 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 20:16:53.218 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:16:53.271 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:18:00.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:18:00.296 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:19:54.752 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 20:19:55.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:21:49.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-19 20:21:50.362 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:23:47.272 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:30:47.428 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 81, train_steps: 3240 +2026-06-19 20:30:48.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 20:30:48.261 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:30:48.302 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:31:53.315 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:31:53.318 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:33:49.273 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 20:33:50.171 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:35:44.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.70s (1.89min) +2026-06-19 20:35:44.751 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:37:41.675 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:44:40.002 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 82, train_steps: 3280 +2026-06-19 20:44:40.696 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 20:44:40.798 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:44:40.842 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:45:45.423 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:45:45.426 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:47:44.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 20:47:44.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:49:38.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.41s (1.89min) +2026-06-19 20:49:39.274 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:51:38.422 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:58:38.991 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 83, train_steps: 3320 +2026-06-19 20:58:39.739 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 20:58:40.016 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:58:40.122 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:59:44.538 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:59:44.540 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:01:44.304 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 21:01:45.137 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:03:38.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.28s (1.89min) +2026-06-19 21:03:39.339 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:05:38.816 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:12:38.930 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 84, train_steps: 3360 +2026-06-19 21:12:39.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 21:12:39.739 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:12:39.789 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:13:44.458 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:13:44.461 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:15:43.791 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 21:15:44.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:17:38.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.31s (1.89min) +2026-06-19 21:17:38.863 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:19:38.521 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:26:39.497 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 85, train_steps: 3400 +2026-06-19 21:26:40.167 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 21:26:40.270 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:26:40.309 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:27:45.141 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:27:45.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:29:43.697 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 21:29:44.521 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:31:38.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-19 21:31:38.987 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:33:38.754 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:40:38.647 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 86, train_steps: 3440 +2026-06-19 21:40:39.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 21:40:39.549 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:40:39.658 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:41:44.591 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:41:44.591 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:43:42.867 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 21:43:43.693 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:45:37.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 21:45:37.986 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:47:37.707 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:54:39.052 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 87, train_steps: 3480 +2026-06-19 21:54:39.743 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 21:54:39.921 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:54:39.995 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:55:44.454 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:55:44.454 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:57:43.169 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 21:57:44.005 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:59:37.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.39s (1.89min) +2026-06-19 21:59:38.317 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:01:38.553 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:08:39.981 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 88, train_steps: 3520 +2026-06-19 22:08:40.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 22:08:40.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:08:41.026 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:09:45.367 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:09:45.368 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:11:43.755 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 22:11:44.586 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:13:38.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.57s (1.89min) +2026-06-19 22:13:39.103 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:15:38.303 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:22:38.706 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 89, train_steps: 3560 +2026-06-19 22:22:39.424 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 22:22:39.558 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:22:39.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:23:44.316 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:23:44.319 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:25:43.683 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 22:25:44.585 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:27:38.162 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 22:27:38.886 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:29:38.493 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:36:39.446 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 90, train_steps: 3600 +2026-06-19 22:37:15.308 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=90 (missing=0, unexpected=0) +2026-06-19 22:37:17.307 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 22:37:17.437 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:37:17.494 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:38:23.296 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:38:23.300 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:40:19.240 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 22:40:20.098 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:42:14.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.70s (1.89min) +2026-06-19 22:42:14.863 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:44:13.689 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:51:14.855 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 91, train_steps: 3640 +2026-06-19 22:51:15.700 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 22:51:15.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:51:16.001 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:52:20.393 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:52:20.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:54:15.937 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 22:54:16.774 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:56:12.672 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.67s (1.93min) +2026-06-19 22:56:13.387 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:58:13.110 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:05:14.269 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 92, train_steps: 3680 +2026-06-19 23:05:15.118 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 23:05:15.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:05:15.407 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:06:19.760 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:06:19.766 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:08:16.058 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 23:08:16.904 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:10:11.751 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.62s (1.91min) +2026-06-19 23:10:12.470 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:12:11.280 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:19:13.489 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 93, train_steps: 3720 +2026-06-19 23:19:14.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 23:19:14.286 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:19:14.333 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:20:18.934 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:20:18.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:22:14.354 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 23:22:15.373 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:24:10.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.33s (1.92min) +2026-06-19 23:24:11.707 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:26:11.284 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:33:13.960 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 94, train_steps: 3760 +2026-06-19 23:33:14.641 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 23:33:14.835 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:33:14.960 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:34:19.553 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:34:19.553 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:36:16.135 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 23:36:16.960 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:38:12.062 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.89s (1.91min) +2026-06-19 23:38:12.783 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:40:11.319 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:47:11.877 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 95, train_steps: 3800 +2026-06-19 23:47:12.590 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-19 23:47:12.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:47:12.855 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:48:17.098 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:48:17.101 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:50:12.004 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-19 23:50:13.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:52:08.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.82s (1.91min) +2026-06-19 23:52:08.790 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:54:07.675 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:01:09.778 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 96, train_steps: 3840 +2026-06-20 00:01:10.476 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 00:01:10.688 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:01:10.931 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:02:15.437 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:02:15.441 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:04:11.783 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 00:04:12.765 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:06:07.928 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 00:06:08.643 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:08:07.686 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:15:08.506 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 97, train_steps: 3880 +2026-06-20 00:15:09.218 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 00:15:09.322 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:15:09.364 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:16:13.808 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:16:13.811 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:18:11.238 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 00:18:12.150 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:20:07.671 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.20s (1.92min) +2026-06-20 00:20:08.400 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:22:07.124 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:29:07.008 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 98, train_steps: 3920 +2026-06-20 00:29:07.692 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 00:29:07.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:29:08.029 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:30:12.801 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:30:12.801 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:32:10.419 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 00:32:11.252 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:34:06.389 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 00:34:07.117 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:36:03.849 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:43:05.346 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 99, train_steps: 3960 +2026-06-20 00:43:06.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 00:43:06.142 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:43:06.176 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:44:11.651 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:44:11.652 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:46:09.899 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 00:46:10.754 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:48:05.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.75s (1.91min) +2026-06-20 00:48:06.439 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:50:03.396 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:57:04.806 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 100, train_steps: 4000 +2026-06-20 00:58:47.616 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=100 (missing=0, unexpected=0) +2026-06-20 00:58:49.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 00:58:49.790 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:58:49.881 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:59:55.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:59:55.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:01:50.303 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 01:01:51.132 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:03:44.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-20 01:03:45.434 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:05:42.219 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:12:39.886 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 101, train_steps: 4040 +2026-06-20 01:12:40.567 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 01:12:40.664 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:12:40.694 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:13:45.831 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:13:45.831 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:15:43.499 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 01:15:44.501 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:17:38.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.62s (1.89min) +2026-06-20 01:17:38.930 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:19:38.322 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:26:37.806 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 102, train_steps: 4080 +2026-06-20 01:26:38.504 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 01:26:38.676 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:26:38.752 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:27:43.013 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:27:43.016 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:29:41.183 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 01:29:42.465 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:31:36.383 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-20 01:31:37.098 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:33:37.224 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:40:38.539 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 103, train_steps: 4120 +2026-06-20 01:40:39.244 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 01:40:39.361 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:40:39.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:41:44.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:41:44.153 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:43:43.334 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 01:43:44.325 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:45:37.767 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-20 01:45:38.487 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:47:38.263 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:54:37.791 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 104, train_steps: 4160 +2026-06-20 01:54:38.704 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 01:54:38.911 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:54:39.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:55:43.557 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:55:43.557 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:57:42.889 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 01:57:43.889 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:59:37.314 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-20 01:59:38.030 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:01:38.017 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:08:38.456 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 105, train_steps: 4200 +2026-06-20 02:08:39.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 02:08:39.258 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:08:39.297 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:09:43.998 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:09:44.001 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:11:43.945 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 02:11:44.833 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:13:38.217 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 02:13:38.925 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:15:38.266 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:22:37.852 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 106, train_steps: 4240 +2026-06-20 02:22:38.564 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 02:22:38.698 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:22:38.726 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:23:43.368 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:23:43.372 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:25:40.847 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 02:25:41.757 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:27:35.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.90s (1.90min) +2026-06-20 02:27:36.582 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:29:36.810 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:36:37.096 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 107, train_steps: 4280 +2026-06-20 02:36:37.801 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 02:36:37.925 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:36:37.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:37:42.537 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:37:42.540 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:39:39.229 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 02:39:40.227 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:41:35.347 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 02:41:36.071 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:43:35.143 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:50:35.008 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 108, train_steps: 4320 +2026-06-20 02:50:35.727 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 02:50:35.917 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:50:36.016 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:51:40.659 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:51:40.660 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:53:37.573 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 02:53:38.406 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:55:33.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.45s (1.91min) +2026-06-20 02:55:33.972 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:57:30.646 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:04:30.701 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 109, train_steps: 4360 +2026-06-20 03:04:31.406 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 03:04:31.615 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:04:31.752 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:05:36.097 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:05:36.100 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:07:33.407 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 03:07:34.387 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:09:29.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-20 03:09:30.014 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:11:26.573 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:18:25.352 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 110, train_steps: 4400 +2026-06-20 03:19:01.541 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=110 (missing=0, unexpected=0) +2026-06-20 03:19:03.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 03:19:03.367 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:19:03.387 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:20:11.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:20:11.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:22:07.524 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 03:22:08.474 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:24:03.966 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.41s (1.92min) +2026-06-20 03:24:04.689 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:26:02.471 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:33:05.395 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 111, train_steps: 4440 +2026-06-20 03:33:06.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 03:33:06.172 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:33:06.215 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:34:11.826 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:34:11.826 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:36:10.098 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 03:36:11.090 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:38:06.021 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 03:38:06.731 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:40:04.482 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:47:04.316 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 112, train_steps: 4480 +2026-06-20 03:47:05.043 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 03:47:05.207 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:47:05.278 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:48:10.751 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:48:10.751 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:50:04.983 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 03:50:06.252 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:52:00.943 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-20 03:52:01.657 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:54:00.961 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:01:04.009 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 113, train_steps: 4520 +2026-06-20 04:01:04.874 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 04:01:04.972 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:01:05.019 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:02:10.092 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:02:10.096 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:04:06.611 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 04:04:07.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:06:02.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.15s (1.92min) +2026-06-20 04:06:03.532 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:08:01.584 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:15:03.784 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 114, train_steps: 4560 +2026-06-20 04:15:04.473 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 04:15:04.592 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:15:04.637 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:16:09.744 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:16:09.744 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:18:04.152 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 04:18:05.448 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:20:00.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.72s (1.91min) +2026-06-20 04:20:00.894 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:22:00.097 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:29:01.468 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 115, train_steps: 4600 +2026-06-20 04:29:02.193 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 04:29:02.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:29:02.489 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:30:06.908 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:30:06.911 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:32:02.494 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 04:32:03.487 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:33:58.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.33s (1.91min) +2026-06-20 04:33:58.775 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:35:55.624 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:42:56.470 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 116, train_steps: 4640 +2026-06-20 04:42:57.159 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 04:42:57.286 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:42:57.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:44:01.983 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:44:01.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:45:58.380 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 04:45:59.278 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:47:53.698 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.25s (1.90min) +2026-06-20 04:47:54.407 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:49:50.887 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:56:50.236 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 117, train_steps: 4680 +2026-06-20 04:56:50.953 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 04:56:51.139 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:56:51.229 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:57:55.699 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:57:55.705 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:59:51.798 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 04:59:52.639 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:01:46.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.49s (1.89min) +2026-06-20 05:01:47.078 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:03:44.147 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:10:42.595 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 118, train_steps: 4720 +2026-06-20 05:10:43.296 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 05:10:43.465 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:10:43.545 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:11:48.121 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:11:48.124 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:13:46.957 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 05:13:47.810 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:15:41.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.26s (1.89min) +2026-06-20 05:15:42.000 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:17:38.993 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:24:37.905 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 119, train_steps: 4760 +2026-06-20 05:24:38.622 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 05:24:38.755 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:24:38.796 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:25:43.355 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:25:43.358 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:27:40.119 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 05:27:41.109 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:29:34.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-20 05:29:35.559 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:31:33.104 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:38:32.154 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 120, train_steps: 4800 +2026-06-20 05:39:08.637 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=120 (missing=0, unexpected=0) +2026-06-20 05:39:10.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 05:39:11.050 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:39:11.115 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:40:17.607 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:40:17.611 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:42:12.566 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 05:42:13.398 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:44:08.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.96s (1.92min) +2026-06-20 05:44:09.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:46:10.086 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:53:12.067 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 121, train_steps: 4840 +2026-06-20 05:53:12.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 05:53:12.909 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:53:12.945 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:54:17.902 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:54:17.905 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:56:13.156 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 05:56:14.016 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:58:09.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.97s (1.92min) +2026-06-20 05:58:10.075 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:00:10.575 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:07:11.955 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 122, train_steps: 4880 +2026-06-20 06:07:12.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 06:07:12.789 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:07:12.827 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:08:17.216 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:08:17.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:10:13.195 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 06:10:14.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:12:08.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.25s (1.90min) +2026-06-20 06:12:09.325 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:14:10.579 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:21:11.875 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 123, train_steps: 4920 +2026-06-20 06:21:12.561 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 06:21:12.726 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:21:12.786 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:22:17.077 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:22:17.080 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:24:12.071 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 06:24:13.113 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:26:08.788 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.47s (1.92min) +2026-06-20 06:26:09.511 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:28:09.832 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:35:11.851 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 124, train_steps: 4960 +2026-06-20 06:35:12.537 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 06:35:12.633 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:35:12.671 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:36:16.929 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:36:16.932 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:38:12.129 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 06:38:13.045 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:40:08.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.44s (1.92min) +2026-06-20 06:40:09.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:42:09.771 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:49:11.652 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 125, train_steps: 5000 +2026-06-20 06:50:15.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 06:50:15.353 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:50:15.417 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:51:19.869 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:51:19.872 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:53:15.926 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 06:53:16.749 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:55:11.654 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.68s (1.91min) +2026-06-20 06:55:12.371 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:57:11.762 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:04:11.921 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 126, train_steps: 5040 +2026-06-20 07:04:12.630 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 07:04:12.773 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:04:12.817 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:05:17.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:05:17.602 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:07:13.988 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 07:07:15.181 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:09:10.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.08s (1.92min) +2026-06-20 07:09:11.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:11:11.163 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:18:11.931 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 127, train_steps: 5080 +2026-06-20 07:18:12.629 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 07:18:12.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:18:12.780 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:19:17.341 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:19:17.341 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:21:13.050 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 07:21:13.945 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:23:09.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.50s (1.93min) +2026-06-20 07:23:10.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:25:10.157 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:32:11.172 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 128, train_steps: 5120 +2026-06-20 07:32:11.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 07:32:12.004 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:32:12.045 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:33:16.493 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:33:16.494 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:35:13.016 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 07:35:14.007 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:37:08.641 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.56s (1.91min) +2026-06-20 07:37:09.372 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:39:09.928 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:46:11.989 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 129, train_steps: 5160 +2026-06-20 07:46:12.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 07:46:12.875 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:46:12.934 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:47:17.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:47:17.719 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:49:13.070 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 07:49:14.055 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:51:09.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-20 07:51:09.758 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:53:10.302 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:00:11.786 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 130, train_steps: 5200 +2026-06-20 08:00:49.500 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=130 (missing=0, unexpected=0) +2026-06-20 08:00:51.244 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 08:00:51.330 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:00:51.421 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:01:58.698 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:01:58.701 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:03:52.509 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 08:03:53.275 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:05:47.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.72s (1.90min) +2026-06-20 08:05:48.018 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:07:44.842 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:14:45.322 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 131, train_steps: 5240 +2026-06-20 08:14:46.025 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 08:14:46.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:14:46.303 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:15:51.043 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:15:51.047 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:17:46.989 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 08:17:47.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:19:41.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.38s (1.89min) +2026-06-20 08:19:42.139 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:21:39.466 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:28:37.829 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 132, train_steps: 5280 +2026-06-20 08:28:38.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 08:28:38.658 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:28:38.692 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:29:43.382 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:29:43.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:31:40.849 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 08:31:41.763 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:33:35.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.44s (1.89min) +2026-06-20 08:33:36.118 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:35:35.261 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:42:34.335 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 133, train_steps: 5320 +2026-06-20 08:42:35.018 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 08:42:35.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:42:35.324 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:43:39.892 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:43:39.899 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:45:35.888 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 08:45:36.727 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:47:32.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.77s (1.93min) +2026-06-20 08:47:33.449 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:49:30.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:56:29.440 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 134, train_steps: 5360 +2026-06-20 08:56:30.113 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 08:56:30.206 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:56:30.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:57:35.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:57:35.102 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:59:32.040 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 08:59:32.880 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:01:26.319 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 09:01:27.047 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:03:23.804 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:10:22.991 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 135, train_steps: 5400 +2026-06-20 09:10:23.673 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 09:10:23.788 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:10:23.829 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:11:28.658 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:11:28.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:13:23.932 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 09:13:24.781 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:15:18.588 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.61s (1.89min) +2026-06-20 09:15:19.310 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:17:17.100 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:24:16.806 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 136, train_steps: 5440 +2026-06-20 09:24:17.531 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 09:24:17.668 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:24:17.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:25:22.249 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:25:22.252 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:27:18.656 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 09:27:19.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:29:12.958 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.15s (1.89min) +2026-06-20 09:29:13.686 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:31:13.175 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:38:13.227 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 137, train_steps: 5480 +2026-06-20 09:38:13.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 09:38:14.027 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:38:14.076 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:39:18.676 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:39:18.683 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:41:14.371 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 09:41:15.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:43:09.522 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-20 09:43:10.250 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:45:10.502 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:52:12.154 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 138, train_steps: 5520 +2026-06-20 09:52:12.866 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 09:52:13.005 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:52:13.039 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:53:17.675 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:53:17.678 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:55:13.383 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 09:55:14.429 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:57:10.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.46s (1.92min) +2026-06-20 09:57:10.920 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:59:11.150 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:06:12.676 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 139, train_steps: 5560 +2026-06-20 10:06:13.363 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 10:06:13.491 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:06:13.533 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:07:17.974 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:07:17.977 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:09:13.264 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 10:09:14.092 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:11:08.675 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.20s (1.90min) +2026-06-20 10:11:09.397 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:13:09.778 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:20:12.680 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 140, train_steps: 5600 +2026-06-20 10:20:50.192 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=140 (missing=0, unexpected=0) +2026-06-20 10:20:51.812 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 10:20:51.939 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:20:51.970 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:21:58.333 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:21:58.336 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:23:53.040 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 10:23:53.996 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:25:47.600 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-20 10:25:48.313 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:27:45.362 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:34:44.001 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 141, train_steps: 5640 +2026-06-20 10:34:44.690 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 10:34:44.908 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:34:45.046 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:35:49.483 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:35:49.486 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:37:46.846 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 10:37:47.834 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:39:41.415 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.51s (1.89min) +2026-06-20 10:39:42.147 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:41:39.461 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:48:38.465 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 142, train_steps: 5680 +2026-06-20 10:48:39.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 10:48:39.303 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:48:39.354 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:49:44.095 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:49:44.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:51:43.153 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 10:51:44.050 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:53:37.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-20 10:53:37.851 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:55:37.991 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:02:37.920 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 143, train_steps: 5720 +2026-06-20 11:02:38.631 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 11:02:38.749 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:02:38.780 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:03:43.746 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:03:43.746 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:05:42.169 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 11:05:43.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:07:36.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 11:07:37.312 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:09:36.840 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:16:37.095 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 144, train_steps: 5760 +2026-06-20 11:16:37.798 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 11:16:37.939 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:16:37.974 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:17:42.529 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:17:42.531 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:19:38.707 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 11:19:39.779 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:21:33.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.47s (1.89min) +2026-06-20 11:21:34.087 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:23:30.883 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:30:30.396 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 145, train_steps: 5800 +2026-06-20 11:30:31.096 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 11:30:31.231 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:30:31.271 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:31:35.632 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:31:35.635 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:33:33.207 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 11:33:34.054 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:35:28.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.89s (1.90min) +2026-06-20 11:35:28.899 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:37:25.686 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:44:23.925 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 146, train_steps: 5840 +2026-06-20 11:44:24.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 11:44:24.733 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:44:24.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:45:29.258 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:45:29.260 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:47:23.247 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 11:47:24.093 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:49:19.133 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-20 11:49:19.855 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:51:17.037 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:58:15.930 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 147, train_steps: 5880 +2026-06-20 11:58:16.617 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 11:58:16.788 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:58:16.872 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:59:21.806 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:59:21.816 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:01:16.232 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 12:01:17.234 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:03:12.586 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.29s (1.92min) +2026-06-20 12:03:13.293 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:05:12.902 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:12:12.699 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 148, train_steps: 5920 +2026-06-20 12:12:13.404 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 12:12:13.544 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:12:13.575 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:13:18.269 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:13:18.271 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:15:13.144 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 12:15:14.104 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:17:09.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.69s (1.91min) +2026-06-20 12:17:09.788 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:19:10.586 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:26:11.387 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 149, train_steps: 5960 +2026-06-20 12:26:12.136 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 12:26:12.327 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:26:12.419 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:27:16.744 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:27:16.747 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:29:13.029 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 12:29:14.014 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:31:08.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.72s (1.91min) +2026-06-20 12:31:09.533 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:33:09.830 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:40:10.984 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 150, train_steps: 6000 +2026-06-20 12:41:53.494 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=150 (missing=0, unexpected=0) +2026-06-20 12:41:56.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 12:41:56.341 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:41:56.449 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:43:02.584 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:43:02.587 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:44:59.744 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 12:45:00.746 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:46:55.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.70s (1.91min) +2026-06-20 12:46:56.235 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:48:52.986 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:55:51.603 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 151, train_steps: 6040 +2026-06-20 12:55:52.305 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 12:55:52.405 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:55:52.450 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:56:56.830 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:56:56.833 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:58:50.721 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 12:58:51.703 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:00:46.189 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.41s (1.91min) +2026-06-20 13:00:46.907 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:02:43.650 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:09:41.902 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 152, train_steps: 6080 +2026-06-20 13:09:42.596 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 13:09:42.723 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:09:42.757 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:10:47.419 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:10:47.425 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:12:43.910 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 13:12:44.910 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:14:38.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-20 13:14:39.491 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:16:38.547 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:23:37.213 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 153, train_steps: 6120 +2026-06-20 13:23:38.092 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 13:23:38.221 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:23:38.260 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:24:43.275 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:24:43.275 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:26:40.127 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 13:26:41.277 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:28:34.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.39s (1.89min) +2026-06-20 13:28:35.472 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:30:32.919 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:37:31.696 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 154, train_steps: 6160 +2026-06-20 13:37:32.394 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 13:37:32.573 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:37:32.649 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:38:37.029 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:38:37.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:40:33.788 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 13:40:34.762 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:42:29.329 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.47s (1.91min) +2026-06-20 13:42:30.061 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:44:27.161 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:51:24.974 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 155, train_steps: 6200 +2026-06-20 13:51:25.699 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 13:51:25.818 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:51:25.856 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:52:30.474 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:52:30.477 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:54:24.447 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 13:54:25.421 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:56:19.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.63s (1.89min) +2026-06-20 13:56:19.859 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:58:16.747 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:05:15.341 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 156, train_steps: 6240 +2026-06-20 14:05:16.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 14:05:16.129 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:05:16.175 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:06:21.063 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:06:21.067 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:08:15.931 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 14:08:16.914 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:10:11.722 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.74s (1.91min) +2026-06-20 14:10:12.456 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:12:11.102 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:19:12.239 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 157, train_steps: 6280 +2026-06-20 14:19:12.923 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 14:19:13.042 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:19:13.082 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:20:17.693 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:20:17.696 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:22:12.398 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 14:22:13.497 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:24:08.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.22s (1.92min) +2026-06-20 14:24:09.689 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:26:09.902 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:33:10.827 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 158, train_steps: 6320 +2026-06-20 14:33:11.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 14:33:11.632 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:33:11.675 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:34:16.228 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:34:16.228 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:36:10.935 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 14:36:12.011 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:38:07.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.88s (1.91min) +2026-06-20 14:38:07.784 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:40:05.255 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:47:04.567 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 159, train_steps: 6360 +2026-06-20 14:47:05.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 14:47:05.435 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:47:05.471 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:48:10.753 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:48:10.753 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:50:04.048 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 14:50:05.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:51:59.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.86s (1.90min) +2026-06-20 14:51:59.928 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:53:57.320 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:00:59.183 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 160, train_steps: 6400 +2026-06-20 15:01:37.364 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=160 (missing=0, unexpected=0) +2026-06-20 15:01:39.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 15:01:39.480 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:01:39.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:02:45.175 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:02:45.178 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:04:42.879 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 15:04:43.722 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:06:37.408 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.48s (1.89min) +2026-06-20 15:06:38.136 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:08:38.302 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:15:37.526 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 161, train_steps: 6440 +2026-06-20 15:15:38.207 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 15:15:38.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:15:38.415 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:16:42.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:16:42.885 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:18:40.375 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 15:18:41.402 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:20:35.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-20 15:20:36.331 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:22:36.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:29:36.898 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 162, train_steps: 6480 +2026-06-20 15:29:37.570 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 15:29:37.749 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:29:37.836 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:30:42.290 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:30:42.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:32:37.444 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 15:32:38.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:34:33.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.48s (1.91min) +2026-06-20 15:34:33.826 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:36:31.315 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:43:31.483 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 163, train_steps: 6520 +2026-06-20 15:43:32.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 15:43:32.380 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:43:32.507 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:44:36.925 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:44:36.932 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:46:33.676 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 15:46:34.663 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:48:29.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.47s (1.91min) +2026-06-20 15:48:29.934 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:50:26.535 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:57:25.503 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 164, train_steps: 6560 +2026-06-20 15:57:26.227 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 15:57:26.325 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:57:26.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:58:31.004 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:58:31.005 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:00:26.401 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 16:00:27.371 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:02:21.064 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.61s (1.89min) +2026-06-20 16:02:21.792 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:04:18.346 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:11:18.377 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 165, train_steps: 6600 +2026-06-20 16:11:19.085 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 16:11:19.180 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:11:19.224 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:12:24.191 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:12:24.193 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:14:19.105 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 16:14:20.206 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:16:14.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.75s (1.90min) +2026-06-20 16:16:14.869 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:18:14.123 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:25:13.940 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 166, train_steps: 6640 +2026-06-20 16:25:14.618 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 16:25:14.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:25:14.757 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:26:19.742 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:26:19.742 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:28:13.831 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 16:28:14.916 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:30:08.844 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.78s (1.90min) +2026-06-20 16:30:09.574 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:32:10.304 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:39:12.154 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 167, train_steps: 6680 +2026-06-20 16:39:12.849 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 16:39:13.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:39:13.093 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:40:17.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:40:17.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:42:12.095 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 16:42:13.332 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:44:08.745 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.36s (1.92min) +2026-06-20 16:44:09.476 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:46:10.018 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:53:11.386 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 168, train_steps: 6720 +2026-06-20 16:53:12.092 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 16:53:12.281 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:53:12.385 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:54:16.892 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:54:16.892 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:56:11.767 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 16:56:12.630 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:58:07.736 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-20 16:58:08.463 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:00:07.579 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:07:09.193 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 169, train_steps: 6760 +2026-06-20 17:07:09.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 17:07:09.994 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:07:10.061 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:08:14.791 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:08:14.797 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:10:11.903 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 17:10:12.763 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:12:08.254 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.20s (1.92min) +2026-06-20 17:12:08.968 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:14:09.317 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:21:11.118 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 170, train_steps: 6800 +2026-06-20 17:21:47.210 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=170 (missing=0, unexpected=0) +2026-06-20 17:21:49.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 17:21:49.446 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:21:49.588 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:22:54.828 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:22:54.831 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:24:49.925 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 17:24:50.777 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:26:44.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.05s (1.88min) +2026-06-20 17:26:44.743 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:28:41.424 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:35:39.738 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 21, sp_rank 0: update step: 171, train_steps: 6840 +2026-06-20 17:35:40.431 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 21, dp_rank 10: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=5, rank_in_group=0 +2026-06-20 17:35:40.548 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:35:40.617 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:36:45.542 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 21 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:36:45.546 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 21 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:38:42.537 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 21: Using local metric names without global sync: ['avg'] +2026-06-20 17:38:43.560 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 21: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:40:36.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 21: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.89s (1.88min) +2026-06-20 17:40:37.377 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 21, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-29 01:28:17.254 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 01:28:17.254 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 01:52:51.688 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 01:52:51.688 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 02:20:51.228 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 02:20:51.230 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 02:46:50.618 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 02:46:50.619 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:06:52.474 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:06:52.474 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:35:13.945 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:35:13.946 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:58:06.831 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:58:06.832 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:27:49.873 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:27:49.874 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:51:26.971 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:51:26.973 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 05:15:18.658 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 05:15:18.659 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 05:38:45.767 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 05:38:45.769 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:02:37.806 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:02:37.807 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:24:20.268 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:24:20.268 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:49:03.938 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:49:03.939 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 07:12:16.571 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 07:12:16.572 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 07:35:01.510 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 07:35:01.510 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:04:31.369 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:04:31.369 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:33:44.001 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:33:44.001 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:00:59.787 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:00:59.787 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:22:25.897 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:22:25.897 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:47:19.922 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:47:19.923 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 10:09:58.422 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 10:09:58.422 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 10:37:07.729 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 10:37:07.730 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:02:16.838 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:02:16.839 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:28:19.880 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:28:19.881 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:56:03.931 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:56:03.932 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 12:18:29.685 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 12:18:29.686 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 12:46:12.805 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 12:46:12.806 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 13:13:00.572 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 13:13:00.572 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 13:35:55.426 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 13:35:55.426 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:04:07.872 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:04:07.873 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:27:03.836 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:27:03.837 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:49:03.140 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:49:03.142 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 15:23:13.369 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 15:23:13.370 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 15:50:22.280 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 15:50:22.280 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 16:15:40.909 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 16:15:40.910 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 16:43:17.790 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 16:43:17.791 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:11:03.084 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:11:03.085 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:32:10.193 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:32:10.194 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:53:41.029 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:53:41.031 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 18:17:17.599 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 18:17:17.600 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 18:38:51.258 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 18:38:51.259 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_26.log b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_26.log new file mode 100644 index 0000000000000000000000000000000000000000..66d69dda47556bdb2a1662bbc1c8d857e10e9b9d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_26.log @@ -0,0 +1,3241 @@ +2026-06-19 01:22:18.211 | INFO | __main__:__init__:933 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-19 01:22:18.890 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-19 01:22:18.891 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-19 01:30:17.294 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 01:30:31.482 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 01:30:31.485 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:30:31.487 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:31:52.639 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_6/index_2282_seed_5650_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 01:31:57.595 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 01:32:15.073 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:32:15.081 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:32:15.081 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:33:22.249 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_6/index_2282_seed_5651_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 01:33:26.052 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 01:33:43.448 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:33:45.515 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 01:34:00.983 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 01:35:56.997 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.94s (1.93min) +2026-06-19 01:35:57.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 01:37:54.182 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 01:37:57.220 | INFO | hyvideo.utils.grpo_utils:train_one_step:3127 - Rank 26, sp_rank 0: init adaptive KL controller (init=1.000e-07, min=1.000e-07, max=1.000e-05) +2026-06-19 01:45:11.974 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 1, train_steps: 40 +2026-06-19 01:45:12.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 01:45:12.675 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 01:45:12.690 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:45:12.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:46:18.632 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000001/group_6/index_5897_seed_5252_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 01:46:21.472 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 01:46:40.428 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:46:40.435 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:46:40.435 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:47:47.693 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000001/group_6/index_5897_seed_5253_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 01:47:50.605 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 01:48:08.390 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:48:10.945 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 01:48:11.995 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 01:50:07.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.70s (1.93min) +2026-06-19 01:50:08.466 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 01:52:06.954 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 01:59:06.790 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 2, train_steps: 80 +2026-06-19 01:59:07.381 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 01:59:07.516 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 01:59:07.555 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:59:07.619 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:00:12.941 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000002/group_6/index_912_seed_17519_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 02:00:17.969 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 02:00:35.325 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:00:35.333 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:00:35.334 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:01:46.251 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000002/group_6/index_912_seed_17520_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 02:01:49.939 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 02:02:07.945 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:02:10.528 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:02:11.694 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:04:06.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.88s (1.91min) +2026-06-19 02:04:07.374 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:06:03.753 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:13:04.860 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 3, train_steps: 120 +2026-06-19 02:13:05.465 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 02:13:05.580 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:13:05.597 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:13:05.641 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:14:11.255 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_6/index_1434_seed_11439_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 02:14:16.169 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 02:14:33.015 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:14:33.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:14:33.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:15:40.476 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_6/index_1434_seed_11440_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 02:15:45.241 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 02:16:02.487 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:16:04.655 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:16:05.758 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:17:59.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.77s (1.90min) +2026-06-19 02:18:00.448 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:19:59.349 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:26:59.880 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 4, train_steps: 160 +2026-06-19 02:27:00.547 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 02:27:00.767 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:27:00.913 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:27:01.007 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:28:06.438 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000004/group_6/index_5817_seed_23060_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 02:28:11.176 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 02:28:28.006 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:28:28.015 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:28:28.015 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:29:35.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000004/group_6/index_5817_seed_23061_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 02:29:39.218 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 02:29:58.075 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:30:00.316 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:30:01.310 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:31:55.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.62s (1.91min) +2026-06-19 02:31:56.651 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:33:53.222 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:40:53.221 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 5, train_steps: 200 +2026-06-19 02:40:53.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 02:40:54.035 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:40:54.070 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:40:54.160 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:41:59.982 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_6/index_143_seed_28537_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 02:42:02.698 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 02:42:20.038 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:42:20.047 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:42:20.047 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:43:28.515 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_6/index_143_seed_28538_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 02:43:31.143 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 02:43:48.590 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:43:51.812 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:43:52.771 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:45:46.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.50s (1.89min) +2026-06-19 02:45:46.993 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:47:43.352 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:54:40.651 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 6, train_steps: 240 +2026-06-19 02:54:41.360 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 02:54:41.494 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:54:41.532 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:54:41.655 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:55:47.787 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_6/index_7633_seed_12740_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 02:55:50.488 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 02:56:07.775 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:56:07.784 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:56:07.784 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:57:19.442 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_6/index_7633_seed_12741_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 02:57:24.488 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 02:57:41.643 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:57:42.949 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:57:43.850 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:59:37.354 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.38s (1.89min) +2026-06-19 02:59:38.043 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:01:37.728 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:08:37.484 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 7, train_steps: 280 +2026-06-19 03:08:38.148 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 03:08:38.361 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:08:38.493 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:08:38.570 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:09:43.908 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_6/index_7957_seed_26759_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 03:09:48.766 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 03:10:06.058 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:10:06.067 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:10:06.067 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:11:14.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_6/index_7957_seed_26760_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 03:11:19.463 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 03:11:36.881 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:11:38.337 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:11:39.597 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:13:34.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.68s (1.91min) +2026-06-19 03:13:34.976 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:15:32.041 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:22:31.145 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 8, train_steps: 320 +2026-06-19 03:22:31.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 03:22:31.978 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:22:31.995 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:22:32.107 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:23:38.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000008/group_6/index_5420_seed_27395_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 03:23:42.701 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 03:24:00.466 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:24:00.474 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:24:00.475 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:25:09.097 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000008/group_6/index_5420_seed_27396_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 03:25:14.160 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 03:25:31.458 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:25:33.100 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:25:34.010 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:27:28.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-19 03:27:29.431 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:29:26.903 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:36:25.650 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 9, train_steps: 360 +2026-06-19 03:36:26.347 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 03:36:26.531 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:36:26.628 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:36:26.723 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:37:32.557 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_6/index_3172_seed_17183_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 03:37:36.314 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 03:37:54.067 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:37:54.076 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:37:54.076 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:39:01.938 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_6/index_3172_seed_17184_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 03:39:06.982 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 03:39:24.446 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:39:26.010 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:39:26.959 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:41:20.400 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.36s (1.89min) +2026-06-19 03:41:21.105 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:43:17.559 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:50:16.945 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 10, train_steps: 400 +2026-06-19 03:50:53.313 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=10 (missing=0, unexpected=0) +2026-06-19 03:50:56.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 03:50:56.283 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:50:56.318 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:50:56.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:52:03.406 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_6/index_9033_seed_29281_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 03:52:07.234 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 03:52:23.467 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:52:23.476 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:52:23.477 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:53:32.432 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_6/index_9033_seed_29282_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 03:53:37.169 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 03:53:53.697 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:53:57.828 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:53:58.743 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:55:52.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.65s (1.89min) +2026-06-19 03:55:53.210 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:57:49.762 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:04:48.374 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 11, train_steps: 440 +2026-06-19 04:04:49.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 04:04:49.287 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:04:49.399 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:04:49.476 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:05:55.107 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_6/index_9751_seed_14506_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 04:05:59.837 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 04:06:15.951 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:06:15.958 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:06:15.958 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:07:24.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_6/index_9751_seed_14507_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 04:07:28.552 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 04:07:45.464 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:07:48.853 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:07:49.847 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:09:43.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.25s (1.89min) +2026-06-19 04:09:43.797 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:11:40.267 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:18:37.299 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 12, train_steps: 480 +2026-06-19 04:18:37.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 04:18:38.186 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:18:38.328 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:18:38.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:19:44.120 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_6/index_804_seed_22854_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 04:19:48.834 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 04:20:04.836 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:20:04.846 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:20:04.846 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:21:12.855 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_6/index_804_seed_22855_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 04:21:17.566 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 04:21:34.022 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:21:37.180 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:21:38.113 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:23:33.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.19s (1.92min) +2026-06-19 04:23:34.070 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:25:31.242 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:32:29.814 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 13, train_steps: 520 +2026-06-19 04:32:30.512 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 04:32:30.618 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:32:30.666 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:32:30.714 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:33:36.308 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_6/index_690_seed_465_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 04:33:41.319 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 04:33:57.386 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:33:57.395 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:33:57.395 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:35:05.934 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_6/index_690_seed_466_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 04:35:09.637 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 04:35:26.647 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:35:30.313 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:35:31.318 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:37:24.102 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.77s (1.88min) +2026-06-19 04:37:24.792 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:39:21.077 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:46:18.591 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 14, train_steps: 560 +2026-06-19 04:46:19.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 04:46:19.389 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:46:19.425 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:46:19.514 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:47:25.750 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_6/index_7955_seed_31118_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 04:47:28.544 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 04:47:44.870 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:47:44.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:47:44.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:48:54.181 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_6/index_7955_seed_31119_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 04:48:57.231 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 04:49:14.165 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:49:17.445 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:49:18.420 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:51:12.963 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.49s (1.91min) +2026-06-19 04:51:13.650 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:53:12.993 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:00:11.828 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 15, train_steps: 600 +2026-06-19 05:00:12.525 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 05:00:12.662 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:00:12.696 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:00:12.787 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:01:18.669 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000015/group_6/index_568_seed_12215_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 05:01:23.607 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 05:01:40.822 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:01:40.829 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:01:40.829 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:02:47.790 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000015/group_6/index_568_seed_12216_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 05:02:51.567 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 05:03:08.050 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:03:11.349 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:03:12.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:05:07.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.64s (1.91min) +2026-06-19 05:05:08.061 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:07:05.978 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:14:04.904 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 16, train_steps: 640 +2026-06-19 05:14:05.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 05:14:05.759 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:14:05.865 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:14:05.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:15:11.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_6/index_1808_seed_21466_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 05:15:14.992 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 05:15:31.437 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:15:31.446 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:15:31.447 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:16:42.903 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_6/index_1808_seed_21467_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 05:16:47.387 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 05:17:04.329 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:17:06.988 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:17:07.939 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:19:03.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.02s (1.92min) +2026-06-19 05:19:03.742 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:21:01.960 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:28:03.320 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 17, train_steps: 680 +2026-06-19 05:28:04.007 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 05:28:04.213 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:28:04.313 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:28:04.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:29:10.388 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_6/index_1286_seed_7809_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 05:29:15.261 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 05:29:31.658 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:29:31.668 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:29:31.668 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:30:39.402 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_6/index_1286_seed_7810_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 05:30:43.002 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 05:30:59.796 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:31:03.926 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:31:05.041 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:32:58.689 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.48s (1.89min) +2026-06-19 05:32:59.387 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:34:56.264 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:41:58.171 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 18, train_steps: 720 +2026-06-19 05:41:58.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 05:41:58.929 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:41:58.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:41:59.089 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:43:05.129 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_6/index_7019_seed_1460_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 05:43:10.049 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 05:43:26.190 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:43:26.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:43:26.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:44:35.541 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_6/index_7019_seed_1461_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 05:44:40.442 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 05:44:56.669 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:45:00.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:45:01.141 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:46:55.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.88s (1.90min) +2026-06-19 05:46:55.797 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:48:52.329 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:55:50.489 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 19, train_steps: 760 +2026-06-19 05:55:51.178 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 05:55:51.388 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:55:51.517 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:55:51.593 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:56:57.117 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_6/index_8767_seed_30086_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 05:57:00.925 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 05:57:17.554 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:57:17.565 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:57:17.565 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:58:25.482 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_6/index_8767_seed_30087_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 05:58:30.416 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 05:58:47.245 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:58:49.068 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:58:49.982 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:00:43.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.10s (1.89min) +2026-06-19 06:00:43.876 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:02:40.340 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:09:39.758 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 20, train_steps: 800 +2026-06-19 06:10:17.413 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=20 (missing=0, unexpected=0) +2026-06-19 06:10:18.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 06:10:18.938 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:10:19.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:10:19.104 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:11:26.388 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_6/index_9698_seed_14950_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 06:11:30.112 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 06:11:46.599 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:11:46.609 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:11:46.609 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:12:54.562 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_6/index_9698_seed_14951_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 06:12:57.578 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 06:13:14.455 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:13:19.050 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:13:20.177 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:15:13.224 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.90s (1.88min) +2026-06-19 06:15:13.910 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:17:12.688 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:24:12.562 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 21, train_steps: 840 +2026-06-19 06:24:13.271 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 06:24:13.452 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:24:13.532 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:24:13.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:25:19.221 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_6/index_7112_seed_14918_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 06:25:24.072 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 06:25:40.401 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:25:40.411 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:25:40.411 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:26:47.746 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_6/index_7112_seed_14919_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 06:26:52.462 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 06:27:09.089 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:27:11.261 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:27:12.275 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:29:06.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.51s (1.91min) +2026-06-19 06:29:07.668 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:31:04.578 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:38:04.277 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 22, train_steps: 880 +2026-06-19 06:38:04.948 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 06:38:05.113 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:38:05.198 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:38:05.302 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:39:10.960 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_6/index_8300_seed_14647_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 06:39:14.788 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 06:39:31.393 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:39:31.402 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:39:31.403 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:40:39.752 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_6/index_8300_seed_14648_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 06:40:42.735 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 06:40:59.662 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:41:03.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:41:04.319 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:42:58.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.99s (1.90min) +2026-06-19 06:42:59.165 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:44:56.505 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:51:58.171 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 23, train_steps: 920 +2026-06-19 06:51:58.843 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 06:51:59.048 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:51:59.196 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:51:59.289 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:53:04.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000023/group_6/index_7364_seed_13225_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 06:53:07.886 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 06:53:24.595 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:53:24.604 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:53:24.605 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:54:34.936 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000023/group_6/index_7364_seed_13226_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 06:54:40.158 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 06:54:57.483 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:54:59.789 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:55:00.750 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:56:54.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.11s (1.90min) +2026-06-19 06:56:55.622 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:58:52.319 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:05:51.320 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 24, train_steps: 960 +2026-06-19 07:05:52.004 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 07:05:52.115 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:05:52.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:05:52.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:06:58.135 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_6/index_1762_seed_1223_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 07:07:02.919 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 07:07:19.624 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:07:19.633 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:07:19.633 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:08:27.630 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_6/index_1762_seed_1224_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 07:08:31.399 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 07:08:48.625 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:08:52.217 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:08:53.196 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:10:46.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-19 07:10:46.871 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:12:44.153 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:19:41.979 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 25, train_steps: 1000 +2026-06-19 07:20:52.105 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 07:20:52.297 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:20:52.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:20:52.480 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:21:58.387 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_6/index_7585_seed_7158_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 07:22:01.392 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 07:22:18.645 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:22:18.655 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:22:18.655 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:23:28.840 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_6/index_7585_seed_7159_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 07:23:32.557 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 07:23:49.988 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:23:52.666 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:23:53.653 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:25:46.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.81s (1.88min) +2026-06-19 07:25:47.206 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:27:44.380 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:34:43.688 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 26, train_steps: 1040 +2026-06-19 07:34:44.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 07:34:44.520 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:34:44.593 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:34:44.711 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:35:50.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_6/index_6973_seed_12051_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 07:35:55.491 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 07:36:12.100 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:36:12.109 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:36:12.109 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:37:20.981 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_6/index_6973_seed_12052_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 07:37:25.621 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 07:37:42.522 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:37:45.443 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:37:46.378 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:39:39.639 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.14s (1.89min) +2026-06-19 07:39:40.348 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:41:39.070 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:48:39.671 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 27, train_steps: 1080 +2026-06-19 07:48:40.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 07:48:40.508 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:48:40.599 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:48:40.702 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:49:47.852 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_6/index_544_seed_3827_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 07:49:52.494 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 07:50:09.062 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:50:09.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:50:09.074 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:51:19.251 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_6/index_544_seed_3828_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 07:51:23.850 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 07:51:40.846 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:51:43.285 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:51:44.178 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:53:38.298 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.99s (1.90min) +2026-06-19 07:53:39.007 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:55:38.422 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:02:38.168 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 28, train_steps: 1120 +2026-06-19 08:02:38.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 08:02:39.063 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:02:39.165 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:02:39.267 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:03:45.001 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_6/index_8124_seed_15930_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 08:03:48.953 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 08:04:06.093 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:04:06.106 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:04:06.106 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:05:17.391 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_6/index_8124_seed_15931_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 08:05:22.316 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 08:05:40.137 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:05:41.901 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:05:43.000 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:07:37.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.24s (1.90min) +2026-06-19 08:07:38.092 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:09:37.932 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:16:37.855 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 29, train_steps: 1160 +2026-06-19 08:16:38.692 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 08:16:38.900 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:16:39.201 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:16:39.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:17:44.902 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_6/index_7400_seed_16051_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 08:17:48.759 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 08:18:05.961 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:18:05.971 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:18:05.971 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:19:17.877 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_6/index_7400_seed_16052_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 08:19:20.878 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 08:19:38.622 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:19:42.881 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:19:43.883 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:21:37.264 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 08:21:37.961 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:23:37.406 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:30:39.548 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 30, train_steps: 1200 +2026-06-19 08:31:15.812 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=30 (missing=0, unexpected=0) +2026-06-19 08:31:18.259 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 08:31:18.442 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:31:18.537 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:31:18.619 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:32:26.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000030/group_6/index_2034_seed_1422_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 08:32:29.394 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 08:32:46.359 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:32:46.369 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:32:46.369 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:33:55.904 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000030/group_6/index_2034_seed_1423_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 08:33:59.039 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 08:34:16.155 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:34:19.980 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:34:21.104 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:36:14.510 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-19 08:36:15.227 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:38:14.171 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:45:14.286 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 31, train_steps: 1240 +2026-06-19 08:45:14.993 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 08:45:15.211 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:45:15.361 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:45:15.440 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:46:20.985 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_6/index_946_seed_12652_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 08:46:24.042 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 08:46:40.959 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:46:40.969 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:46:40.969 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:47:50.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_6/index_946_seed_12653_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 08:47:53.726 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 08:48:10.859 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:48:15.288 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:48:16.427 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:50:10.323 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.84s (1.90min) +2026-06-19 08:50:11.023 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:52:11.149 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:59:13.167 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 32, train_steps: 1280 +2026-06-19 08:59:13.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 08:59:13.953 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:59:14.034 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:59:14.147 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:00:20.029 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000032/group_6/index_1203_seed_12795_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 09:00:23.153 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 09:00:40.761 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:00:40.771 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:00:40.771 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:01:48.668 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000032/group_6/index_1203_seed_12796_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 09:01:52.565 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 09:02:09.763 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:02:13.436 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:02:14.645 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:04:08.713 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.91s (1.90min) +2026-06-19 09:04:09.420 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:06:10.406 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:13:13.286 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 33, train_steps: 1320 +2026-06-19 09:13:13.953 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 09:13:14.170 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:13:14.281 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:13:14.362 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:14:20.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000033/group_6/index_1178_seed_12372_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 09:14:25.232 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 09:14:41.780 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:14:41.789 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:14:41.790 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:15:49.857 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000033/group_6/index_1178_seed_12373_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 09:15:53.557 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 09:16:10.895 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:16:14.822 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:16:15.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:18:11.066 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.06s (1.92min) +2026-06-19 09:18:11.786 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:20:11.625 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:27:13.803 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 34, train_steps: 1360 +2026-06-19 09:27:14.444 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 09:27:14.553 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:27:14.581 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:27:14.625 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:28:20.502 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_6/index_281_seed_2516_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 09:28:23.393 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 09:28:40.762 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:28:40.771 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:28:40.771 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:29:49.756 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_6/index_281_seed_2517_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 09:29:54.648 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 09:30:11.535 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:30:13.695 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:30:14.917 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:32:08.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.81s (1.90min) +2026-06-19 09:32:09.593 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:34:09.778 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:41:12.395 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 35, train_steps: 1400 +2026-06-19 09:41:13.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 09:41:13.212 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:41:13.279 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:41:13.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:42:19.154 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000035/group_6/index_6787_seed_10319_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 09:42:23.772 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 09:42:40.804 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:42:40.810 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:42:40.810 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:43:48.545 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000035/group_6/index_6787_seed_10320_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 09:43:52.256 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 09:44:09.587 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:44:13.080 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:44:14.091 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:46:10.686 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 116.56s (1.94min) +2026-06-19 09:46:11.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:48:10.894 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:55:11.963 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 36, train_steps: 1440 +2026-06-19 09:55:12.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 09:55:12.760 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:55:12.777 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:55:12.808 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:56:19.110 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_6/index_2473_seed_5919_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 09:56:24.081 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 09:56:40.861 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:56:40.871 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:56:40.874 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:57:48.155 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_6/index_2473_seed_5920_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 09:57:51.846 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 09:58:09.288 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:58:12.646 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:58:13.843 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:00:08.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-19 10:00:09.671 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:02:10.006 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:09:11.703 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 37, train_steps: 1480 +2026-06-19 10:09:12.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 10:09:12.607 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:09:12.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:09:12.830 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:10:18.558 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_6/index_9371_seed_25347_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 10:10:21.520 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 10:10:38.274 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:10:38.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:10:38.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:11:47.863 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_6/index_9371_seed_25348_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 10:11:52.880 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 10:12:09.621 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:12:13.042 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:12:13.987 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:14:09.030 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.94s (1.92min) +2026-06-19 10:14:09.737 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:16:10.367 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:23:12.361 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 38, train_steps: 1520 +2026-06-19 10:23:13.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 10:23:13.218 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:23:13.302 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:23:13.420 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:24:19.202 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000038/group_6/index_6279_seed_3098_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 10:24:23.007 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 10:24:40.778 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:24:40.789 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:24:40.789 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:25:48.706 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000038/group_6/index_6279_seed_3099_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 10:25:51.708 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 10:26:08.977 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:26:12.667 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:26:13.850 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:28:09.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.16s (1.92min) +2026-06-19 10:28:09.954 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:30:10.744 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:37:11.953 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 39, train_steps: 1560 +2026-06-19 10:37:12.646 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 10:37:12.841 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:37:12.963 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:37:13.060 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:38:18.981 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000039/group_6/index_7150_seed_4012_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 10:38:22.069 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 10:38:40.405 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:38:40.415 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:38:40.416 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:41:17.625 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000039/group_6/index_7150_seed_4013_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 10:41:22.581 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 10:41:40.433 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:41:42.526 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:41:43.652 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:43:36.471 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.70s (1.88min) +2026-06-19 10:43:37.180 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:45:37.174 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:52:37.856 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 40, train_steps: 1600 +2026-06-19 10:53:13.839 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=40 (missing=0, unexpected=0) +2026-06-19 10:53:15.773 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 10:53:15.941 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:53:16.028 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:53:16.104 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:54:23.367 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_6/index_9896_seed_241_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 10:54:27.037 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 10:54:42.739 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:54:42.749 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:54:42.749 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:55:52.353 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_6/index_9896_seed_242_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 10:55:56.914 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 10:56:12.572 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:56:16.410 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:56:17.326 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:58:11.717 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.27s (1.90min) +2026-06-19 10:58:12.416 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:00:11.325 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:07:12.574 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 41, train_steps: 1640 +2026-06-19 11:07:13.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 11:07:13.485 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:07:13.612 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:07:13.717 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:08:19.392 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_6/index_6276_seed_23876_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 11:08:23.787 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 11:08:40.806 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:08:40.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:08:40.816 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:09:48.762 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_6/index_6276_seed_23877_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 11:09:52.716 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 11:10:10.252 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:10:12.668 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:10:13.902 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:12:08.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.28s (1.90min) +2026-06-19 11:12:09.087 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:14:08.756 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:21:10.889 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 42, train_steps: 1680 +2026-06-19 11:21:11.804 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 11:21:11.973 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:21:12.049 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:21:12.201 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:22:18.143 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000042/group_6/index_511_seed_31058_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 11:22:21.265 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 11:22:39.220 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:22:39.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:22:39.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:23:47.452 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000042/group_6/index_511_seed_31059_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 11:23:52.537 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 11:24:09.750 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:24:12.073 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:24:13.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:26:07.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.32s (1.91min) +2026-06-19 11:26:08.279 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:28:06.794 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:35:06.773 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 43, train_steps: 1720 +2026-06-19 11:35:07.458 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 11:35:07.623 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:35:07.682 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:35:07.813 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:36:13.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_6/index_773_seed_6239_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 11:36:17.191 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 11:36:34.496 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:36:34.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:36:34.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:37:46.314 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_6/index_773_seed_6240_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 11:37:49.626 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 11:38:06.871 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:38:10.299 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:38:11.238 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:40:05.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.38s (1.91min) +2026-06-19 11:40:06.549 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:42:03.522 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:49:04.428 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 44, train_steps: 1760 +2026-06-19 11:49:05.114 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 11:49:05.334 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:49:05.465 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:49:05.546 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:50:11.294 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000044/group_6/index_3332_seed_8325_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 11:50:15.156 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 11:50:32.097 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:50:32.106 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:50:32.107 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:51:41.535 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000044/group_6/index_3332_seed_8326_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 11:51:44.475 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 11:52:01.558 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:52:06.491 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:52:07.498 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:54:01.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.69s (1.89min) +2026-06-19 11:54:01.910 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:56:01.233 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:03:03.876 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 45, train_steps: 1800 +2026-06-19 12:03:04.556 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 12:03:04.662 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:03:04.714 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:03:04.766 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:04:10.845 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000045/group_6/index_6294_seed_30556_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 12:04:13.801 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 12:04:31.125 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:04:31.135 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:04:31.135 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:05:40.360 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000045/group_6/index_6294_seed_30557_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 12:05:43.457 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 12:06:00.896 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:06:05.347 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:06:06.326 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:08:01.417 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.05s (1.92min) +2026-06-19 12:08:02.101 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:10:01.380 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:17:03.594 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 46, train_steps: 1840 +2026-06-19 12:17:04.279 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 12:17:04.374 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:17:04.425 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:17:04.472 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:18:10.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_6/index_5013_seed_32871_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 12:18:13.664 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 12:18:30.890 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:18:30.899 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:18:30.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:19:40.593 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_6/index_5013_seed_32872_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 12:19:45.713 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 12:20:02.822 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:20:05.333 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:20:06.492 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:22:01.733 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.17s (1.92min) +2026-06-19 12:22:02.442 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:24:01.449 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:31:02.919 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 47, train_steps: 1880 +2026-06-19 12:31:03.608 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 12:31:03.826 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:31:03.969 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:31:04.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:32:09.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000047/group_6/index_3662_seed_17668_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 12:32:14.072 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 12:32:31.043 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:32:31.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:32:31.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:33:39.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000047/group_6/index_3662_seed_17669_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 12:33:44.122 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 12:34:01.406 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:34:03.182 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:34:04.262 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:35:58.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-19 12:35:58.983 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:37:55.889 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:44:57.678 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 48, train_steps: 1920 +2026-06-19 12:44:58.355 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 12:44:58.515 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:44:58.578 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:44:58.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:46:04.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_6/index_2529_seed_16087_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 12:46:08.560 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 12:46:25.923 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:46:25.933 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:46:25.933 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:47:34.747 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_6/index_2529_seed_16088_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 12:47:37.794 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 12:47:57.546 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:47:58.931 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:47:59.954 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:49:54.620 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-19 12:49:55.343 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:51:52.288 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:58:52.090 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 49, train_steps: 1960 +2026-06-19 12:58:52.782 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 12:58:52.941 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:58:52.998 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:58:53.110 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:59:58.997 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_6/index_9415_seed_19876_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 13:00:03.977 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 13:00:20.827 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:00:20.836 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:00:20.836 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:01:30.223 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_6/index_9415_seed_19877_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 13:01:33.936 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 13:01:51.294 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:01:54.369 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:01:55.355 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:03:48.446 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.04s (1.88min) +2026-06-19 13:03:49.149 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:05:45.465 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:12:45.279 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 50, train_steps: 2000 +2026-06-19 13:14:21.535 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=50 (missing=0, unexpected=0) +2026-06-19 13:14:23.203 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 13:14:23.358 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:14:23.435 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:14:23.535 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:15:31.277 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_6/index_7465_seed_30695_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 13:15:36.247 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 13:15:53.196 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:15:53.205 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:15:53.206 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:17:00.888 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_6/index_7465_seed_30696_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 13:17:05.605 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 13:17:22.880 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:17:25.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:17:26.118 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:19:20.081 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.84s (1.90min) +2026-06-19 13:19:20.807 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:21:17.522 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:28:18.465 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 51, train_steps: 2040 +2026-06-19 13:28:19.186 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 13:28:19.401 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:28:19.537 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:28:19.632 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:29:25.269 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000051/group_6/index_2159_seed_3488_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 13:29:29.949 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 13:29:46.982 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:29:46.992 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:29:46.992 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:30:54.029 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000051/group_6/index_2159_seed_3489_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 13:30:57.852 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 13:31:15.401 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:31:18.029 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:31:19.078 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:33:13.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.83s (1.90min) +2026-06-19 13:33:13.742 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:35:13.228 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:42:13.616 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 52, train_steps: 2080 +2026-06-19 13:42:14.278 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 13:42:14.483 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:42:14.580 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:42:14.674 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:43:20.781 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_6/index_5733_seed_30410_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 13:43:25.809 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 13:43:42.509 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:43:42.518 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:43:42.518 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:44:49.995 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_6/index_5733_seed_30411_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 13:44:54.677 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 13:45:11.787 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:45:15.287 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:45:16.350 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:47:10.481 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.03s (1.90min) +2026-06-19 13:47:11.197 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:49:11.263 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:56:12.594 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 53, train_steps: 2120 +2026-06-19 13:56:13.303 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 13:56:13.414 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:56:13.457 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:56:13.487 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:57:19.398 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_6/index_4856_seed_15265_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 13:57:23.538 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 13:57:41.195 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:57:41.205 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:57:41.206 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:58:48.727 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_6/index_4856_seed_15266_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 13:58:51.864 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 13:59:09.881 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:59:13.130 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:59:14.121 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:01:08.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.41s (1.91min) +2026-06-19 14:01:09.529 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:03:10.474 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:10:12.163 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 54, train_steps: 2160 +2026-06-19 14:10:12.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 14:10:13.036 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:10:13.097 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:10:13.228 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:11:18.966 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_6/index_4610_seed_9214_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 14:11:23.956 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 14:11:41.201 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:11:41.209 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:11:41.210 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:12:47.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_6/index_4610_seed_9215_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 14:12:52.486 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 14:13:10.013 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:13:12.866 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:13:13.825 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:15:08.493 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.42s (1.91min) +2026-06-19 14:15:09.218 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:17:09.304 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:24:11.910 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 55, train_steps: 2200 +2026-06-19 14:24:12.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 14:24:12.784 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:24:12.876 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:24:12.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:25:18.618 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_6/index_9062_seed_14566_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 14:25:23.335 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 14:25:40.865 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:25:40.874 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:25:40.874 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:26:48.291 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_6/index_9062_seed_14567_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 14:26:52.114 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 14:27:10.181 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:27:12.754 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:27:13.925 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:29:09.888 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.82s (1.93min) +2026-06-19 14:29:10.605 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:31:10.762 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:38:12.116 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 56, train_steps: 2240 +2026-06-19 14:38:12.811 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 14:38:13.017 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:38:13.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:38:13.236 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:39:19.001 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_6/index_5111_seed_22368_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 14:39:24.007 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 14:39:41.098 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:39:41.107 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:39:41.110 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:40:48.823 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_6/index_5111_seed_22369_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 14:40:52.489 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 14:41:10.165 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:41:13.092 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:41:14.086 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:43:08.781 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.66s (1.91min) +2026-06-19 14:43:09.475 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:45:09.716 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:52:11.847 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 57, train_steps: 2280 +2026-06-19 14:52:12.526 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 14:52:12.622 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:52:12.673 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:52:12.720 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:53:18.705 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_6/index_846_seed_23254_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 14:53:21.656 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 14:53:39.803 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:53:39.813 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:53:39.813 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:54:48.617 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_6/index_846_seed_23255_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 14:54:53.651 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 14:55:10.965 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:55:13.057 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:55:14.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:57:08.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.79s (1.91min) +2026-06-19 14:57:09.597 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:59:09.878 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:06:12.283 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 58, train_steps: 2320 +2026-06-19 15:06:12.985 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 15:06:13.132 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:06:13.185 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:06:13.299 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:07:19.164 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_6/index_1185_seed_24315_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 15:07:23.913 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 15:07:42.087 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:07:42.096 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:07:42.096 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:08:48.902 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_6/index_1185_seed_24316_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 15:08:53.575 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 15:09:10.341 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:09:13.062 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:09:13.963 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:11:08.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.71s (1.91min) +2026-06-19 15:11:09.498 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:13:09.808 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:20:12.055 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 59, train_steps: 2360 +2026-06-19 15:20:12.738 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 15:20:12.916 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:20:13.011 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:20:13.136 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:21:18.700 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_6/index_3106_seed_10063_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 15:21:23.431 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 15:21:40.987 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:21:40.996 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:21:40.999 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:22:47.757 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_6/index_3106_seed_10064_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 15:22:52.487 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 15:23:09.751 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:23:12.372 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:23:13.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:25:09.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.28s (1.92min) +2026-06-19 15:25:09.742 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:27:10.365 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:34:12.066 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 60, train_steps: 2400 +2026-06-19 15:34:49.613 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=60 (missing=0, unexpected=0) +2026-06-19 15:34:50.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 15:34:50.772 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:34:50.863 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:34:50.943 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:35:58.569 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_6/index_3324_seed_6707_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 15:36:02.389 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 15:36:18.211 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:36:18.221 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:36:18.221 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:37:28.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_6/index_3324_seed_6708_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 15:37:32.960 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 15:37:48.905 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:37:53.529 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:37:54.633 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:39:47.707 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-19 15:39:48.403 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:41:45.232 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:48:44.589 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 61, train_steps: 2440 +2026-06-19 15:48:45.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 15:48:45.408 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:48:45.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:48:45.481 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:49:51.373 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_6/index_1961_seed_8607_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 15:49:54.850 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 15:50:10.860 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:50:10.870 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:50:10.870 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:51:23.209 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_6/index_1961_seed_8608_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 15:51:27.912 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 15:51:44.003 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:51:47.609 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:51:48.520 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:53:41.681 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.04s (1.88min) +2026-06-19 15:53:42.399 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:55:39.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:02:39.001 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 62, train_steps: 2480 +2026-06-19 16:02:39.688 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 16:02:39.866 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:02:39.933 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:02:40.060 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:03:46.453 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000062/group_6/index_5843_seed_2020_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 16:03:51.071 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 16:04:07.668 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:04:07.677 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:04:07.677 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:05:18.703 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000062/group_6/index_5843_seed_2021_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 16:05:23.355 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 16:05:40.981 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:05:44.132 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:05:45.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:07:38.761 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.59s (1.89min) +2026-06-19 16:07:39.471 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:09:38.852 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:16:39.315 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 63, train_steps: 2520 +2026-06-19 16:16:40.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 16:16:40.178 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:16:40.216 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:16:40.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:17:47.055 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_6/index_6544_seed_25754_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 16:17:51.614 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 16:18:08.389 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:18:08.399 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:18:08.399 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:19:19.896 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_6/index_6544_seed_25755_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 16:19:23.452 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 16:19:40.826 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:19:45.197 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:19:46.198 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:21:39.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.36s (1.89min) +2026-06-19 16:21:40.303 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:23:38.879 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:30:40.936 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 64, train_steps: 2560 +2026-06-19 16:30:41.626 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 16:30:41.827 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:30:41.927 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:30:41.993 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:31:48.524 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_6/index_9052_seed_10600_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 16:31:51.653 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 16:32:08.074 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:32:08.084 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:32:08.084 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:33:20.810 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_6/index_9052_seed_10601_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 16:33:23.787 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 16:33:40.842 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:33:45.471 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:33:46.456 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:35:40.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.63s (1.89min) +2026-06-19 16:35:40.844 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:37:39.397 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:44:41.298 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 65, train_steps: 2600 +2026-06-19 16:44:41.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 16:44:42.128 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:44:42.207 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:44:42.332 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:45:48.992 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_6/index_744_seed_1537_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 16:45:53.942 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 16:46:10.308 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:46:10.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:46:10.318 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:47:22.623 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_6/index_744_seed_1538_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 16:47:26.172 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 16:47:43.373 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:47:47.482 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:47:48.460 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:49:42.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.72s (1.90min) +2026-06-19 16:49:42.959 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:51:40.035 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:58:41.284 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 66, train_steps: 2640 +2026-06-19 16:58:41.989 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 16:58:42.107 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:58:42.137 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:58:42.186 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:59:48.674 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_6/index_3146_seed_14650_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 16:59:51.722 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 17:00:08.433 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:00:08.443 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:00:08.443 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:01:19.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_6/index_3146_seed_14651_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 17:01:22.596 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 17:01:39.826 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:01:43.599 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:01:44.572 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:03:37.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.34s (1.89min) +2026-06-19 17:03:38.682 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:05:38.381 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:12:40.125 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 67, train_steps: 2680 +2026-06-19 17:12:40.908 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 17:12:41.203 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:12:41.290 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:12:41.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:13:48.478 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_6/index_2275_seed_32831_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 17:13:51.570 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 17:14:08.120 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:14:08.136 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:14:08.136 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:15:20.385 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_6/index_2275_seed_32832_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 17:15:23.395 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 17:15:40.792 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:15:44.529 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:15:45.518 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:17:38.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 17:17:39.679 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:19:39.111 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:26:39.550 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 68, train_steps: 2720 +2026-06-19 17:26:40.216 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 17:26:40.427 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:26:40.551 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:26:40.642 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:27:47.504 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_6/index_8701_seed_16831_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 17:27:50.427 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 17:28:07.128 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:28:07.138 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:28:07.140 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:29:20.091 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_6/index_8701_seed_16832_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 17:29:23.011 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 17:29:40.805 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:29:43.886 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:29:44.880 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:31:38.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.76s (1.90min) +2026-06-19 17:31:39.420 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:33:38.461 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:40:38.370 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 69, train_steps: 2760 +2026-06-19 17:40:39.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 17:40:39.181 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:40:39.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:40:39.388 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:41:45.444 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000069/group_6/index_1350_seed_9347_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 17:41:48.433 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 17:42:05.121 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:42:05.131 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:42:05.131 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:43:16.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000069/group_6/index_1350_seed_9348_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 17:43:20.085 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 17:43:39.491 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:43:41.069 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:43:42.131 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:45:35.854 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.55s (1.89min) +2026-06-19 17:45:36.570 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:47:37.115 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:54:38.578 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 70, train_steps: 2800 +2026-06-19 17:55:14.468 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=70 (missing=0, unexpected=0) +2026-06-19 17:55:16.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 17:55:17.172 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:55:17.261 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:55:17.359 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:56:25.248 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_6/index_5899_seed_17584_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 17:56:30.116 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 17:56:46.201 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:56:46.210 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:56:46.210 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:57:54.298 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_6/index_5899_seed_17585_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 17:57:58.974 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 17:58:15.426 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:58:18.228 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:58:19.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:00:13.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.58s (1.89min) +2026-06-19 18:00:13.838 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:02:13.851 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:09:13.888 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 71, train_steps: 2840 +2026-06-19 18:09:14.615 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 18:09:14.751 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:09:14.787 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:09:14.889 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:10:20.471 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_6/index_8653_seed_10897_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 18:10:24.209 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 18:10:41.292 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:10:41.302 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:10:41.302 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:11:51.733 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_6/index_8653_seed_10898_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 18:11:54.805 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 18:12:11.911 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:12:16.021 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:12:17.040 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:14:12.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.67s (1.93min) +2026-06-19 18:14:13.460 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:16:12.693 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:23:13.829 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 72, train_steps: 2880 +2026-06-19 18:23:14.506 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 18:23:14.706 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:23:14.822 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:23:14.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:24:20.889 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_6/index_3063_seed_31368_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 18:24:25.879 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 18:24:43.515 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:24:43.524 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:24:43.524 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:25:51.126 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_6/index_3063_seed_31369_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 18:25:55.112 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 18:26:13.594 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:26:14.711 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:26:15.868 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:28:10.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.45s (1.91min) +2026-06-19 18:28:11.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:30:10.986 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:37:12.789 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 73, train_steps: 2920 +2026-06-19 18:37:13.455 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 18:37:13.579 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:37:13.617 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:37:13.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:38:19.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_6/index_3958_seed_1160_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 18:38:22.655 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 18:38:40.810 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:38:40.820 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:38:40.820 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:39:49.912 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_6/index_3958_seed_1161_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 18:39:54.682 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 18:40:12.545 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:40:13.249 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:40:14.181 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:42:09.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.16s (1.92min) +2026-06-19 18:42:10.284 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:44:10.445 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:51:12.763 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 74, train_steps: 2960 +2026-06-19 18:51:13.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 18:51:13.634 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:51:13.711 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:51:13.823 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:52:20.042 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_6/index_3510_seed_30296_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 18:52:23.945 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 18:52:40.879 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:52:40.886 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:52:40.886 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:53:49.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_6/index_3510_seed_30297_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 18:53:52.649 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 18:54:10.013 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:54:13.982 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:54:14.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:56:10.312 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.21s (1.92min) +2026-06-19 18:56:11.031 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:58:11.441 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:05:14.028 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 75, train_steps: 3000 +2026-06-19 19:06:23.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 19:06:23.353 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:06:23.395 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:06:23.492 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:07:29.744 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000075/group_6/index_2022_seed_10217_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 19:07:34.603 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 19:07:51.251 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:07:51.260 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:07:51.260 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:08:59.010 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000075/group_6/index_2022_seed_10218_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 19:09:02.738 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 19:09:19.795 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:09:24.209 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:09:25.213 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:11:18.397 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.14s (1.89min) +2026-06-19 19:11:19.100 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:13:16.252 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:20:16.320 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 76, train_steps: 3040 +2026-06-19 19:20:16.990 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 19:20:17.202 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:20:17.298 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:20:17.395 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:21:23.092 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000076/group_6/index_7052_seed_31561_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 19:21:27.968 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 19:21:44.816 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:21:44.825 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:21:44.826 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:22:53.642 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000076/group_6/index_7052_seed_31562_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 19:22:58.235 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 19:23:15.057 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:23:18.455 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:23:19.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:25:13.441 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.66s (1.89min) +2026-06-19 19:25:14.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:27:12.732 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:34:13.268 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 77, train_steps: 3080 +2026-06-19 19:34:13.949 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 19:34:14.071 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:34:14.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:34:14.181 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:35:19.817 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_6/index_8250_seed_24058_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 19:35:23.404 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 19:35:40.785 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:35:40.794 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:35:40.795 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:36:49.302 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_6/index_8250_seed_24059_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 19:36:52.439 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 19:37:09.594 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:37:13.074 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:37:14.078 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:39:08.734 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.60s (1.91min) +2026-06-19 19:39:09.441 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:41:09.768 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:48:12.650 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 78, train_steps: 3120 +2026-06-19 19:48:13.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 19:48:13.536 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:48:13.615 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:48:13.738 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:49:19.651 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_6/index_443_seed_23495_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 19:49:22.671 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 19:49:40.106 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:49:40.116 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:49:40.116 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:50:49.860 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_6/index_443_seed_23496_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 19:50:52.940 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 19:51:10.145 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:51:13.323 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:51:14.318 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:53:09.267 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-19 19:53:09.975 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:55:09.984 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:02:13.176 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 79, train_steps: 3160 +2026-06-19 20:02:13.885 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 20:02:14.100 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:02:14.227 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:02:14.334 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:03:20.168 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000079/group_6/index_1357_seed_13449_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 20:03:23.220 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 20:03:40.414 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:03:40.424 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:03:40.424 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:04:49.756 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000079/group_6/index_1357_seed_13450_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 20:04:54.689 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 20:05:11.154 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:05:13.093 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:05:14.034 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:07:09.660 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.53s (1.93min) +2026-06-19 20:07:10.385 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:09:11.031 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:16:13.955 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 80, train_steps: 3200 +2026-06-19 20:16:50.694 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=80 (missing=0, unexpected=0) +2026-06-19 20:16:53.092 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 20:16:53.284 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:16:53.392 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:16:53.496 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:18:02.068 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_6/index_3973_seed_27956_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 20:18:06.780 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 20:18:23.208 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:18:23.217 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:18:23.220 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:19:30.970 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_6/index_3973_seed_27957_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 20:19:35.651 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 20:19:52.536 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:19:54.752 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:19:55.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:21:49.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-19 20:21:50.363 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:23:47.271 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:30:47.423 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 81, train_steps: 3240 +2026-06-19 20:30:48.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 20:30:48.283 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:30:48.306 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:30:48.380 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:31:54.435 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_6/index_7556_seed_27043_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 20:31:59.108 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 20:32:15.398 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:32:15.407 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:32:15.410 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:33:25.062 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_6/index_7556_seed_27044_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 20:33:28.745 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 20:33:45.907 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:33:49.273 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:33:50.300 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:35:44.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.70s (1.89min) +2026-06-19 20:35:44.753 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:37:41.676 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:44:40.150 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 82, train_steps: 3280 +2026-06-19 20:44:40.696 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 20:44:40.902 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:44:41.042 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:44:41.141 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:45:47.941 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_6/index_8579_seed_8903_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 20:45:50.919 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 20:46:07.564 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:46:07.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:46:07.575 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:47:19.947 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_6/index_8579_seed_8904_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 20:47:23.112 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 20:47:40.759 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:47:44.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:47:45.114 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:49:38.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.42s (1.89min) +2026-06-19 20:49:39.275 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:51:38.431 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:58:38.999 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 83, train_steps: 3320 +2026-06-19 20:58:39.739 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 20:58:39.974 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:58:40.012 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:58:40.107 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:59:46.485 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_6/index_5643_seed_19982_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 20:59:49.555 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 21:00:06.241 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:00:06.251 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:00:06.251 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:01:19.576 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_6/index_5643_seed_19983_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 21:01:22.992 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 21:01:40.803 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:01:44.304 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:01:45.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:03:38.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.28s (1.89min) +2026-06-19 21:03:39.339 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:05:38.817 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:12:39.007 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 84, train_steps: 3360 +2026-06-19 21:12:39.655 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 21:12:39.864 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:12:39.999 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:12:40.082 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:13:46.998 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_6/index_5653_seed_15348_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 21:13:50.133 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 21:14:06.981 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:14:06.990 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:14:06.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:15:19.543 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_6/index_5653_seed_15349_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 21:15:24.333 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 21:15:41.303 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:15:43.791 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:15:44.757 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:17:38.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.31s (1.89min) +2026-06-19 21:17:38.863 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:19:38.521 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:26:39.498 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 85, train_steps: 3400 +2026-06-19 21:26:40.168 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 21:26:40.357 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:26:40.459 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:26:40.571 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:27:47.463 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_6/index_4210_seed_11238_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 21:27:52.060 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 21:28:08.403 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:28:08.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:28:08.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:29:19.994 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_6/index_4210_seed_11239_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 21:29:24.568 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 21:29:41.954 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:29:43.697 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:29:44.738 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:31:38.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-19 21:31:38.988 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:33:38.754 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:40:38.641 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 86, train_steps: 3440 +2026-06-19 21:40:39.367 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 21:40:39.572 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:40:39.700 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:40:39.804 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:41:46.937 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_6/index_504_seed_20305_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 21:41:52.009 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 21:42:08.362 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:42:08.369 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:42:08.369 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:43:18.654 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_6/index_504_seed_20306_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 21:43:22.340 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 21:43:40.125 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:43:42.867 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:43:43.863 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:45:37.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 21:45:37.987 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:47:37.706 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:54:39.053 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 87, train_steps: 3480 +2026-06-19 21:54:39.742 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 21:54:39.843 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:54:39.906 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:54:39.974 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:55:46.569 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000087/group_6/index_6729_seed_2085_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 21:55:51.361 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 21:56:07.644 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:56:07.651 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:56:07.651 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:57:19.080 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000087/group_6/index_6729_seed_2086_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 21:57:23.802 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 21:57:40.828 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:57:43.169 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:57:44.133 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:59:37.604 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.39s (1.89min) +2026-06-19 21:59:38.318 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:01:38.553 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:08:39.988 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 88, train_steps: 3520 +2026-06-19 22:08:40.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 22:08:40.817 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:08:40.857 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:08:40.937 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:09:47.738 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_6/index_8179_seed_27128_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 22:09:52.527 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 22:10:09.038 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:10:09.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:10:09.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:11:19.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_6/index_8179_seed_27129_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 22:11:23.230 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 22:11:40.537 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:11:43.755 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:11:44.770 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:13:38.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.57s (1.89min) +2026-06-19 22:13:39.103 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:15:38.303 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:22:38.710 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 89, train_steps: 3560 +2026-06-19 22:22:39.425 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 22:22:39.566 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:22:39.610 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:22:39.713 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:23:46.609 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_6/index_7044_seed_25565_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 22:23:49.687 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 22:24:06.532 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:24:06.542 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:24:06.542 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:25:19.808 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_6/index_7044_seed_25566_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 22:25:22.925 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 22:25:40.433 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:25:43.683 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:25:44.742 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:27:38.162 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 22:27:38.886 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:29:38.493 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:36:39.589 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 90, train_steps: 3600 +2026-06-19 22:37:15.707 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=90 (missing=0, unexpected=0) +2026-06-19 22:37:17.308 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 22:37:17.549 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:37:17.665 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:37:17.742 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:38:24.829 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_6/index_1563_seed_20683_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 22:38:29.100 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 22:38:45.827 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:38:45.836 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:38:45.836 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:39:55.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_6/index_1563_seed_20684_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 22:39:58.873 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 22:40:15.963 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:40:19.241 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:40:20.277 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:42:14.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.70s (1.89min) +2026-06-19 22:42:14.863 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:44:13.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:51:14.849 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 91, train_steps: 3640 +2026-06-19 22:51:15.699 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 22:51:15.825 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:51:15.855 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:51:15.903 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:52:21.796 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000091/group_6/index_8210_seed_6562_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 22:52:24.865 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 22:52:43.104 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:52:43.114 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:52:43.114 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:53:51.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000091/group_6/index_8210_seed_6563_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 22:53:55.172 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 22:54:13.801 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:54:15.937 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:54:16.959 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:56:12.672 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.67s (1.93min) +2026-06-19 22:56:13.388 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:58:13.111 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:05:14.269 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 92, train_steps: 3680 +2026-06-19 23:05:15.118 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 23:05:15.300 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:05:15.398 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:05:15.515 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:06:21.108 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000092/group_6/index_7031_seed_3569_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 23:06:24.190 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 23:06:41.605 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:06:41.614 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:06:41.615 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:07:51.060 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000092/group_6/index_7031_seed_3570_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 23:07:56.070 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 23:08:13.106 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:08:16.058 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:08:17.045 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:10:11.751 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.62s (1.91min) +2026-06-19 23:10:12.471 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:12:11.282 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:19:13.479 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 93, train_steps: 3720 +2026-06-19 23:19:14.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 23:19:14.277 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:19:14.315 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:19:14.377 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:20:20.299 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_6/index_3519_seed_6356_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 23:20:24.072 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 23:20:42.930 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:20:42.938 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:20:42.939 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:21:50.204 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_6/index_3519_seed_6357_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 23:21:54.015 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 23:22:12.760 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:22:14.354 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:22:15.540 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:24:10.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.33s (1.92min) +2026-06-19 23:24:11.708 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:26:11.283 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:33:13.943 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 94, train_steps: 3760 +2026-06-19 23:33:14.640 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 23:33:14.796 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:33:14.850 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:33:14.977 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:34:20.959 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_6/index_8066_seed_8255_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 23:34:24.042 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 23:34:41.488 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:34:41.498 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:34:41.498 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:35:51.482 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_6/index_8066_seed_8256_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 23:35:54.503 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 23:36:11.872 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:36:16.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:36:17.132 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:38:12.062 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.89s (1.91min) +2026-06-19 23:38:12.784 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:40:11.319 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:47:11.877 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 95, train_steps: 3800 +2026-06-19 23:47:12.590 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-19 23:47:12.781 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:47:12.893 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:47:12.971 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:48:18.634 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_6/index_3152_seed_15750_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-19 23:48:21.605 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 23:48:40.106 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:48:40.116 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:48:40.118 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:49:48.281 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_6/index_3152_seed_15751_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-19 23:49:53.298 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 23:50:10.299 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:50:12.004 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:50:13.206 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:52:08.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.82s (1.91min) +2026-06-19 23:52:08.791 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:54:07.675 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:01:09.779 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 96, train_steps: 3840 +2026-06-20 00:01:10.475 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 00:01:10.670 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:01:10.889 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:01:11.004 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:02:17.006 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_6/index_9675_seed_11507_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 00:02:21.032 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 00:02:39.842 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:02:39.852 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:02:39.852 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:03:47.367 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_6/index_9675_seed_11508_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 00:03:50.400 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 00:04:07.871 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:04:11.783 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:04:12.923 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:06:07.928 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 00:06:08.644 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:08:07.687 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:15:08.506 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 97, train_steps: 3880 +2026-06-20 00:15:09.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 00:15:09.412 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:15:09.524 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:15:09.638 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:16:15.315 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_6/index_1254_seed_32195_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 00:16:20.220 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 00:16:37.366 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:16:37.376 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:16:37.376 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:17:47.297 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_6/index_1254_seed_32196_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 00:17:51.980 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 00:18:08.817 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:18:11.238 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:18:12.224 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:20:07.671 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.20s (1.92min) +2026-06-20 00:20:08.401 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:22:07.125 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:29:07.007 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 98, train_steps: 3920 +2026-06-20 00:29:07.691 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 00:29:07.811 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:29:07.852 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:29:07.925 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:30:13.982 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000098/group_6/index_2843_seed_5001_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 00:30:17.645 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 00:30:34.958 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:30:34.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:30:34.971 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:31:46.435 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000098/group_6/index_2843_seed_5002_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 00:31:49.652 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 00:32:07.183 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:32:10.419 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:32:11.429 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:34:06.389 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 00:34:07.118 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:36:03.850 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:43:05.348 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 99, train_steps: 3960 +2026-06-20 00:43:06.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 00:43:06.207 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:43:06.311 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:43:06.412 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:44:12.158 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_6/index_6955_seed_19658_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 00:44:15.177 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 00:44:32.710 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:44:32.719 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:44:32.720 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:45:46.201 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_6/index_6955_seed_19659_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 00:45:49.359 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 00:46:06.436 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:46:09.899 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:46:10.918 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:48:05.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.75s (1.91min) +2026-06-20 00:48:06.440 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:50:03.397 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:57:04.807 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 100, train_steps: 4000 +2026-06-20 00:58:47.617 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=100 (missing=0, unexpected=0) +2026-06-20 00:58:49.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 00:58:49.778 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:58:49.894 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:58:49.972 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:59:57.331 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_6/index_6339_seed_7677_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 01:00:02.218 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 01:00:19.208 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:00:19.218 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:00:19.218 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:01:26.812 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_6/index_6339_seed_7678_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 01:01:31.569 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 01:01:48.613 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:01:50.303 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:01:51.251 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:03:44.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-20 01:03:45.434 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:05:42.218 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:12:39.859 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 101, train_steps: 4040 +2026-06-20 01:12:40.570 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 01:12:40.780 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:12:40.887 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:12:40.945 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:13:48.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_6/index_7173_seed_22520_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 01:13:53.126 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 01:14:10.547 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:14:10.556 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:14:10.557 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:15:20.050 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_6/index_7173_seed_22521_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 01:15:23.765 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 01:15:41.132 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:15:43.499 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:15:44.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:17:38.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.62s (1.89min) +2026-06-20 01:17:38.931 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:19:38.323 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:26:37.810 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 102, train_steps: 4080 +2026-06-20 01:26:38.503 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 01:26:38.622 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:26:38.657 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:26:38.726 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:27:44.650 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_6/index_2376_seed_23527_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 01:27:47.659 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 01:28:05.652 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:28:05.662 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:28:05.664 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:29:17.180 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_6/index_2376_seed_23528_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 01:29:22.263 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 01:29:40.850 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:29:41.183 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:29:42.431 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:31:36.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-20 01:31:37.097 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:33:37.223 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:40:38.537 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 103, train_steps: 4120 +2026-06-20 01:40:39.244 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 01:40:39.378 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:40:39.409 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:40:39.506 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:41:45.994 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000103/group_6/index_9395_seed_9105_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 01:41:50.552 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 01:42:07.402 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:42:07.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:42:07.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:43:18.927 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000103/group_6/index_9395_seed_9106_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 01:43:22.741 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 01:43:40.786 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:43:43.334 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:43:44.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:45:37.767 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-20 01:45:38.487 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:47:38.264 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:54:37.797 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 104, train_steps: 4160 +2026-06-20 01:54:38.703 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 01:54:38.866 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:54:38.938 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:54:39.065 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:55:45.524 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_6/index_9439_seed_31599_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 01:55:50.433 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 01:56:07.000 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:56:07.010 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:56:07.010 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:57:18.848 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_6/index_9439_seed_31600_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 01:57:23.469 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 01:57:40.864 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:57:42.889 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:57:43.814 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:59:37.314 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-20 01:59:38.031 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:01:38.017 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:08:38.446 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 105, train_steps: 4200 +2026-06-20 02:08:39.148 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 02:08:39.355 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:08:39.468 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:08:39.561 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:09:46.131 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000105/group_6/index_7611_seed_32412_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 02:09:50.782 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 02:10:07.424 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:10:07.435 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:10:07.435 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:11:18.906 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000105/group_6/index_7611_seed_32413_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 02:11:22.652 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 02:11:40.720 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:11:43.945 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:11:44.871 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:13:38.217 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 02:13:38.925 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:15:38.266 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:22:37.852 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 106, train_steps: 4240 +2026-06-20 02:22:38.565 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 02:22:38.769 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:22:38.881 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:22:38.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:23:44.664 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000106/group_6/index_6148_seed_29968_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 02:23:48.330 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 02:24:05.032 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:24:05.042 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:24:05.043 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:25:16.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000106/group_6/index_6148_seed_29969_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 02:25:21.710 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 02:25:38.322 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:25:40.847 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:25:41.831 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:27:35.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.90s (1.90min) +2026-06-20 02:27:36.583 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:29:36.810 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:36:37.095 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 107, train_steps: 4280 +2026-06-20 02:36:37.802 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 02:36:38.008 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:36:38.136 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:36:38.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:37:43.945 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_6/index_166_seed_19416_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 02:37:47.630 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 02:38:04.660 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:38:04.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:38:04.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:39:15.112 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_6/index_166_seed_19417_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 02:39:18.183 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 02:39:35.380 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:39:39.229 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:39:40.504 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:41:35.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 02:41:36.072 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:43:35.144 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:50:34.996 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 108, train_steps: 4320 +2026-06-20 02:50:35.727 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 02:50:35.903 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:50:35.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:50:36.075 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:51:42.160 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_6/index_8646_seed_8041_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 02:51:45.185 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 02:52:02.011 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:52:02.021 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:52:02.021 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:53:13.220 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_6/index_8646_seed_8042_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 02:53:16.346 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 02:53:33.047 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:53:37.574 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:53:38.596 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:55:33.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.45s (1.91min) +2026-06-20 02:55:33.972 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:57:30.647 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:04:30.706 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 109, train_steps: 4360 +2026-06-20 03:04:31.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 03:04:31.615 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:04:31.728 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:04:31.822 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:05:37.478 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000109/group_6/index_4799_seed_7923_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 03:05:41.006 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 03:05:58.049 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:05:58.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:05:58.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:07:09.532 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000109/group_6/index_4799_seed_7924_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 03:07:12.644 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 03:07:29.827 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:07:33.407 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:07:34.402 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:09:29.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-20 03:09:30.015 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:11:26.575 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:18:25.335 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 110, train_steps: 4400 +2026-06-20 03:19:01.602 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=110 (missing=0, unexpected=0) +2026-06-20 03:19:03.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 03:19:03.463 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:19:03.553 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:19:03.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:20:11.492 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_6/index_5595_seed_3660_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 03:20:14.548 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 03:20:32.310 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:20:32.320 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:20:32.320 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:21:43.420 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_6/index_5595_seed_3661_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 03:21:46.566 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 03:22:03.730 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:22:07.524 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:22:08.511 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:24:03.966 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.41s (1.92min) +2026-06-20 03:24:04.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:26:02.472 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:33:05.395 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 111, train_steps: 4440 +2026-06-20 03:33:06.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 03:33:06.230 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:33:06.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:33:06.393 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:34:12.262 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_6/index_214_seed_11700_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 03:34:15.451 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 03:34:32.406 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:34:32.416 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:34:32.416 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:35:46.483 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_6/index_214_seed_11701_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 03:35:49.667 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 03:36:06.759 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:36:10.098 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:36:11.106 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:38:06.021 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 03:38:06.732 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:40:04.482 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:47:04.318 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 112, train_steps: 4480 +2026-06-20 03:47:05.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 03:47:05.221 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:47:05.316 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:47:05.427 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:48:11.088 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_6/index_5372_seed_21863_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 03:48:16.018 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 03:48:32.445 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:48:32.457 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:48:32.460 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:49:40.732 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_6/index_5372_seed_21864_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 03:49:44.368 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 03:50:01.508 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:50:04.983 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:50:06.258 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:52:00.943 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-20 03:52:01.658 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:54:00.962 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:01:04.002 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 113, train_steps: 4520 +2026-06-20 04:01:04.874 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 04:01:05.027 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:01:05.087 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:01:05.190 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:02:11.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000113/group_6/index_2557_seed_33177_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 04:02:16.063 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 04:02:32.509 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:02:32.517 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:02:32.518 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:03:41.947 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000113/group_6/index_2557_seed_33178_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 04:03:46.496 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 04:04:03.134 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:04:06.611 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:04:07.546 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:06:02.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.15s (1.92min) +2026-06-20 04:06:03.532 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:08:01.585 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:15:03.779 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 114, train_steps: 4560 +2026-06-20 04:15:04.473 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 04:15:04.674 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:15:04.790 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:15:04.906 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:16:10.324 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000114/group_6/index_6913_seed_16864_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 04:16:14.064 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 04:16:31.190 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:16:31.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:16:31.200 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:17:40.375 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000114/group_6/index_6913_seed_16865_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 04:17:45.340 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 04:18:02.374 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:18:04.152 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:18:05.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:20:00.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.72s (1.91min) +2026-06-20 04:20:00.895 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:22:00.099 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:29:01.472 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 115, train_steps: 4600 +2026-06-20 04:29:02.193 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 04:29:02.418 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:29:02.536 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:29:02.615 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:30:08.404 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_6/index_988_seed_15344_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 04:30:13.009 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 04:30:29.613 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:30:29.622 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:30:29.623 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:31:38.919 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_6/index_988_seed_15345_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 04:31:43.557 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 04:32:00.269 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:32:02.494 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:32:03.425 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:33:58.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.33s (1.91min) +2026-06-20 04:33:58.775 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:35:55.624 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:42:56.470 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 116, train_steps: 4640 +2026-06-20 04:42:57.160 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 04:42:57.357 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:42:57.466 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:42:57.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:44:03.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_6/index_6496_seed_7879_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 04:44:08.031 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 04:44:25.076 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:44:25.085 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:44:25.086 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:45:33.813 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_6/index_6496_seed_7880_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 04:45:38.715 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 04:45:58.009 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:45:58.380 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:45:59.409 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:47:53.698 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.25s (1.90min) +2026-06-20 04:47:54.407 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:49:50.887 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:56:50.228 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 117, train_steps: 4680 +2026-06-20 04:56:50.955 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 04:56:51.149 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:56:51.236 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:56:51.311 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:57:57.358 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_6/index_7743_seed_5837_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 04:58:00.371 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 04:58:17.535 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:58:17.544 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:58:17.545 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:59:27.959 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_6/index_7743_seed_5838_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 04:59:32.995 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 04:59:50.194 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:59:51.798 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:59:52.766 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:01:46.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.49s (1.89min) +2026-06-20 05:01:47.078 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:03:44.145 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:10:42.586 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 118, train_steps: 4720 +2026-06-20 05:10:43.295 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 05:10:43.412 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:10:43.447 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:10:43.513 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:11:49.498 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000118/group_6/index_3792_seed_4172_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 05:11:54.141 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 05:12:10.961 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:12:10.970 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:12:10.970 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:13:22.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000118/group_6/index_3792_seed_4173_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 05:13:27.459 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 05:13:44.494 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:13:46.957 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:13:47.888 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:15:41.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.26s (1.89min) +2026-06-20 05:15:42.001 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:17:38.993 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:24:37.902 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 119, train_steps: 4760 +2026-06-20 05:24:38.630 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 05:24:38.833 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:24:38.942 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:24:39.038 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:25:44.666 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_6/index_6731_seed_10063_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 05:25:48.637 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 05:26:05.897 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:26:05.907 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:26:05.907 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:27:14.959 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_6/index_6731_seed_10064_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 05:27:19.928 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 05:27:36.928 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:27:40.119 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:27:41.260 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:29:34.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-20 05:29:35.560 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:31:33.103 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:38:32.156 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 120, train_steps: 4800 +2026-06-20 05:39:09.226 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=120 (missing=0, unexpected=0) +2026-06-20 05:39:10.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 05:39:11.073 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:39:11.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:39:11.273 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:40:18.964 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000120/group_6/index_9560_seed_14658_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 05:40:23.701 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 05:40:40.921 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:40:40.928 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:40:40.928 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:41:48.592 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000120/group_6/index_9560_seed_14659_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 05:41:52.316 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 05:42:09.690 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:42:12.566 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:42:13.580 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:44:08.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.96s (1.92min) +2026-06-20 05:44:09.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:46:10.085 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:53:12.067 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 121, train_steps: 4840 +2026-06-20 05:53:12.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 05:53:12.944 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:53:13.020 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:53:13.120 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:54:19.301 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_6/index_5563_seed_24024_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 05:54:22.412 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 05:54:40.102 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:54:40.112 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:54:40.112 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:55:48.785 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_6/index_5563_seed_24025_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 05:55:53.775 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 05:56:10.823 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:56:13.156 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:56:14.081 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:58:09.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.97s (1.92min) +2026-06-20 05:58:10.076 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:00:10.573 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:07:11.950 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 122, train_steps: 4880 +2026-06-20 06:07:12.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 06:07:12.832 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:07:12.925 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:07:13.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:08:18.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_6/index_2814_seed_13594_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 06:08:23.009 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 06:08:40.853 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:08:40.863 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:08:40.864 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:09:48.700 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_6/index_2814_seed_13595_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 06:09:53.442 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 06:10:10.268 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:10:13.195 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:10:14.106 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:12:08.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.25s (1.90min) +2026-06-20 06:12:09.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:14:10.578 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:21:11.866 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 123, train_steps: 4920 +2026-06-20 06:21:12.561 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 06:21:12.767 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:21:12.857 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:21:12.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:22:18.584 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_6/index_488_seed_12202_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 06:22:23.218 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 06:22:40.781 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:22:40.791 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:22:40.791 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:23:48.166 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_6/index_488_seed_12203_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 06:23:52.052 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 06:24:09.270 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:24:12.071 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:24:13.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:26:08.788 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.47s (1.92min) +2026-06-20 06:26:09.511 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:28:09.832 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:35:11.857 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 124, train_steps: 4960 +2026-06-20 06:35:12.539 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 06:35:12.703 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:35:12.773 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:35:12.875 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:36:18.491 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_6/index_4925_seed_25524_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 06:36:21.603 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 06:36:39.516 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:36:39.526 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:36:39.526 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:37:48.135 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_6/index_4925_seed_25525_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 06:37:53.197 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 06:38:10.261 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:38:12.129 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:38:13.178 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:40:08.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.44s (1.92min) +2026-06-20 06:40:09.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:42:09.771 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:49:11.645 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 125, train_steps: 5000 +2026-06-20 06:50:15.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 06:50:15.377 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:50:15.447 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:50:15.554 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:51:21.077 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_6/index_4363_seed_21271_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 06:51:24.749 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 06:51:41.975 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:51:41.984 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:51:41.985 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:52:51.196 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_6/index_4363_seed_21272_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 06:52:56.340 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 06:53:13.300 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:53:15.926 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:53:16.863 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:55:11.654 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.68s (1.91min) +2026-06-20 06:55:12.372 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:57:11.761 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:04:11.921 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 126, train_steps: 5040 +2026-06-20 07:04:12.630 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 07:04:12.771 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:04:12.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:04:12.912 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:05:18.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_6/index_1651_seed_11286_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 07:05:23.370 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 07:05:40.403 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:05:40.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:05:40.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:06:49.262 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_6/index_1651_seed_11287_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 07:06:53.038 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 07:07:10.407 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:07:13.988 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:07:15.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:09:10.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.08s (1.92min) +2026-06-20 07:09:11.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:11:11.165 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:18:11.936 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 127, train_steps: 5080 +2026-06-20 07:18:12.633 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 07:18:12.848 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:18:12.978 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:18:13.040 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:19:18.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_6/index_1528_seed_25570_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 07:19:21.867 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 07:19:39.809 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:19:39.819 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:19:39.819 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:20:49.022 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_6/index_1528_seed_25571_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 07:20:54.143 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 07:21:11.500 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:21:13.050 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:21:14.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:23:09.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.50s (1.93min) +2026-06-20 07:23:10.327 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:25:10.157 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:32:11.177 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 128, train_steps: 5120 +2026-06-20 07:32:11.887 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 07:32:12.100 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:32:12.212 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:32:12.307 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:33:17.950 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_6/index_4822_seed_14382_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 07:33:21.648 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 07:33:39.558 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:33:39.568 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:33:39.570 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:34:48.960 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_6/index_4822_seed_14383_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 07:34:54.124 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 07:35:11.078 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:35:13.016 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:35:13.985 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:37:08.641 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.56s (1.91min) +2026-06-20 07:37:09.373 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:39:09.929 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:46:11.982 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 129, train_steps: 5160 +2026-06-20 07:46:12.714 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 07:46:12.929 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:46:13.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:46:13.138 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:47:19.050 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_6/index_326_seed_21189_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 07:47:23.643 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 07:47:40.821 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:47:40.831 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:47:40.831 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:48:48.939 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_6/index_326_seed_21190_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 07:48:52.657 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 07:49:10.071 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:49:13.070 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:49:14.093 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:51:09.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-20 07:51:09.758 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:53:10.302 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:00:11.790 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 130, train_steps: 5200 +2026-06-20 08:00:49.393 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=130 (missing=0, unexpected=0) +2026-06-20 08:00:51.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 08:00:51.430 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:00:51.535 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:00:51.630 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:02:00.176 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_6/index_5043_seed_13472_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 08:02:03.341 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 08:02:20.576 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:02:20.585 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:02:20.586 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:03:28.868 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_6/index_5043_seed_13473_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 08:03:32.270 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 08:03:49.262 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:03:52.509 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:03:53.468 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:05:47.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.72s (1.90min) +2026-06-20 08:05:48.018 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:07:44.843 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:14:45.318 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 131, train_steps: 5240 +2026-06-20 08:14:46.027 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 08:14:46.228 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:14:46.341 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:14:46.419 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:15:52.328 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_6/index_4870_seed_17363_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 08:15:56.132 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 08:16:12.974 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:16:12.984 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:16:12.984 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:17:23.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_6/index_4870_seed_17364_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 08:17:26.339 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 08:17:43.347 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:17:46.990 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:17:48.007 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:19:41.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.38s (1.89min) +2026-06-20 08:19:42.139 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:21:39.465 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:28:37.831 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 132, train_steps: 5280 +2026-06-20 08:28:38.521 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 08:28:38.702 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:28:38.794 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:28:38.888 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:29:44.981 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000132/group_6/index_227_seed_2423_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 08:29:50.034 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 08:30:06.985 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:30:06.995 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:30:06.996 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:31:15.560 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000132/group_6/index_227_seed_2424_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 08:31:20.335 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 08:31:37.451 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:31:40.849 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:31:41.781 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:33:35.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.44s (1.89min) +2026-06-20 08:33:36.119 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:35:35.260 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:42:34.338 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 133, train_steps: 5320 +2026-06-20 08:42:35.018 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 08:42:35.204 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:42:35.299 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:42:35.398 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:43:41.415 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_6/index_816_seed_638_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 08:43:45.277 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 08:44:02.375 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:44:02.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:44:02.385 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:45:11.686 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_6/index_816_seed_639_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 08:45:16.707 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 08:45:33.782 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:45:35.888 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:45:36.852 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:47:32.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.77s (1.93min) +2026-06-20 08:47:33.450 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:49:30.689 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:56:29.427 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 134, train_steps: 5360 +2026-06-20 08:56:30.116 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 08:56:30.308 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:56:30.409 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:56:30.489 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:57:36.524 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000134/group_6/index_4824_seed_9838_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 08:57:40.846 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 08:57:57.668 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:57:57.678 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:57:57.678 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:59:07.587 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000134/group_6/index_4824_seed_9839_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 08:59:10.804 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 08:59:27.647 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:59:32.041 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:59:33.045 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:01:26.319 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 09:01:27.048 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:03:23.803 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:10:22.984 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 135, train_steps: 5400 +2026-06-20 09:10:23.678 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 09:10:23.895 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:10:24.015 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:10:24.093 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:11:30.186 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000135/group_6/index_2110_seed_17452_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 09:11:35.164 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 09:11:52.381 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:11:52.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:11:52.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:13:00.477 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000135/group_6/index_2110_seed_17453_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 09:13:05.595 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 09:13:22.492 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:13:23.932 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:13:24.891 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:15:18.588 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.61s (1.89min) +2026-06-20 09:15:19.310 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:17:17.098 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:24:16.799 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 136, train_steps: 5440 +2026-06-20 09:24:17.533 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 09:24:17.696 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:24:17.754 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:24:17.856 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:25:23.457 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_6/index_8206_seed_19902_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 09:25:27.307 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 09:25:44.377 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:25:44.386 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:25:44.387 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:26:54.142 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_6/index_8206_seed_19903_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 09:26:59.153 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 09:27:16.229 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:27:18.656 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:27:19.630 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:29:12.959 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.15s (1.89min) +2026-06-20 09:29:13.686 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:31:13.176 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:38:13.216 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 137, train_steps: 5480 +2026-06-20 09:38:13.936 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 09:38:14.116 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:38:14.189 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:38:14.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:39:19.889 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000137/group_6/index_5631_seed_27289_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 09:39:24.608 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 09:39:41.554 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:39:41.563 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:39:41.564 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:40:50.309 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000137/group_6/index_5631_seed_27290_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 09:40:55.101 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 09:41:12.142 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:41:14.371 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:41:15.468 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:43:09.522 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-20 09:43:10.251 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:45:10.502 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:52:12.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 138, train_steps: 5520 +2026-06-20 09:52:12.868 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 09:52:13.054 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:52:13.139 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:52:13.416 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:53:19.176 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_6/index_8167_seed_9189_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 09:53:23.836 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 09:53:40.811 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:53:40.822 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:53:40.823 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:54:49.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_6/index_8167_seed_9190_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 09:54:54.188 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 09:55:11.348 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:55:13.383 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:55:14.557 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:57:10.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.46s (1.92min) +2026-06-20 09:57:10.921 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:59:11.150 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:06:12.670 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 139, train_steps: 5560 +2026-06-20 10:06:13.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 10:06:13.534 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:06:13.607 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:06:13.704 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:07:19.449 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_6/index_3007_seed_23683_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 10:07:24.067 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 10:07:40.811 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:07:40.822 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:07:40.822 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:08:49.993 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_6/index_3007_seed_23684_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 10:08:54.595 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 10:09:11.458 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:09:13.264 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:09:14.178 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:11:08.675 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.20s (1.90min) +2026-06-20 10:11:09.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:13:09.780 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:20:12.683 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 140, train_steps: 5600 +2026-06-20 10:20:49.812 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=140 (missing=0, unexpected=0) +2026-06-20 10:20:51.812 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 10:20:51.938 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:20:51.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:20:52.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:21:59.478 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000140/group_6/index_7988_seed_27710_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 10:22:04.147 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 10:22:20.817 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:22:20.826 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:22:20.827 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:23:28.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000140/group_6/index_7988_seed_27711_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 10:23:33.461 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 10:23:50.328 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:23:53.040 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:23:53.948 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:25:47.600 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-20 10:25:48.313 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:27:45.362 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:34:43.997 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 141, train_steps: 5640 +2026-06-20 10:34:44.690 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 10:34:44.886 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:34:44.971 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:34:45.071 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:35:50.668 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_6/index_3905_seed_15426_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 10:35:54.502 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 10:36:11.716 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:36:11.726 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:36:11.726 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:37:22.977 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_6/index_3905_seed_15427_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 10:37:26.128 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 10:37:43.320 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:37:46.846 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:37:47.854 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:39:41.415 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.51s (1.89min) +2026-06-20 10:39:42.148 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:41:39.462 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:48:38.468 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 142, train_steps: 5680 +2026-06-20 10:48:39.157 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 10:48:39.306 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:48:39.359 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:48:39.460 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:49:46.344 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000142/group_6/index_6226_seed_10444_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 10:49:49.501 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 10:50:06.153 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:50:06.163 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:50:06.163 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:51:19.263 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000142/group_6/index_6226_seed_10445_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 10:51:22.410 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 10:51:40.434 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:51:43.153 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:51:44.140 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:53:37.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-20 10:53:37.851 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:55:37.992 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:02:37.924 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 143, train_steps: 5720 +2026-06-20 11:02:38.634 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 11:02:38.826 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:02:38.919 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:02:38.999 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:03:45.152 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000143/group_6/index_2232_seed_11696_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 11:03:48.375 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 11:04:05.523 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:04:05.534 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:04:05.534 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:05:18.135 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000143/group_6/index_2232_seed_11697_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 11:05:21.274 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 11:05:40.124 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:05:42.169 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:05:43.313 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:07:36.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 11:07:37.313 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:09:36.842 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:16:37.109 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 144, train_steps: 5760 +2026-06-20 11:16:37.800 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 11:16:37.985 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:16:38.084 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:16:38.166 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:17:43.944 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_6/index_4279_seed_9975_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 11:17:47.045 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 11:18:04.252 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:18:04.263 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:18:04.265 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:19:13.194 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_6/index_4279_seed_9976_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 11:19:16.853 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 11:19:34.236 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:19:38.707 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:19:39.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:21:33.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.47s (1.89min) +2026-06-20 11:21:34.087 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:23:30.883 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:30:30.405 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 145, train_steps: 5800 +2026-06-20 11:30:31.098 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 11:30:31.269 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:30:31.352 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:30:31.453 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:31:37.340 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000145/group_6/index_2721_seed_15667_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 11:31:42.495 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 11:31:59.110 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:31:59.119 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:31:59.120 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:33:08.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000145/group_6/index_2721_seed_15668_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 11:33:13.666 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 11:33:30.519 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:33:33.207 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:33:34.177 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:35:28.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.89s (1.90min) +2026-06-20 11:35:28.900 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:37:25.684 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:44:23.915 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 146, train_steps: 5840 +2026-06-20 11:44:24.630 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 11:44:24.844 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:44:24.964 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:44:25.022 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:45:30.730 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_6/index_5776_seed_12894_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 11:45:35.391 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 11:45:51.807 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:45:51.817 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:45:51.820 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:46:59.139 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_6/index_5776_seed_12895_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 11:47:03.787 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 11:47:20.308 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:47:23.247 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:47:24.175 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:49:19.133 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-20 11:49:19.857 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:51:17.038 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:58:15.932 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 147, train_steps: 5880 +2026-06-20 11:58:16.619 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 11:58:16.818 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:58:16.924 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:58:17.018 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:59:23.260 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_6/index_5281_seed_3306_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 11:59:27.215 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 11:59:44.272 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:59:44.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:59:44.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:00:52.583 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_6/index_5281_seed_3307_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 12:00:57.449 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 12:01:14.285 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:01:16.232 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:01:17.224 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:03:12.586 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.29s (1.92min) +2026-06-20 12:03:13.294 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:05:12.903 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:12:12.695 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 148, train_steps: 5920 +2026-06-20 12:12:13.407 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 12:12:13.592 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:12:13.681 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:12:13.756 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:13:19.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000148/group_6/index_8981_seed_32855_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 12:13:24.219 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 12:13:40.797 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:13:40.804 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:13:40.804 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:14:49.607 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000148/group_6/index_8981_seed_32856_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 12:14:54.266 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 12:15:11.307 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:15:13.144 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:15:14.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:17:09.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.69s (1.91min) +2026-06-20 12:17:09.789 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:19:10.586 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:26:11.389 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 149, train_steps: 5960 +2026-06-20 12:26:12.137 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 12:26:12.348 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:26:12.457 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:26:12.521 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:27:18.149 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000149/group_6/index_2435_seed_18928_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 12:27:21.974 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 12:27:39.532 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:27:39.543 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:27:39.543 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:28:48.604 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000149/group_6/index_2435_seed_18929_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 12:28:51.587 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 12:29:08.690 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:29:13.029 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:29:14.039 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:31:08.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.72s (1.91min) +2026-06-20 12:31:09.533 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:33:09.829 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:40:10.994 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 150, train_steps: 6000 +2026-06-20 12:41:53.921 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=150 (missing=0, unexpected=0) +2026-06-20 12:41:56.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 12:41:56.189 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:41:56.239 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:41:56.311 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:43:03.916 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_6/index_5857_seed_1373_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 12:43:06.890 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 12:43:23.950 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:43:23.959 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:43:23.960 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:44:32.531 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_6/index_5857_seed_1374_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 12:44:35.527 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 12:44:52.806 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:44:59.744 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:45:00.794 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:46:55.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.70s (1.91min) +2026-06-20 12:46:56.236 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:48:52.987 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:55:51.602 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 151, train_steps: 6040 +2026-06-20 12:55:52.311 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 12:55:52.528 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:55:52.640 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:55:52.713 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:56:58.386 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_6/index_7596_seed_25677_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 12:57:01.508 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 12:57:18.726 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:57:18.736 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:57:18.736 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:58:27.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_6/index_7596_seed_25678_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 12:58:30.117 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 12:58:47.487 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:58:50.721 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:58:51.732 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:00:46.190 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.41s (1.91min) +2026-06-20 13:00:46.908 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:02:43.651 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:09:41.911 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 152, train_steps: 6080 +2026-06-20 13:09:42.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 13:09:42.773 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:09:42.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:09:42.951 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:10:49.067 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_6/index_899_seed_8609_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 13:10:53.894 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 13:11:10.418 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:11:10.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:11:10.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:12:20.448 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_6/index_899_seed_8610_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 13:12:25.163 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 13:12:42.109 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:12:43.910 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:12:44.874 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:14:38.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-20 13:14:39.492 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:16:38.549 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:23:37.214 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 153, train_steps: 6120 +2026-06-20 13:23:38.094 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 13:23:38.210 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:23:38.251 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:23:38.319 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:24:44.811 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_6/index_5840_seed_17478_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 13:24:48.660 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 13:25:05.684 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:25:05.695 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:25:05.695 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:26:14.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_6/index_5840_seed_17479_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 13:26:17.894 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 13:26:35.382 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:26:40.127 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:26:41.314 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:28:34.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.39s (1.89min) +2026-06-20 13:28:35.473 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:30:32.906 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:37:31.703 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 154, train_steps: 6160 +2026-06-20 13:37:32.394 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 13:37:32.594 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:37:32.696 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:37:32.775 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:38:38.313 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_6/index_5057_seed_29225_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 13:38:42.687 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 13:38:59.720 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:38:59.730 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:38:59.730 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:40:09.339 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_6/index_5057_seed_29226_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 13:40:14.230 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 13:40:31.238 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:40:33.789 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:40:34.768 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:42:29.329 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.47s (1.91min) +2026-06-20 13:42:30.062 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:44:27.158 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:51:24.976 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 155, train_steps: 6200 +2026-06-20 13:51:25.701 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 13:51:25.855 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:51:25.908 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:51:26.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:52:31.808 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_6/index_3583_seed_14049_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 13:52:36.388 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 13:52:52.951 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:52:52.961 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:52:52.961 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:54:01.163 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_6/index_3583_seed_14050_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 13:54:04.930 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 13:54:21.918 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:54:24.447 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:54:25.473 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:56:19.129 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.63s (1.89min) +2026-06-20 13:56:19.859 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:58:16.748 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:05:15.352 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 156, train_steps: 6240 +2026-06-20 14:05:16.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 14:05:16.139 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:05:16.187 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:05:16.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:06:22.461 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_6/index_41_seed_8558_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 14:06:25.514 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 14:06:42.208 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:06:42.218 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:06:42.218 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:07:51.593 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_6/index_41_seed_8559_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 14:07:54.761 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 14:08:11.794 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:08:15.931 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:08:16.937 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:10:11.722 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.74s (1.91min) +2026-06-20 14:10:12.457 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:12:11.103 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:19:12.240 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 157, train_steps: 6280 +2026-06-20 14:19:12.924 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 14:19:13.055 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:19:13.097 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:19:13.186 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:20:19.027 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000157/group_6/index_8999_seed_19585_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 14:20:23.888 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 14:20:40.822 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:20:40.832 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:20:40.832 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:21:47.980 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000157/group_6/index_8999_seed_19586_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 14:21:51.724 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 14:22:09.265 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:22:12.398 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:22:13.527 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:24:08.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.22s (1.92min) +2026-06-20 14:24:09.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:26:09.903 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:33:10.836 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 158, train_steps: 6320 +2026-06-20 14:33:11.521 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 14:33:11.740 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:33:11.851 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:33:11.915 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:34:17.599 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_6/index_6311_seed_6541_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 14:34:20.635 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 14:34:37.468 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:34:37.479 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:34:37.479 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:35:47.266 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_6/index_6311_seed_6542_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 14:35:50.287 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 14:36:07.358 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:36:10.935 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:36:12.022 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:38:07.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.88s (1.91min) +2026-06-20 14:38:07.784 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:40:05.256 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:47:04.576 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 159, train_steps: 6360 +2026-06-20 14:47:05.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 14:47:05.424 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:47:05.462 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:47:05.551 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:48:11.448 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_6/index_8853_seed_32705_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 14:48:16.420 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 14:48:32.926 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:48:32.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:48:32.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:49:40.274 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_6/index_8853_seed_32706_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 14:49:44.909 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 14:50:01.734 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:50:04.049 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:50:05.197 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:51:59.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.86s (1.90min) +2026-06-20 14:51:59.929 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:53:57.323 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:00:59.174 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 160, train_steps: 6400 +2026-06-20 15:01:36.919 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=160 (missing=0, unexpected=0) +2026-06-20 15:01:39.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 15:01:39.394 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:01:39.440 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:01:39.487 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:02:47.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_6/index_3107_seed_8018_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 15:02:52.654 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 15:03:09.344 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:03:09.354 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:03:09.357 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:04:18.892 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_6/index_3107_seed_8019_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 15:04:21.878 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 15:04:39.534 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:04:42.879 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:04:43.892 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:06:37.409 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.48s (1.89min) +2026-06-20 15:06:38.136 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:08:38.301 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:15:37.526 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 161, train_steps: 6440 +2026-06-20 15:15:38.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 15:15:38.407 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:15:38.512 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:15:38.594 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:16:44.458 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_6/index_5837_seed_12355_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 15:16:47.610 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 15:17:04.105 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:17:04.115 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:17:04.115 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:18:16.515 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_6/index_5837_seed_12356_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 15:18:19.593 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 15:18:36.260 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:18:40.375 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:18:41.564 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:20:35.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-20 15:20:36.332 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:22:36.155 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:29:36.902 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 162, train_steps: 6480 +2026-06-20 15:29:37.570 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 15:29:37.761 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:29:37.863 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:29:37.964 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:30:43.651 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_6/index_5656_seed_17855_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 15:30:48.463 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 15:31:04.990 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:31:04.999 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:31:04.999 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:32:11.859 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_6/index_5656_seed_17856_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 15:32:16.483 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 15:32:33.092 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:32:37.444 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:32:38.612 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:34:33.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.48s (1.91min) +2026-06-20 15:34:33.827 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:36:31.316 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:43:31.491 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 163, train_steps: 6520 +2026-06-20 15:43:32.172 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 15:43:32.328 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:43:32.389 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:43:32.504 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:44:38.442 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000163/group_6/index_8385_seed_6503_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 15:44:43.249 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 15:44:59.330 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:44:59.339 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:44:59.340 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:46:09.445 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000163/group_6/index_8385_seed_6504_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 15:46:14.186 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 15:46:30.331 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:46:33.676 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:46:34.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:48:29.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.47s (1.91min) +2026-06-20 15:48:29.934 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:50:26.537 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:57:25.504 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 164, train_steps: 6560 +2026-06-20 15:57:26.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 15:57:26.456 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:57:26.575 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:57:26.666 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:58:32.049 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_6/index_7174_seed_19678_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 15:58:35.686 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 15:58:52.575 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:58:52.582 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:58:52.582 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:00:02.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_6/index_7174_seed_19679_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 16:00:07.568 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 16:00:24.208 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:00:26.402 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:00:27.342 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:02:21.064 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.61s (1.89min) +2026-06-20 16:02:21.793 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:04:18.344 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:11:18.367 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 165, train_steps: 6600 +2026-06-20 16:11:19.088 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 16:11:19.271 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:11:19.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:11:19.470 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:12:25.504 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_6/index_9009_seed_15532_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 16:12:30.203 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 16:12:46.328 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:12:46.338 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:12:46.338 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:13:54.678 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_6/index_9009_seed_15533_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 16:13:58.486 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 16:14:15.224 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:14:19.106 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:14:20.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:16:14.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.75s (1.90min) +2026-06-20 16:16:14.870 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:18:14.124 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:25:13.931 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 166, train_steps: 6640 +2026-06-20 16:25:14.621 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 16:25:14.756 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:25:14.797 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:25:14.890 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:26:21.070 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_6/index_4999_seed_12980_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 16:26:25.911 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 16:26:42.433 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:26:42.442 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:26:42.442 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:27:50.011 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_6/index_4999_seed_12981_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 16:27:54.658 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 16:28:11.289 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:28:13.831 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:28:14.906 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:30:08.844 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.78s (1.90min) +2026-06-20 16:30:09.574 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:32:10.304 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:39:12.150 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 167, train_steps: 6680 +2026-06-20 16:39:12.849 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 16:39:13.050 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:39:13.136 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:39:13.226 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:40:19.049 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_6/index_4636_seed_3708_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 16:40:22.841 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 16:40:41.121 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:40:41.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:40:41.135 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:41:48.215 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_6/index_4636_seed_3709_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 16:41:53.074 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 16:42:10.063 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:42:12.095 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:42:13.258 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:44:08.745 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.36s (1.92min) +2026-06-20 16:44:09.477 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:46:10.018 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:53:11.380 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 168, train_steps: 6720 +2026-06-20 16:53:12.091 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 16:53:12.288 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:53:12.412 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:53:12.499 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:54:18.383 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_6/index_2521_seed_30428_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 16:54:22.850 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 16:54:40.127 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:54:40.137 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:54:40.137 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:55:47.462 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_6/index_2521_seed_30429_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 16:55:51.150 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 16:56:07.911 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:56:11.767 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:56:12.801 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:58:07.736 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-20 16:58:08.464 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:00:07.578 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:07:09.251 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 169, train_steps: 6760 +2026-06-20 17:07:09.881 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 17:07:10.136 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 17:07:10.282 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:07:10.381 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:08:16.515 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_6/index_9575_seed_10915_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 17:08:21.372 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 17:08:37.944 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:08:37.953 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:08:37.954 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:09:48.191 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_6/index_9575_seed_10916_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 17:09:52.927 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 17:10:09.730 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:10:11.904 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 17:10:12.888 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:12:08.255 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.20s (1.92min) +2026-06-20 17:12:08.969 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:14:09.318 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:21:11.045 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 170, train_steps: 6800 +2026-06-20 17:21:47.184 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=170 (missing=0, unexpected=0) +2026-06-20 17:21:49.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 17:21:49.322 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 17:21:49.373 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:21:49.422 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:22:56.501 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_6/index_7679_seed_30336_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 17:23:01.131 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 17:23:17.404 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:23:17.414 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:23:17.417 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:24:25.599 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_6/index_7679_seed_30337_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 17:24:30.201 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 17:24:46.593 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:24:49.925 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 17:24:50.858 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:26:44.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.05s (1.88min) +2026-06-20 17:26:44.743 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:28:41.424 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:35:39.787 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 26, sp_rank 0: update step: 171, train_steps: 6840 +2026-06-20 17:35:40.432 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 26, dp_rank 13: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=6, rank_in_group=1 +2026-06-20 17:35:40.626 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 17:35:40.872 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:35:40.959 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:36:47.796 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000171/group_6/index_4592_seed_31490_dp_13_rank_26_sample_2_slot_0.mp4 +2026-06-20 17:36:51.814 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 17:37:08.265 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:37:08.277 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 26 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:37:08.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 26 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:38:19.301 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 26: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000171/group_6/index_4592_seed_31491_dp_13_rank_26_sample_3_slot_1.mp4 +2026-06-20 17:38:22.468 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 17:38:40.187 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:38:42.537 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 26: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 17:38:43.733 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 26: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:40:36.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 26: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.89s (1.88min) +2026-06-20 17:40:37.378 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 26, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-29 01:30:50.946 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 01:30:51.612 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 01:30:51.615 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 01:53:15.070 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 01:53:15.738 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 01:53:15.739 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 02:22:42.603 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 02:22:43.283 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 02:22:43.286 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 02:48:37.989 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 02:48:38.644 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 02:48:38.644 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:09:53.102 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 03:09:53.780 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:09:53.783 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:33:45.973 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 03:33:46.632 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:33:46.633 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:00:51.589 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 04:00:52.245 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:00:52.247 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:27:42.693 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 04:27:43.353 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:27:43.356 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:54:00.343 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 04:54:01.009 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:54:01.010 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 05:18:09.862 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 05:18:10.532 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 05:18:10.535 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 05:41:14.591 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 05:41:15.264 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 05:41:15.264 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:04:29.327 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 06:04:29.982 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:04:29.985 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:26:59.349 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 06:27:00.022 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:27:00.023 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:51:34.862 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 06:51:35.534 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:51:35.534 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 07:15:17.461 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 07:15:18.135 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 07:15:18.135 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 07:36:27.380 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 07:36:28.033 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 07:36:28.036 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:08:49.859 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 08:08:50.521 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:08:50.522 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:33:27.295 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 08:33:27.975 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:33:27.976 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:01:15.694 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 09:01:16.357 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:01:16.357 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:24:41.705 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 09:24:42.366 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:24:42.368 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:49:41.127 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 09:49:41.789 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:49:41.791 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 10:12:11.200 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 10:12:11.879 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 10:12:11.882 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 10:38:24.785 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 10:38:25.441 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 10:38:25.442 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:07:03.918 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 11:07:04.579 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:07:04.579 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:29:16.775 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 11:29:17.429 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:29:17.429 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:57:16.387 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 11:57:17.056 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:57:17.056 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 12:21:48.990 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 12:21:49.648 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 12:21:49.648 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 12:48:43.192 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 12:48:43.865 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 12:48:43.865 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 13:14:42.910 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 13:14:43.597 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 13:14:43.597 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 13:39:09.852 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 13:39:10.510 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 13:39:10.512 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:05:47.440 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 14:05:48.113 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:05:48.114 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:30:31.199 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 14:30:31.854 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:30:31.855 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:51:45.123 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 14:51:45.790 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:51:45.790 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 15:23:02.386 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 15:23:03.054 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 15:23:03.057 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 15:48:53.767 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 15:48:54.431 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 15:48:54.433 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 16:17:45.200 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 16:17:45.870 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 16:17:45.872 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 16:43:02.854 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 16:43:03.514 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 16:43:03.515 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:11:39.708 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 17:11:40.376 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:11:40.376 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:34:38.509 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 17:34:39.183 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:34:39.183 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:56:06.182 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 17:56:06.865 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:56:06.866 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 18:18:46.100 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 18:18:46.773 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 18:18:46.773 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 18:40:43.579 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 18:40:44.238 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 18:40:44.241 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_28.log b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_28.log new file mode 100644 index 0000000000000000000000000000000000000000..fac55833af13b513f8d79d0bf5c1fba270a9daf3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_28.log @@ -0,0 +1,3413 @@ +2026-06-19 01:22:17.822 | INFO | __main__:__init__:933 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-19 01:22:18.477 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-19 01:22:18.479 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-19 01:30:17.294 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 01:30:31.848 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 01:30:31.851 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:30:31.851 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:31:52.719 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_7/index_7599_seed_8266_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 01:31:56.553 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 01:32:14.180 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:32:14.189 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:32:14.189 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:33:22.320 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_7/index_7599_seed_8267_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 01:33:25.441 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 01:33:42.702 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:33:45.515 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 01:33:51.016 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 01:34:00.951 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 01:35:56.997 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.94s (1.93min) +2026-06-19 01:35:57.655 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 01:37:54.182 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 01:37:57.221 | INFO | hyvideo.utils.grpo_utils:train_one_step:3127 - Rank 28, sp_rank 0: init adaptive KL controller (init=1.000e-07, min=1.000e-07, max=1.000e-05) +2026-06-19 01:45:11.848 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 1, train_steps: 40 +2026-06-19 01:45:12.575 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 01:45:12.728 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 01:45:12.786 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:45:12.837 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:46:18.605 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000001/group_7/index_9764_seed_19707_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 01:46:21.638 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 01:46:40.408 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:46:40.416 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:46:40.416 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:47:47.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000001/group_7/index_9764_seed_19708_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 01:47:51.444 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 01:48:08.810 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:48:10.945 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 01:48:10.998 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 01:48:11.906 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 01:50:07.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.70s (1.93min) +2026-06-19 01:50:08.466 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 01:52:06.953 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 01:59:06.759 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 2, train_steps: 80 +2026-06-19 01:59:07.380 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 01:59:07.475 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 01:59:07.483 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:59:07.493 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:00:13.152 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000002/group_7/index_4618_seed_27257_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 02:00:16.207 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 02:00:33.283 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:00:33.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:00:33.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:01:46.330 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000002/group_7/index_4618_seed_27258_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 02:01:51.323 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 02:02:09.077 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:02:10.528 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:02:10.622 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 02:02:11.727 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:04:06.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.88s (1.91min) +2026-06-19 02:04:07.374 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:06:03.753 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:13:04.853 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 3, train_steps: 120 +2026-06-19 02:13:05.465 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 02:13:05.565 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:13:05.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:13:05.587 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:14:11.184 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_7/index_9149_seed_30243_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 02:14:15.856 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 02:14:33.577 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:14:33.583 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:14:33.583 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:15:40.393 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_7/index_9149_seed_30244_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 02:15:44.226 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 02:16:01.978 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:16:04.655 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:16:04.784 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 02:16:05.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:17:59.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.77s (1.90min) +2026-06-19 02:18:00.448 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:19:59.350 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:26:59.883 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 4, train_steps: 160 +2026-06-19 02:27:00.546 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 02:27:00.653 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:27:00.699 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:27:00.749 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:28:06.734 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000004/group_7/index_8338_seed_31411_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 02:28:11.726 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 02:28:29.512 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:28:29.522 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:28:29.522 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:29:35.197 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000004/group_7/index_8338_seed_31412_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 02:29:40.475 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 02:29:58.534 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:30:00.316 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:30:00.461 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 02:30:01.324 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:31:55.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.62s (1.91min) +2026-06-19 02:31:56.651 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:33:53.222 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:40:53.219 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 5, train_steps: 200 +2026-06-19 02:40:53.901 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 02:40:54.051 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:40:54.105 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:40:54.232 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:42:00.078 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_7/index_3286_seed_20131_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 02:42:04.778 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 02:42:22.399 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:42:22.409 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:42:22.409 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:43:28.745 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_7/index_3286_seed_20132_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 02:43:33.659 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 02:43:51.581 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:43:51.812 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:43:51.941 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 02:43:52.786 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:45:46.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.50s (1.89min) +2026-06-19 02:45:46.994 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:47:43.352 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:54:40.659 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 6, train_steps: 240 +2026-06-19 02:54:41.360 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 02:54:41.475 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:54:41.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:54:41.615 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:55:48.612 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_7/index_2712_seed_22675_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 02:55:52.432 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 02:56:10.335 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:56:10.341 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:56:10.341 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:57:19.380 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_7/index_2712_seed_22676_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 02:57:23.234 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 02:57:40.786 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:57:42.949 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:57:43.088 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 02:57:43.826 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:59:37.354 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.38s (1.89min) +2026-06-19 02:59:38.043 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:01:37.729 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:08:37.478 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 7, train_steps: 280 +2026-06-19 03:08:38.148 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 03:08:38.345 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:08:38.462 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:08:38.541 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:09:44.106 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_7/index_9335_seed_32631_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 03:09:47.147 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 03:10:04.425 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:10:04.435 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:10:04.435 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:11:14.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_7/index_9335_seed_32632_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 03:11:19.792 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 03:11:37.504 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:11:38.337 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:11:38.471 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 03:11:39.606 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:13:34.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.68s (1.91min) +2026-06-19 03:13:34.977 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:15:32.042 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:22:31.149 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 8, train_steps: 320 +2026-06-19 03:22:31.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 03:22:31.946 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:22:31.981 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:22:32.068 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:23:38.728 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000008/group_7/index_9763_seed_21216_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 03:23:42.972 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 03:24:00.473 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:24:00.479 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:24:00.479 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:25:09.055 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000008/group_7/index_9763_seed_21217_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 03:25:14.141 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 03:25:32.149 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:25:33.100 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:25:33.239 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 03:25:34.106 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:27:28.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-19 03:27:29.431 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:29:26.904 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:36:25.642 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 9, train_steps: 360 +2026-06-19 03:36:26.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 03:36:26.560 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:36:26.677 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:36:26.753 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:37:32.621 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_7/index_9813_seed_26727_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 03:37:37.365 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 03:37:54.798 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:37:54.807 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:37:54.807 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:39:01.956 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_7/index_9813_seed_26728_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 03:39:06.729 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 03:39:24.244 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:39:26.010 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:39:26.123 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 03:39:26.994 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:41:20.400 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.36s (1.89min) +2026-06-19 03:41:21.105 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:43:17.561 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:50:16.947 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 10, train_steps: 400 +2026-06-19 03:50:53.287 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=10 (missing=0, unexpected=0) +2026-06-19 03:50:56.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 03:50:56.363 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:50:56.471 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:50:56.580 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:52:03.461 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_7/index_1435_seed_11862_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 03:52:08.266 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 03:52:24.448 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:52:24.456 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:52:24.456 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:53:32.508 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_7/index_1435_seed_11863_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 03:53:37.099 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 03:53:53.588 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:53:57.828 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:53:57.864 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 03:53:58.703 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:55:52.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.65s (1.89min) +2026-06-19 03:55:53.210 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:57:49.762 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:04:48.376 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 11, train_steps: 440 +2026-06-19 04:04:49.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 04:04:49.242 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:04:49.312 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:04:49.414 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:05:55.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_7/index_3839_seed_16052_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 04:05:58.152 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 04:06:15.642 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:06:15.651 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:06:15.651 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:07:24.760 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_7/index_3839_seed_16053_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 04:07:27.616 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 04:07:45.196 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:07:48.853 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:07:48.998 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 04:07:49.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:09:43.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.25s (1.89min) +2026-06-19 04:09:43.797 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:11:40.269 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:18:37.290 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 12, train_steps: 480 +2026-06-19 04:18:37.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 04:18:38.140 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:18:38.224 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:18:38.344 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:19:44.114 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_7/index_8397_seed_23661_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 04:19:47.012 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 04:20:04.298 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:20:04.308 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:20:04.308 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:21:12.808 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_7/index_8397_seed_23662_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 04:21:17.531 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 04:21:35.553 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:21:37.180 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:21:37.315 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 04:21:38.096 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:23:33.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.19s (1.92min) +2026-06-19 04:23:34.071 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:25:31.242 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:32:29.816 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 13, train_steps: 520 +2026-06-19 04:32:30.513 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 04:32:30.687 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:32:30.774 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:32:30.897 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:33:36.790 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_7/index_4505_seed_26748_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 04:33:39.983 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 04:33:57.618 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:33:57.625 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:33:57.625 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:35:06.166 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_7/index_4505_seed_26749_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 04:35:11.257 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 04:35:29.517 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:35:30.313 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:35:30.444 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 04:35:31.319 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:37:24.102 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.77s (1.88min) +2026-06-19 04:37:24.791 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:39:21.077 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:46:18.583 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 14, train_steps: 560 +2026-06-19 04:46:19.247 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 04:46:19.462 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:46:19.609 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:46:19.689 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:47:25.786 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_7/index_6678_seed_31784_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 04:47:30.523 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 04:47:48.048 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:47:48.054 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:47:48.054 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:48:54.212 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_7/index_6678_seed_31785_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 04:48:58.327 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 04:49:16.027 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:49:17.445 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:49:17.562 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 04:49:18.354 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:51:12.963 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.49s (1.91min) +2026-06-19 04:51:13.650 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:53:12.995 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:00:11.830 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 15, train_steps: 600 +2026-06-19 05:00:12.525 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 05:00:12.707 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:00:12.803 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:00:12.915 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:01:18.880 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000015/group_7/index_9584_seed_9621_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 05:01:21.932 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 05:01:40.811 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:01:40.820 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:01:40.820 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:02:47.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000015/group_7/index_9584_seed_9622_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 05:02:50.975 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 05:03:08.563 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:03:11.349 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:03:11.419 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 05:03:12.504 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:05:07.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.64s (1.91min) +2026-06-19 05:05:08.061 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:07:05.977 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:14:04.975 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 16, train_steps: 640 +2026-06-19 05:14:05.573 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 05:14:05.687 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:14:05.720 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:14:05.768 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:15:11.582 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_7/index_8728_seed_3853_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 05:15:14.690 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 05:15:31.866 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:15:31.875 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:15:31.876 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:16:43.003 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_7/index_8728_seed_3854_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 05:16:48.128 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 05:17:05.691 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:17:06.988 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:17:07.099 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 05:17:07.970 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:19:03.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.02s (1.92min) +2026-06-19 05:19:03.741 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:21:01.962 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:28:03.316 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 17, train_steps: 680 +2026-06-19 05:28:04.006 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 05:28:04.130 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:28:04.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:28:04.240 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:29:10.207 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_7/index_8287_seed_15154_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 05:29:13.987 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 05:29:31.359 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:29:31.368 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:29:31.368 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:30:39.478 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_7/index_8287_seed_15155_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 05:30:44.638 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 05:31:02.393 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:31:03.926 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:31:04.033 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 05:31:05.053 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:32:58.689 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.48s (1.89min) +2026-06-19 05:32:59.387 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:34:56.265 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:41:58.095 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 18, train_steps: 720 +2026-06-19 05:41:58.777 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 05:41:58.959 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:41:59.071 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:41:59.181 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:43:04.949 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_7/index_2825_seed_30778_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 05:43:08.925 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 05:43:26.396 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:43:26.405 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:43:26.407 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:44:35.363 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_7/index_2825_seed_30779_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 05:44:38.613 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 05:44:57.673 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:45:00.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:45:00.339 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 05:45:01.114 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:46:55.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.88s (1.90min) +2026-06-19 05:46:55.797 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:48:52.329 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:55:50.496 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 19, train_steps: 760 +2026-06-19 05:55:51.177 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 05:55:51.278 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:55:51.314 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:55:51.361 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:56:57.512 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_7/index_3009_seed_26885_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 05:57:02.607 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 05:57:20.454 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:57:20.463 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:57:20.463 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:58:25.697 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_7/index_3009_seed_26886_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 05:58:30.596 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 05:58:48.396 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:58:49.068 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:58:49.194 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 05:58:49.966 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:00:43.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.10s (1.89min) +2026-06-19 06:00:43.876 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:02:40.341 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:09:39.755 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 20, train_steps: 800 +2026-06-19 06:10:17.110 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=20 (missing=0, unexpected=0) +2026-06-19 06:10:18.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 06:10:18.968 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:10:19.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:10:19.136 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:11:26.532 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_7/index_8665_seed_15911_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 06:11:31.411 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 06:11:49.000 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:11:49.009 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:11:49.011 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:12:54.780 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_7/index_8665_seed_15912_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 06:12:59.638 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 06:13:17.368 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:13:19.050 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:13:19.175 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 06:13:20.169 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:15:13.224 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.90s (1.88min) +2026-06-19 06:15:13.910 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:17:12.688 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:24:12.560 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 21, train_steps: 840 +2026-06-19 06:24:13.271 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 06:24:13.440 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:24:13.515 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:24:13.612 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:25:19.162 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_7/index_9748_seed_6100_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 06:25:23.889 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 06:25:41.518 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:25:41.527 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:25:41.527 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:26:47.701 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_7/index_9748_seed_6101_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 06:26:51.816 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 06:27:09.894 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:27:11.261 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:27:11.372 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 06:27:12.260 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:29:06.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.51s (1.91min) +2026-06-19 06:29:07.668 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:31:04.578 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:38:04.279 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 22, train_steps: 880 +2026-06-19 06:38:04.948 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 06:38:05.157 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:38:05.285 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:38:05.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:39:11.041 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_7/index_980_seed_30193_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 06:39:16.188 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 06:39:34.048 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:39:34.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:39:34.057 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:40:39.943 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_7/index_980_seed_30194_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 06:40:44.788 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 06:41:02.686 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:41:03.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:41:03.196 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 06:41:04.336 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:42:58.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.99s (1.90min) +2026-06-19 06:42:59.164 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:44:56.505 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:51:58.166 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 23, train_steps: 920 +2026-06-19 06:51:58.842 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 06:51:58.976 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:51:58.995 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:51:59.066 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:53:04.817 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000023/group_7/index_6384_seed_29768_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 06:53:09.651 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 06:53:27.358 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:53:27.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:53:27.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:54:34.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000023/group_7/index_6384_seed_29769_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 06:54:39.205 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 06:54:58.063 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:54:59.789 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:54:59.936 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 06:55:00.725 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:56:54.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.11s (1.90min) +2026-06-19 06:56:55.622 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:58:52.319 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:05:51.318 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 24, train_steps: 960 +2026-06-19 07:05:52.006 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 07:05:52.211 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:05:52.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:05:52.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:06:58.169 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_7/index_3392_seed_29718_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 07:07:01.263 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 07:07:18.822 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:07:18.828 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:07:18.829 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:08:27.610 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_7/index_3392_seed_29719_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 07:08:32.799 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 07:08:51.008 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:08:52.217 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:08:52.349 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 07:08:53.210 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:10:46.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-19 07:10:46.871 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:12:44.153 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:19:41.973 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 25, train_steps: 1000 +2026-06-19 07:20:52.105 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 07:20:52.299 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:20:52.412 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:20:52.495 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:21:58.415 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_7/index_6896_seed_8720_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 07:22:02.611 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 07:22:20.922 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:22:20.927 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:22:20.927 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:23:28.851 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_7/index_6896_seed_8721_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 07:23:33.880 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 07:23:52.318 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:23:52.666 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:23:52.728 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 07:23:53.674 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:25:46.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.81s (1.88min) +2026-06-19 07:25:47.206 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:27:44.380 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:34:43.675 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 26, train_steps: 1040 +2026-06-19 07:34:44.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 07:34:44.449 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:34:44.492 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:34:44.539 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:35:50.243 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_7/index_4002_seed_18276_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 07:35:55.027 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 07:36:13.123 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:36:13.132 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:36:13.132 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:37:20.909 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_7/index_4002_seed_18277_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 07:37:25.734 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 07:37:44.146 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:37:45.443 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:37:45.574 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 07:37:46.443 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:39:39.639 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.14s (1.89min) +2026-06-19 07:39:40.348 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:41:39.072 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:48:39.672 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 27, train_steps: 1080 +2026-06-19 07:48:40.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 07:48:40.544 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:48:40.669 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:48:40.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:49:47.852 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_7/index_7479_seed_26136_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 07:49:52.650 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 07:50:10.783 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:50:10.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:50:10.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:51:19.235 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_7/index_7479_seed_26137_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 07:51:24.087 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 07:51:42.484 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:51:43.285 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:51:43.415 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 07:51:44.285 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:53:38.298 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.99s (1.90min) +2026-06-19 07:53:39.007 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:55:38.423 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:02:38.169 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 28, train_steps: 1120 +2026-06-19 08:02:38.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 08:02:39.017 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:02:39.053 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:02:39.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:03:45.535 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_7/index_5718_seed_11900_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 08:03:50.454 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 08:04:08.769 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:04:08.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:04:08.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:05:17.612 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_7/index_5718_seed_11901_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 08:05:21.727 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 08:05:40.908 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:05:41.901 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:05:42.031 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 08:05:43.028 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:07:37.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.24s (1.90min) +2026-06-19 08:07:38.091 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:09:37.933 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:16:37.855 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 29, train_steps: 1160 +2026-06-19 08:16:38.691 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 08:16:38.825 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:16:38.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:16:39.108 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:17:44.842 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_7/index_4247_seed_11172_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 08:17:49.871 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 08:18:09.471 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:18:09.480 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:18:09.480 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:19:18.191 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_7/index_4247_seed_11173_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 08:19:22.340 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 08:19:41.588 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:19:42.881 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:19:42.944 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 08:19:43.774 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:21:37.264 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 08:21:37.961 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:23:37.406 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:30:39.615 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 30, train_steps: 1200 +2026-06-19 08:31:15.807 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=30 (missing=0, unexpected=0) +2026-06-19 08:31:18.259 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 08:31:18.446 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:31:18.555 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:31:18.635 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:32:26.399 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000030/group_7/index_1115_seed_27292_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 08:32:29.444 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 08:32:47.341 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:32:47.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:32:47.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:33:55.693 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000030/group_7/index_1115_seed_27293_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 08:33:58.860 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 08:34:17.000 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:34:19.980 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:34:20.131 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 08:34:21.039 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:36:14.510 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-19 08:36:15.227 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:38:14.172 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:45:14.285 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 31, train_steps: 1240 +2026-06-19 08:45:14.992 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 08:45:15.179 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:45:15.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:45:15.392 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:46:21.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_7/index_5156_seed_19054_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 08:46:24.065 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 08:46:42.028 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:46:42.037 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:46:42.037 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:47:50.702 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_7/index_5156_seed_19055_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 08:47:53.754 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 08:48:11.834 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:48:15.288 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:48:15.432 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 08:48:16.334 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:50:10.323 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.84s (1.90min) +2026-06-19 08:50:11.023 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:52:11.149 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:59:13.095 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 32, train_steps: 1280 +2026-06-19 08:59:13.782 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 08:59:13.992 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:59:14.106 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:59:14.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:00:20.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000032/group_7/index_8255_seed_16740_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 09:00:23.252 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 09:00:42.273 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:00:42.282 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:00:42.282 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:01:48.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000032/group_7/index_8255_seed_16741_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 09:01:53.883 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 09:02:13.426 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:02:13.436 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:02:13.489 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 09:02:14.657 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:04:08.713 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.91s (1.90min) +2026-06-19 09:04:09.420 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:06:10.407 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:13:13.271 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 33, train_steps: 1320 +2026-06-19 09:13:13.952 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 09:13:14.127 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:13:14.206 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:13:14.316 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:14:20.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000033/group_7/index_1567_seed_29388_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 09:14:25.052 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 09:14:43.106 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:14:43.116 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:14:43.116 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:15:50.094 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000033/group_7/index_1567_seed_29389_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 09:15:53.740 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 09:16:12.178 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:16:14.822 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:16:14.965 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 09:16:15.817 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:18:11.066 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.06s (1.92min) +2026-06-19 09:18:11.785 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:20:11.625 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:27:13.735 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 34, train_steps: 1360 +2026-06-19 09:27:14.445 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 09:27:14.637 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:27:14.749 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:27:14.854 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:28:20.504 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_7/index_4110_seed_23549_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 09:28:25.559 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 09:28:44.001 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:28:44.010 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:28:44.011 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:29:49.869 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_7/index_4110_seed_23550_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 09:29:53.559 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 09:30:11.995 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:30:13.695 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:30:13.755 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 09:30:14.858 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:32:08.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.81s (1.90min) +2026-06-19 09:32:09.592 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:34:09.778 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:41:12.391 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 35, train_steps: 1400 +2026-06-19 09:41:13.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 09:41:13.160 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:41:13.195 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:41:13.243 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:42:19.212 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000035/group_7/index_9612_seed_1509_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 09:42:22.169 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 09:42:40.799 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:42:40.808 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:42:40.808 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:43:48.609 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000035/group_7/index_9612_seed_1510_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 09:43:53.853 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 09:44:12.494 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:44:13.080 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:44:13.168 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 09:44:14.104 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:46:10.686 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 116.56s (1.94min) +2026-06-19 09:46:11.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:48:10.895 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:55:11.959 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 36, train_steps: 1440 +2026-06-19 09:55:12.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 09:55:12.792 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:55:12.865 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:55:12.985 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:56:18.979 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_7/index_9993_seed_20994_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 09:56:23.753 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 09:56:42.030 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:56:42.036 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:56:42.036 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:57:48.119 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_7/index_9993_seed_20995_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 09:57:52.937 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 09:58:11.083 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:58:12.646 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:58:12.757 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 09:58:13.851 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:00:08.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-19 10:00:09.671 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:02:10.007 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:09:11.699 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 37, train_steps: 1480 +2026-06-19 10:09:12.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 10:09:12.589 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:09:12.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:09:12.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:10:18.575 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_7/index_883_seed_4565_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 10:10:23.351 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 10:10:41.538 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:10:41.547 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:10:41.548 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:11:47.899 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_7/index_883_seed_4566_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 10:11:51.672 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 10:12:09.993 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:12:13.042 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:12:13.153 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 10:12:13.956 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:14:09.030 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.94s (1.92min) +2026-06-19 10:14:09.737 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:16:10.368 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:23:12.355 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 38, train_steps: 1520 +2026-06-19 10:23:13.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 10:23:13.147 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:23:13.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:23:13.243 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:24:19.240 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000038/group_7/index_2559_seed_1507_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 10:24:24.435 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 10:24:42.707 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:24:42.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:24:42.717 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:25:48.666 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000038/group_7/index_2559_seed_1508_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 10:25:53.411 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 10:26:11.780 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:26:12.667 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:26:12.687 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 10:26:13.877 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:28:09.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.16s (1.92min) +2026-06-19 10:28:09.954 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:30:10.746 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:37:11.942 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 39, train_steps: 1560 +2026-06-19 10:37:12.646 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 10:37:12.823 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:37:12.927 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:37:13.045 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:38:19.039 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000039/group_7/index_3930_seed_13380_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 10:38:23.003 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 10:38:41.116 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:38:41.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:38:41.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:41:17.602 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000039/group_7/index_3930_seed_13381_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 10:41:20.411 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 10:41:39.230 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:41:42.526 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:41:42.633 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 10:41:43.634 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:43:36.471 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.70s (1.88min) +2026-06-19 10:43:37.180 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:45:37.175 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:52:37.854 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 40, train_steps: 1600 +2026-06-19 10:53:14.164 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=40 (missing=0, unexpected=0) +2026-06-19 10:53:15.773 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 10:53:15.964 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:53:16.075 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:53:16.134 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:54:23.621 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_7/index_6349_seed_21465_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 10:54:26.499 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 10:54:44.148 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:54:44.158 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:54:44.158 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:55:52.480 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_7/index_6349_seed_21466_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 10:55:57.620 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 10:56:15.782 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:56:16.410 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:56:16.554 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 10:56:17.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:58:11.717 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.27s (1.90min) +2026-06-19 10:58:12.416 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:00:11.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:07:12.580 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 41, train_steps: 1640 +2026-06-19 11:07:13.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 11:07:13.473 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:07:13.593 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:07:13.701 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:08:19.459 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_7/index_1241_seed_13759_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 11:08:24.302 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 11:08:42.696 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:08:42.705 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:08:42.705 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:09:48.763 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_7/index_1241_seed_13760_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 11:09:52.915 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 11:10:10.978 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:10:12.668 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:10:12.773 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 11:10:13.777 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:12:08.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.28s (1.90min) +2026-06-19 11:12:09.087 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:14:08.757 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:21:10.883 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 42, train_steps: 1680 +2026-06-19 11:21:11.804 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 11:21:11.932 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:21:11.961 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:21:12.028 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:22:18.153 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000042/group_7/index_1454_seed_29054_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 11:22:23.406 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 11:22:41.690 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:22:41.698 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:22:41.698 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:23:47.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000042/group_7/index_1454_seed_29055_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 11:23:51.261 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 11:24:09.585 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:24:12.073 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:24:12.191 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 11:24:13.152 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:26:07.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.32s (1.91min) +2026-06-19 11:26:08.279 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:28:06.794 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:35:06.840 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 43, train_steps: 1720 +2026-06-19 11:35:07.458 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 11:35:07.584 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:35:07.612 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:35:07.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:36:13.686 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_7/index_3572_seed_3089_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 11:36:18.878 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 11:36:36.795 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:36:36.804 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:36:36.805 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:37:46.409 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_7/index_3572_seed_3090_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 11:37:50.421 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 11:38:08.308 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:38:10.299 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:38:10.421 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 11:38:11.214 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:40:05.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.38s (1.91min) +2026-06-19 11:40:06.549 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:42:03.523 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:49:04.428 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 44, train_steps: 1760 +2026-06-19 11:49:05.113 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 11:49:05.231 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:49:05.260 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:49:05.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:50:11.427 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000044/group_7/index_8774_seed_17447_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 11:50:14.475 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 11:50:32.426 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:50:32.435 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:50:32.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:51:41.848 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000044/group_7/index_8774_seed_17448_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 11:51:44.796 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 11:52:02.750 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:52:06.491 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:52:06.616 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 11:52:07.380 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:54:01.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.69s (1.89min) +2026-06-19 11:54:01.910 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:56:01.233 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:03:03.874 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 45, train_steps: 1800 +2026-06-19 12:03:04.556 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 12:03:04.684 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:03:04.729 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:03:04.806 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:04:10.816 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000045/group_7/index_8755_seed_21604_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 12:04:15.872 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 12:04:34.064 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:04:34.070 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:04:34.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:05:40.402 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000045/group_7/index_8755_seed_21605_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 12:05:44.628 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 12:06:02.889 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:06:05.347 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:06:05.485 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 12:06:06.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:08:01.417 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.05s (1.92min) +2026-06-19 12:08:02.101 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:10:01.381 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:17:03.585 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 46, train_steps: 1840 +2026-06-19 12:17:04.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 12:17:04.450 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:17:04.523 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:17:04.640 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:18:10.610 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_7/index_220_seed_12920_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 12:18:15.782 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 12:18:34.190 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:18:34.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:18:34.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:19:40.657 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_7/index_220_seed_12921_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 12:19:44.748 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 12:20:03.351 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:20:05.333 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:20:05.417 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 12:20:06.472 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:22:01.733 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.17s (1.92min) +2026-06-19 12:22:02.442 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:24:01.449 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:31:02.925 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 47, train_steps: 1880 +2026-06-19 12:31:03.607 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 12:31:03.729 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:31:03.768 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:31:03.840 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:32:09.504 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000047/group_7/index_8301_seed_4154_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 12:32:12.382 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 12:32:30.497 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:32:30.506 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:32:30.508 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:33:39.397 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000047/group_7/index_8301_seed_4155_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 12:33:44.614 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 12:34:03.172 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:34:03.182 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:34:03.194 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 12:34:04.363 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:35:58.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-19 12:35:58.982 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:37:55.890 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:44:57.681 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 48, train_steps: 1920 +2026-06-19 12:44:58.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 12:44:58.559 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:44:58.688 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:44:58.775 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:46:04.682 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_7/index_319_seed_10819_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 12:46:08.783 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 12:46:26.756 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:46:26.766 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:46:26.768 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:47:34.760 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_7/index_319_seed_10820_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 12:47:40.474 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 12:47:58.795 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:47:58.931 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:47:58.942 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 12:47:59.978 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:49:54.620 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-19 12:49:55.343 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:51:52.288 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:58:52.085 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 49, train_steps: 1960 +2026-06-19 12:58:52.782 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 12:58:52.906 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:58:52.950 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:58:53.019 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:59:58.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_7/index_9039_seed_23557_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 13:00:03.005 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 13:00:20.878 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:00:20.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:00:20.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:01:30.147 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_7/index_9039_seed_23558_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 13:01:35.358 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 13:01:53.790 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:01:54.369 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:01:54.448 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 13:01:55.379 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:03:48.446 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.04s (1.88min) +2026-06-19 13:03:49.148 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:05:45.465 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:12:45.274 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 50, train_steps: 2000 +2026-06-19 13:14:21.270 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=50 (missing=0, unexpected=0) +2026-06-19 13:14:23.203 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 13:14:23.569 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:14:23.629 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:14:23.642 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:15:31.440 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_7/index_5162_seed_10638_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 13:15:36.300 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 13:15:54.401 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:15:54.410 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:15:54.410 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:17:00.982 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_7/index_5162_seed_10639_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 13:17:05.762 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 13:17:23.822 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:17:25.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:17:25.339 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 13:17:26.205 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:19:20.081 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.84s (1.90min) +2026-06-19 13:19:20.807 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:21:17.523 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:28:18.462 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 51, train_steps: 2040 +2026-06-19 13:28:19.184 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 13:28:19.304 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:28:19.343 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:28:19.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:29:25.263 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000051/group_7/index_461_seed_27679_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 13:29:30.077 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 13:29:48.320 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:29:48.329 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:29:48.329 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:30:53.963 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000051/group_7/index_461_seed_27680_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 13:30:58.892 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 13:31:17.444 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:31:18.029 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:31:18.188 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 13:31:19.065 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:33:13.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.83s (1.90min) +2026-06-19 13:33:13.742 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:35:13.228 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:42:13.609 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 52, train_steps: 2080 +2026-06-19 13:42:14.277 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 13:42:14.468 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:42:14.562 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:42:14.659 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:43:21.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_7/index_885_seed_21556_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 13:43:25.885 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 13:43:44.038 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:43:44.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:43:44.049 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:44:50.119 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_7/index_885_seed_21557_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 13:44:54.402 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 13:45:12.894 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:45:15.287 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:45:15.307 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 13:45:16.336 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:47:10.481 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.03s (1.90min) +2026-06-19 13:47:11.197 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:49:11.263 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:56:12.598 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 53, train_steps: 2120 +2026-06-19 13:56:13.305 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 13:56:13.517 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:56:13.630 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:56:13.723 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:57:19.701 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_7/index_4701_seed_3969_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 13:57:22.789 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 13:57:41.049 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:57:41.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:57:41.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:58:48.779 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_7/index_4701_seed_3970_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 13:58:52.461 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 13:59:10.821 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:59:13.130 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:59:13.228 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 13:59:14.050 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:01:08.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.41s (1.91min) +2026-06-19 14:01:09.529 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:03:10.473 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:10:12.231 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 54, train_steps: 2160 +2026-06-19 14:10:12.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 14:10:13.004 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:10:13.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:10:13.115 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:11:18.948 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_7/index_6946_seed_27632_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 14:11:21.863 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 14:11:40.772 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:11:40.781 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:11:40.781 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:12:47.852 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_7/index_6946_seed_27633_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 14:12:52.798 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 14:13:11.114 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:13:12.866 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:13:12.995 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 14:13:13.883 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:15:08.493 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.42s (1.91min) +2026-06-19 14:15:09.218 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:17:09.305 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:24:11.904 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 55, train_steps: 2200 +2026-06-19 14:24:12.602 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 14:24:12.770 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:24:12.840 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:24:12.960 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:25:18.926 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_7/index_6874_seed_29258_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 14:25:23.684 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 14:25:42.043 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:25:42.049 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:25:42.049 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:26:48.278 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_7/index_6874_seed_29259_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 14:26:52.227 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 14:27:10.563 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:27:12.754 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:27:12.903 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 14:27:13.854 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:29:09.888 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.82s (1.93min) +2026-06-19 14:29:10.604 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:31:10.762 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:38:12.109 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 56, train_steps: 2240 +2026-06-19 14:38:12.811 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 14:38:13.004 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:38:13.124 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:38:13.204 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:39:19.135 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_7/index_5754_seed_25259_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 14:39:22.025 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 14:39:40.787 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:39:40.796 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:39:40.797 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:40:48.963 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_7/index_5754_seed_25260_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 14:40:54.168 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 14:41:12.753 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:41:13.092 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:41:13.164 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 14:41:14.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:43:08.781 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.66s (1.91min) +2026-06-19 14:43:09.475 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:45:09.716 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:52:11.846 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 57, train_steps: 2280 +2026-06-19 14:52:12.527 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 14:52:12.712 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:52:12.818 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:52:12.915 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:53:18.709 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_7/index_9633_seed_23391_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 14:53:23.592 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 14:53:42.077 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:53:42.086 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:53:42.086 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:54:48.774 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_7/index_9633_seed_23392_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 14:54:53.640 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 14:55:12.344 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:55:13.057 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:55:13.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 14:55:14.069 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:57:08.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.79s (1.91min) +2026-06-19 14:57:09.597 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:59:09.880 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:06:12.284 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 58, train_steps: 2320 +2026-06-19 15:06:12.986 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 15:06:13.176 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:06:13.281 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:06:13.364 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:07:19.451 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_7/index_1219_seed_3083_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 15:07:23.342 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 15:07:41.168 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:07:41.178 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:07:41.178 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:08:49.047 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_7/index_1219_seed_3084_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 15:08:51.977 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 15:09:09.966 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:09:13.062 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:09:13.170 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 15:09:14.003 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:11:08.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.71s (1.91min) +2026-06-19 15:11:09.498 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:13:09.809 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:20:12.055 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 59, train_steps: 2360 +2026-06-19 15:20:12.737 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 15:20:12.855 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:20:12.890 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:20:12.953 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:21:18.934 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_7/index_958_seed_29498_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 15:21:24.034 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 15:21:41.951 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:21:41.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:21:41.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:22:48.008 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_7/index_958_seed_29499_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 15:22:52.820 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 15:23:10.876 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:23:12.372 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:23:12.514 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 15:23:13.536 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:25:09.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.28s (1.92min) +2026-06-19 15:25:09.743 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:27:10.365 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:34:12.064 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 60, train_steps: 2400 +2026-06-19 15:34:48.189 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=60 (missing=0, unexpected=0) +2026-06-19 15:34:50.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 15:34:50.799 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:34:50.895 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:34:50.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:35:58.562 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_7/index_6082_seed_11620_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 15:36:02.639 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 15:36:20.547 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:36:20.556 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:36:20.556 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:37:28.474 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_7/index_6082_seed_11621_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 15:37:33.823 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 15:37:52.240 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:37:53.529 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:37:53.667 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 15:37:54.736 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:39:47.707 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-19 15:39:48.403 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:41:45.232 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:48:44.589 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 61, train_steps: 2440 +2026-06-19 15:48:45.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 15:48:45.424 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:48:45.465 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:48:45.563 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:49:51.614 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_7/index_3018_seed_25249_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 15:49:56.468 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 15:50:14.754 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:50:14.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:50:14.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:51:23.489 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_7/index_3018_seed_25250_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 15:51:27.585 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 15:51:46.209 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:51:47.609 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:51:47.744 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 15:51:48.621 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:53:41.681 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.04s (1.88min) +2026-06-19 15:53:42.399 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:55:39.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:02:39.004 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 62, train_steps: 2480 +2026-06-19 16:02:39.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 16:02:39.798 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:02:39.843 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:02:39.893 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:03:46.575 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000062/group_7/index_9347_seed_1918_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 16:03:50.662 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 16:04:08.637 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:04:08.645 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:04:08.646 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:05:18.820 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000062/group_7/index_9347_seed_1919_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 16:05:21.867 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 16:05:40.860 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:05:44.132 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:05:44.280 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 16:05:45.053 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:07:38.761 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.59s (1.89min) +2026-06-19 16:07:39.471 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:09:38.852 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:16:39.314 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 63, train_steps: 2520 +2026-06-19 16:16:40.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 16:16:40.243 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:16:40.339 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:16:40.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:17:47.388 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_7/index_6248_seed_27864_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 16:17:52.649 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 16:18:10.897 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:18:10.906 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:18:10.906 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:19:19.830 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_7/index_6248_seed_27865_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 16:19:24.796 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 16:19:43.474 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:19:45.197 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:19:45.294 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 16:19:46.207 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:21:39.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.36s (1.89min) +2026-06-19 16:21:40.303 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:23:38.881 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:30:40.925 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 64, train_steps: 2560 +2026-06-19 16:30:41.626 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 16:30:41.840 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:30:41.944 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:30:42.009 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:31:48.358 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_7/index_9334_seed_26017_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 16:31:53.261 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 16:32:11.399 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:32:11.408 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:32:11.409 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:33:20.929 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_7/index_9334_seed_26018_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 16:33:25.811 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 16:33:44.271 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:33:45.471 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:33:45.526 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 16:33:46.484 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:35:40.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.63s (1.89min) +2026-06-19 16:35:40.844 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:37:39.397 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:44:41.291 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 65, train_steps: 2600 +2026-06-19 16:44:41.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 16:44:42.140 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:44:42.227 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:44:42.347 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:45:49.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_7/index_9382_seed_28302_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 16:45:53.945 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 16:46:13.325 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:46:13.334 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:46:13.334 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:47:22.690 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_7/index_9382_seed_28303_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 16:47:27.581 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 16:47:47.474 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:47:47.482 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:47:47.494 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 16:47:48.497 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:49:42.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.72s (1.90min) +2026-06-19 16:49:42.959 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:51:40.035 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:58:41.270 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 66, train_steps: 2640 +2026-06-19 16:58:41.989 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 16:58:42.124 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:58:42.169 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:58:42.281 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:59:48.640 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_7/index_2001_seed_17926_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 16:59:53.462 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 17:00:12.016 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:00:12.025 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:00:12.025 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:01:19.582 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_7/index_2001_seed_17927_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 17:01:24.426 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 17:01:43.038 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:01:43.599 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:01:43.666 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 17:01:44.594 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:03:37.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.34s (1.89min) +2026-06-19 17:03:38.682 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:05:38.381 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:12:40.119 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 67, train_steps: 2680 +2026-06-19 17:12:40.908 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 17:12:41.227 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:12:41.311 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:12:41.469 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:13:48.306 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_7/index_6448_seed_14522_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 17:13:52.601 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 17:14:10.665 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:14:10.673 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:14:10.674 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:15:20.448 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_7/index_6448_seed_14523_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 17:15:23.688 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 17:15:41.804 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:15:44.529 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:15:44.661 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 17:15:45.457 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:17:38.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 17:17:39.679 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:19:39.111 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:26:39.543 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 68, train_steps: 2720 +2026-06-19 17:26:40.215 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 17:26:40.379 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:26:40.459 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:26:40.581 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:27:47.626 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_7/index_8843_seed_24263_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 17:27:52.724 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 17:28:11.170 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:28:11.176 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:28:11.176 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:29:20.122 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_7/index_8843_seed_24264_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 17:29:24.950 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 17:29:43.518 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:29:43.886 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:29:43.951 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 17:29:44.898 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:31:38.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.76s (1.90min) +2026-06-19 17:31:39.420 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:33:38.461 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:40:38.369 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 69, train_steps: 2760 +2026-06-19 17:40:39.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 17:40:39.131 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:40:39.170 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:40:39.218 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:41:45.581 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000069/group_7/index_6290_seed_16746_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 17:41:49.862 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 17:42:07.917 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:42:07.923 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:42:07.923 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:43:16.950 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000069/group_7/index_6290_seed_16747_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 17:43:19.965 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 17:43:38.632 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:43:41.069 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:43:41.172 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 17:43:42.077 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:45:35.854 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.55s (1.89min) +2026-06-19 17:45:36.570 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:47:37.115 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:54:38.571 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 70, train_steps: 2800 +2026-06-19 17:55:14.367 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=70 (missing=0, unexpected=0) +2026-06-19 17:55:16.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 17:55:17.142 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:55:17.209 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:55:17.309 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:56:25.517 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_7/index_1640_seed_921_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 17:56:30.794 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 17:56:48.772 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:56:48.781 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:56:48.781 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:57:54.463 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_7/index_1640_seed_922_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 17:57:58.379 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 17:58:16.561 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:58:18.228 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:58:18.312 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 17:58:19.347 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:00:13.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.58s (1.89min) +2026-06-19 18:00:13.838 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:02:13.851 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:09:13.890 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 71, train_steps: 2840 +2026-06-19 18:09:14.615 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 18:09:14.779 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:09:14.866 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:09:14.982 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:10:20.888 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_7/index_9153_seed_3716_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 18:10:26.181 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 18:10:45.975 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:10:45.985 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:10:45.985 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:11:51.686 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_7/index_9153_seed_3717_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 18:11:55.719 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 18:12:15.237 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:12:16.021 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:12:16.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 18:12:16.974 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:14:12.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.67s (1.93min) +2026-06-19 18:14:13.460 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:16:12.691 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:23:13.830 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 72, train_steps: 2880 +2026-06-19 18:23:14.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 18:23:14.617 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:23:14.668 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:23:14.720 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:24:20.889 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_7/index_5665_seed_22730_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 18:24:24.061 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 18:24:42.199 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:24:42.209 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:24:42.210 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:25:51.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_7/index_5665_seed_22731_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 18:25:54.340 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 18:26:12.369 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:26:14.711 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:26:14.859 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 18:26:15.758 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:28:10.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.45s (1.91min) +2026-06-19 18:28:11.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:30:10.987 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:37:12.784 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 73, train_steps: 2920 +2026-06-19 18:37:13.456 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 18:37:13.626 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:37:13.709 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:37:13.841 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:38:19.901 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_7/index_6858_seed_6952_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 18:38:22.941 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 18:38:40.926 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:38:40.932 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:38:40.933 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:39:49.865 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_7/index_6858_seed_6953_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 18:39:53.009 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 18:40:11.300 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:40:13.249 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:40:13.371 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 18:40:14.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:42:09.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.16s (1.92min) +2026-06-19 18:42:10.284 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:44:10.445 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:51:12.757 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 74, train_steps: 2960 +2026-06-19 18:51:13.469 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 18:51:13.679 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:51:13.806 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:51:13.904 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:52:19.973 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_7/index_5502_seed_3130_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 18:52:23.059 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 18:52:41.276 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:52:41.285 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:52:41.285 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:53:49.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_7/index_5502_seed_3131_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 18:53:52.791 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 18:54:12.027 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:54:13.982 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:54:14.084 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 18:54:14.870 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:56:10.312 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.21s (1.92min) +2026-06-19 18:56:11.031 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:58:11.441 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:05:14.022 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 75, train_steps: 3000 +2026-06-19 19:06:23.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 19:06:23.421 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:06:23.553 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:06:23.654 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:07:29.743 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000075/group_7/index_6863_seed_12042_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 19:07:32.652 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 19:07:50.943 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:07:50.952 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:07:50.954 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:08:58.958 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000075/group_7/index_6863_seed_12043_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 19:09:04.135 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 19:09:22.966 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:09:24.209 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:09:24.347 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 19:09:25.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:11:18.397 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.14s (1.89min) +2026-06-19 19:11:19.100 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:13:16.252 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:20:16.314 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 76, train_steps: 3040 +2026-06-19 19:20:16.990 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 19:20:17.167 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:20:17.248 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:20:17.347 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:21:23.078 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000076/group_7/index_4184_seed_3111_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 19:21:27.022 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 19:21:44.966 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:21:44.972 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:21:44.972 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:22:53.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000076/group_7/index_4184_seed_3112_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 19:22:58.924 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 19:23:17.191 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:23:18.455 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:23:18.556 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 19:23:19.656 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:25:13.441 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.66s (1.89min) +2026-06-19 19:25:14.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:27:12.732 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:34:13.346 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 77, train_steps: 3080 +2026-06-19 19:34:13.950 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 19:34:14.135 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:34:14.218 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:34:14.344 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:35:20.189 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_7/index_8938_seed_15551_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 19:35:24.169 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 19:35:42.125 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:35:42.135 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:35:42.135 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:36:49.385 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_7/index_8938_seed_15552_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 19:36:52.663 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 19:37:11.314 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:37:13.074 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:37:13.190 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 19:37:13.964 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:39:08.734 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.60s (1.91min) +2026-06-19 19:39:09.441 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:41:09.768 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:48:12.712 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 78, train_steps: 3120 +2026-06-19 19:48:13.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 19:48:13.565 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:48:13.688 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:48:13.767 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:49:19.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_7/index_9727_seed_18006_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 19:49:25.103 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 19:49:43.474 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:49:43.483 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:49:43.483 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:50:49.854 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_7/index_9727_seed_18007_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 19:50:54.663 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 19:51:13.165 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:51:13.323 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:51:13.359 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 19:51:14.358 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:53:09.267 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-19 19:53:09.976 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:55:09.984 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:02:13.170 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 79, train_steps: 3160 +2026-06-19 20:02:13.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 20:02:14.024 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:02:14.060 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:02:14.157 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:03:20.103 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000079/group_7/index_3920_seed_15936_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 20:03:24.620 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 20:03:43.239 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:03:43.249 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:03:43.249 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:04:49.729 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000079/group_7/index_3920_seed_15937_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 20:04:53.752 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 20:05:12.030 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:05:13.093 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:05:13.232 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 20:05:14.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:07:09.660 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.53s (1.93min) +2026-06-19 20:07:10.385 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:09:11.031 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:16:13.962 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 80, train_steps: 3200 +2026-06-19 20:16:51.185 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=80 (missing=0, unexpected=0) +2026-06-19 20:16:53.092 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 20:16:53.267 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:16:53.353 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:16:53.479 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:18:01.851 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_7/index_2621_seed_10017_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 20:18:06.987 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 20:18:24.987 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:18:24.992 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:18:24.993 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:19:30.954 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_7/index_2621_seed_10018_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 20:19:35.079 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 20:19:53.671 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:19:54.752 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:19:54.896 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 20:19:55.786 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:21:49.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-19 20:21:50.363 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:23:47.271 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:30:47.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 81, train_steps: 3240 +2026-06-19 20:30:48.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 20:30:48.242 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:30:48.290 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:30:48.339 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:31:54.745 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_7/index_8865_seed_16004_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 20:31:58.412 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 20:32:16.157 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:32:16.166 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:32:16.166 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:33:25.071 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_7/index_8865_seed_16005_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 20:33:30.316 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 20:33:48.570 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:33:49.273 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:33:49.416 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 20:33:50.227 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:35:44.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.70s (1.89min) +2026-06-19 20:35:44.752 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:37:41.676 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:44:40.141 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 82, train_steps: 3280 +2026-06-19 20:44:40.696 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 20:44:40.870 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:44:40.960 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:44:41.109 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:45:48.042 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_7/index_940_seed_6155_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 20:45:52.068 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 20:46:10.062 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:46:10.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:46:10.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:47:20.030 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_7/index_940_seed_6156_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 20:47:23.120 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 20:47:41.915 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:47:44.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:47:44.269 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 20:47:45.137 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:49:38.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.41s (1.89min) +2026-06-19 20:49:39.275 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:51:38.430 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:58:38.987 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 83, train_steps: 3320 +2026-06-19 20:58:39.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 20:58:40.031 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:58:40.127 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:58:40.242 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:59:46.725 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_7/index_4977_seed_16398_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 20:59:50.964 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 21:00:09.007 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:00:09.017 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:00:09.017 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:01:19.675 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_7/index_4977_seed_16399_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 21:01:24.934 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 21:01:43.494 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:01:44.304 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:01:44.346 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 21:01:45.248 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:03:38.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.28s (1.89min) +2026-06-19 21:03:39.340 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:05:38.817 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:12:38.998 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 84, train_steps: 3360 +2026-06-19 21:12:39.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 21:12:39.811 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:12:39.881 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:12:40.019 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:13:47.008 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_7/index_6200_seed_4437_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 21:13:52.047 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 21:14:10.475 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:14:10.480 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:14:10.481 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:15:19.593 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_7/index_6200_seed_4438_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 21:15:23.564 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 21:15:42.016 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:15:43.791 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:15:43.918 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 21:15:44.695 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:17:38.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.31s (1.89min) +2026-06-19 21:17:38.863 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:19:38.522 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:26:39.497 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 85, train_steps: 3400 +2026-06-19 21:26:40.168 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 21:26:40.292 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:26:40.332 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:26:40.403 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:27:47.739 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_7/index_1316_seed_390_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 21:27:50.829 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 21:28:08.535 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:28:08.543 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:28:08.544 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:29:20.079 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_7/index_1316_seed_391_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 21:29:25.389 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 21:29:43.651 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:29:43.697 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:29:43.764 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 21:29:44.730 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:31:38.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-19 21:31:38.988 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:33:38.754 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:40:38.648 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 86, train_steps: 3440 +2026-06-19 21:40:39.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 21:40:39.467 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:40:39.513 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:40:39.566 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:41:46.745 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_7/index_2634_seed_11410_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 21:41:50.861 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 21:42:08.899 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:42:08.908 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:42:08.908 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:43:18.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_7/index_2634_seed_11411_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 21:43:21.832 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 21:43:40.787 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:43:42.867 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:43:42.975 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 21:43:43.753 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:45:37.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 21:45:37.987 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:47:37.707 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:54:39.044 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 87, train_steps: 3480 +2026-06-19 21:54:39.743 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 21:54:39.902 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:54:39.955 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:54:40.067 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:55:46.567 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000087/group_7/index_8609_seed_32343_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 21:55:49.578 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 21:56:07.406 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:56:07.414 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:56:07.415 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:57:19.075 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000087/group_7/index_8609_seed_32344_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 21:57:24.400 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 21:57:42.862 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:57:43.169 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:57:43.297 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 21:57:44.193 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:59:37.604 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.39s (1.89min) +2026-06-19 21:59:38.317 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:01:38.554 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:08:39.987 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 88, train_steps: 3520 +2026-06-19 22:08:40.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 22:08:40.859 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:08:40.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:08:41.061 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:09:47.922 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_7/index_4075_seed_22434_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 22:09:51.934 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 22:10:10.188 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:10:10.197 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:10:10.197 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:11:19.542 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_7/index_4075_seed_22435_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 22:11:22.705 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 22:11:41.094 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:11:43.755 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:11:43.872 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 22:11:44.713 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:13:38.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.57s (1.89min) +2026-06-19 22:13:39.104 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:15:38.303 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:22:38.712 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 89, train_steps: 3560 +2026-06-19 22:22:39.425 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 22:22:39.582 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:22:39.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:22:39.795 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:23:46.082 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_7/index_1033_seed_16461_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 22:23:51.352 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 22:24:09.966 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:24:09.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:24:09.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:25:19.851 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_7/index_1033_seed_16462_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 22:25:24.022 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 22:25:42.317 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:25:43.683 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:25:43.777 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 22:25:44.665 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:27:38.162 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 22:27:38.886 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:29:38.493 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:36:39.420 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 90, train_steps: 3600 +2026-06-19 22:37:15.538 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=90 (missing=0, unexpected=0) +2026-06-19 22:37:17.307 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 22:37:17.484 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:37:17.578 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:37:17.680 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:38:24.756 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_7/index_4178_seed_19896_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 22:38:29.971 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 22:38:48.474 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:38:48.483 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:38:48.484 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:39:55.076 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_7/index_4178_seed_19897_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 22:40:00.034 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 22:40:18.509 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:40:19.241 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:40:19.303 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 22:40:20.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:42:14.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.70s (1.89min) +2026-06-19 22:42:14.863 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:44:13.691 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:51:14.849 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 91, train_steps: 3640 +2026-06-19 22:51:15.699 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 22:51:15.839 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:51:15.886 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:51:15.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:52:21.883 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000091/group_7/index_6489_seed_33230_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 22:52:26.758 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 22:52:45.178 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:52:45.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:52:45.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:53:51.953 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000091/group_7/index_6489_seed_33231_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 22:53:56.881 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 22:54:15.465 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:54:15.937 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:54:16.089 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 22:54:16.966 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:56:12.672 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.67s (1.93min) +2026-06-19 22:56:13.389 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:58:13.111 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:05:14.261 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 92, train_steps: 3680 +2026-06-19 23:05:15.118 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 23:05:15.332 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:05:15.460 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:05:15.548 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:06:21.337 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000092/group_7/index_8838_seed_22171_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 23:06:25.450 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 23:06:43.772 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:06:43.782 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:06:43.782 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:07:51.279 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000092/group_7/index_8838_seed_22172_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 23:07:56.586 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 23:08:15.081 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:08:16.058 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:08:16.207 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 23:08:17.107 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:10:11.751 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.62s (1.91min) +2026-06-19 23:10:12.471 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:12:11.282 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:19:13.481 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 93, train_steps: 3720 +2026-06-19 23:19:14.152 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 23:19:14.368 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:19:14.488 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:19:14.600 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:20:20.580 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_7/index_3185_seed_29230_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 23:20:25.562 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 23:20:43.986 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:20:43.994 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:20:43.994 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:21:50.181 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_7/index_3185_seed_29231_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 23:21:55.458 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 23:22:13.899 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:22:14.354 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:22:14.381 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 23:22:15.549 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:24:10.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.33s (1.92min) +2026-06-19 23:24:11.707 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:26:11.283 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:33:13.934 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 94, train_steps: 3760 +2026-06-19 23:33:14.639 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 23:33:14.743 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:33:14.784 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:33:14.831 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:34:21.019 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_7/index_1739_seed_24511_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 23:34:25.912 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 23:34:44.732 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:34:44.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:34:44.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:35:51.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_7/index_1739_seed_24512_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 23:35:56.507 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 23:36:15.223 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:36:16.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:36:16.255 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 23:36:17.143 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:38:12.062 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.89s (1.91min) +2026-06-19 23:38:12.784 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:40:11.319 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:47:11.866 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 95, train_steps: 3800 +2026-06-19 23:47:12.590 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-19 23:47:12.756 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:47:12.842 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:47:12.941 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:48:18.415 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_7/index_6920_seed_20675_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-19 23:48:23.257 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 23:48:41.888 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:48:41.897 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:48:41.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:49:48.401 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_7/index_6920_seed_20676_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-19 23:49:52.502 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-19 23:50:10.714 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:50:12.004 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:50:12.114 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-19 23:50:13.149 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:52:08.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.82s (1.91min) +2026-06-19 23:52:08.791 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:54:07.675 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:01:09.778 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 96, train_steps: 3840 +2026-06-20 00:01:10.475 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 00:01:10.620 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:01:10.660 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:01:10.871 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:02:17.209 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_7/index_8960_seed_28160_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 00:02:20.332 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 00:02:39.842 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:02:39.851 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:02:39.852 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:03:47.473 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_7/index_8960_seed_28161_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 00:03:52.744 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 00:04:11.541 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:04:11.783 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:04:11.807 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 00:04:12.972 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:06:07.928 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 00:06:08.644 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:08:07.688 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:15:08.501 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 97, train_steps: 3880 +2026-06-20 00:15:09.218 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 00:15:09.347 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:15:09.385 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:15:09.457 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:16:15.541 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_7/index_8885_seed_16239_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 00:16:20.357 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 00:16:40.799 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:16:40.806 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:16:40.806 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:17:47.416 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_7/index_8885_seed_16240_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 00:17:51.352 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 00:18:09.616 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:18:11.238 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:18:11.329 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 00:18:12.259 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:20:07.671 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.20s (1.92min) +2026-06-20 00:20:08.401 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:22:07.125 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:29:07.007 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 98, train_steps: 3920 +2026-06-20 00:29:07.692 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 00:29:07.879 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:29:07.984 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:29:08.103 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:30:13.978 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000098/group_7/index_8053_seed_15816_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 00:30:17.080 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 00:30:35.224 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:30:35.233 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:30:35.234 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:31:46.461 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000098/group_7/index_8053_seed_15817_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 00:31:51.920 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 00:32:10.408 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:32:10.419 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:32:10.431 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 00:32:11.461 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:34:06.389 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 00:34:07.118 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:36:03.850 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:43:05.346 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 99, train_steps: 3960 +2026-06-20 00:43:06.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 00:43:06.238 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:43:06.362 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:43:06.444 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:44:12.267 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_7/index_8117_seed_12404_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 00:44:17.197 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 00:44:35.874 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:44:35.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:44:35.883 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:45:46.510 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_7/index_8117_seed_12405_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 00:45:51.424 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 00:46:09.418 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:46:09.899 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:46:10.060 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 00:46:10.939 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:48:05.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.75s (1.91min) +2026-06-20 00:48:06.440 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:50:03.397 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:57:04.795 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 100, train_steps: 4000 +2026-06-20 00:58:47.229 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=100 (missing=0, unexpected=0) +2026-06-20 00:58:49.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 00:58:49.791 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:58:49.925 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:58:50.004 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:59:57.438 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_7/index_4642_seed_217_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 01:00:01.480 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 01:00:19.687 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:00:19.696 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:00:19.697 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:01:26.880 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_7/index_4642_seed_218_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 01:01:30.132 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 01:01:48.187 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:01:50.303 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:01:50.452 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 01:01:51.242 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:03:44.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-20 01:03:45.434 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:05:42.218 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:12:39.860 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 101, train_steps: 4040 +2026-06-20 01:12:40.567 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 01:12:40.749 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:12:40.824 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:12:40.916 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:13:48.373 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_7/index_5650_seed_4642_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 01:13:52.454 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 01:14:11.203 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:14:11.209 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:14:11.210 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:15:20.119 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_7/index_5650_seed_4643_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 01:15:23.283 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 01:15:41.836 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:15:43.499 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:15:43.626 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 01:15:44.540 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:17:38.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.62s (1.89min) +2026-06-20 01:17:38.931 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:19:38.324 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:26:37.804 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 102, train_steps: 4080 +2026-06-20 01:26:38.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 01:26:38.711 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:26:38.891 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:26:38.936 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:27:44.565 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_7/index_5132_seed_6375_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 01:27:49.414 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 01:28:07.916 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:28:07.925 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:28:07.925 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:29:17.102 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_7/index_5132_seed_6376_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 01:29:20.996 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 01:29:40.851 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:29:41.183 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:29:41.266 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 01:29:42.417 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:31:36.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-20 01:31:37.097 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:33:37.223 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:40:38.539 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 103, train_steps: 4120 +2026-06-20 01:40:39.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 01:40:39.451 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:40:39.592 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:40:39.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:41:46.299 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000103/group_7/index_1878_seed_29539_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 01:41:51.624 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 01:42:10.452 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:42:10.457 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:42:10.458 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:43:19.148 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000103/group_7/index_1878_seed_29540_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 01:43:23.151 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 01:43:41.486 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:43:43.334 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:43:43.419 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 01:43:44.318 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:45:37.767 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-20 01:45:38.487 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:47:38.263 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:54:37.789 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 104, train_steps: 4160 +2026-06-20 01:54:38.704 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 01:54:38.910 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:54:39.049 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:54:39.113 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:55:45.501 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_7/index_3917_seed_4018_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 01:55:50.743 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 01:56:09.641 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:56:09.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:56:09.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:57:18.841 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_7/index_3917_seed_4019_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 01:57:23.683 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 01:57:41.921 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:57:42.889 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:57:43.035 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 01:57:43.917 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:59:37.314 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-20 01:59:38.031 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:01:38.018 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:08:38.447 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 105, train_steps: 4200 +2026-06-20 02:08:39.149 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 02:08:39.368 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:08:39.499 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:08:39.576 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:09:46.210 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000105/group_7/index_7876_seed_4752_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 02:09:51.101 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 02:10:09.703 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:10:09.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:10:09.713 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:11:18.947 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000105/group_7/index_7876_seed_4753_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 02:11:23.918 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 02:11:42.361 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:11:43.945 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:11:44.073 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 02:11:44.972 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:13:38.217 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 02:13:38.925 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:15:38.268 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:22:37.848 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 106, train_steps: 4240 +2026-06-20 02:22:38.564 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 02:22:38.721 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:22:38.788 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:22:38.914 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:23:44.836 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000106/group_7/index_1463_seed_10757_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 02:23:49.732 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 02:24:07.797 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:24:07.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:24:07.816 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:25:16.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000106/group_7/index_1463_seed_10758_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 02:25:20.815 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 02:25:39.532 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:25:40.847 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:25:40.859 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 02:25:41.801 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:27:35.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.90s (1.90min) +2026-06-20 02:27:36.583 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:29:36.812 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:36:37.098 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 107, train_steps: 4280 +2026-06-20 02:36:37.801 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 02:36:37.964 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:36:38.037 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:36:38.172 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:37:44.025 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_7/index_3508_seed_15128_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 02:37:49.305 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 02:38:07.457 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:38:07.466 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:38:07.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:39:15.311 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_7/index_3508_seed_15129_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 02:39:20.312 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 02:39:39.219 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:39:39.229 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:39:39.241 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 02:39:40.488 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:41:35.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 02:41:36.072 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:43:35.144 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:50:35.002 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 108, train_steps: 4320 +2026-06-20 02:50:35.727 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 02:50:35.914 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:50:36.010 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:50:36.092 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:51:42.418 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_7/index_2244_seed_9315_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 02:51:47.392 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 02:52:05.804 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:52:05.816 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:52:05.816 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:53:13.480 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_7/index_2244_seed_9316_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 02:53:17.892 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 02:53:36.214 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:53:37.574 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:53:37.693 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 02:53:38.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:55:33.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.45s (1.91min) +2026-06-20 02:55:33.972 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:57:30.646 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:04:30.706 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 109, train_steps: 4360 +2026-06-20 03:04:31.405 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 03:04:31.566 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:04:31.619 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:04:31.744 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:05:37.572 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000109/group_7/index_8694_seed_5928_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 03:05:43.056 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 03:06:01.542 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:06:01.552 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:06:01.552 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:07:09.484 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000109/group_7/index_8694_seed_5929_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 03:07:14.299 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 03:07:32.702 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:07:33.407 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:07:33.491 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 03:07:34.425 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:09:29.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-20 03:09:30.015 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:11:26.575 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:18:25.328 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 110, train_steps: 4400 +2026-06-20 03:19:01.515 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=110 (missing=0, unexpected=0) +2026-06-20 03:19:03.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 03:19:03.478 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:19:03.585 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:19:03.664 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:20:11.521 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_7/index_1246_seed_5329_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 03:20:15.497 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 03:20:32.992 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:20:33.001 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:20:33.001 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:21:43.368 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_7/index_1246_seed_5330_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 03:21:46.546 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 03:22:03.745 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:22:07.524 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:22:07.653 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 03:22:08.418 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:24:03.966 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.41s (1.92min) +2026-06-20 03:24:04.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:26:02.472 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:33:05.393 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 111, train_steps: 4440 +2026-06-20 03:33:06.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 03:33:06.262 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:33:06.357 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:33:06.456 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:34:12.274 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_7/index_8494_seed_29814_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 03:34:15.455 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 03:34:32.204 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:34:32.213 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:34:32.213 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:35:46.473 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_7/index_8494_seed_29815_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 03:35:49.649 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 03:36:06.630 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:36:10.098 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:36:10.111 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 03:36:11.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:38:06.021 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 03:38:06.732 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:40:04.481 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:47:04.318 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 112, train_steps: 4480 +2026-06-20 03:47:05.043 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 03:47:05.161 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:47:05.195 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:47:05.260 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:48:11.078 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_7/index_6237_seed_3221_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 03:48:14.251 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 03:48:31.225 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:48:31.235 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:48:31.235 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:49:40.959 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_7/index_6237_seed_3222_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 03:49:44.191 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 03:50:01.229 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:50:04.983 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:50:05.064 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 03:50:06.171 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:52:00.943 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-20 03:52:01.658 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:54:00.963 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:01:04.008 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 113, train_steps: 4520 +2026-06-20 04:01:04.874 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 04:01:05.009 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:01:05.055 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:01:05.139 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:02:11.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000113/group_7/index_2104_seed_11878_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 04:02:16.134 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 04:02:33.309 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:02:33.318 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:02:33.318 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:03:42.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000113/group_7/index_2104_seed_11879_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 04:03:46.668 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 04:04:03.863 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:04:06.611 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:04:06.754 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 04:04:07.636 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:06:02.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.15s (1.92min) +2026-06-20 04:06:03.532 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:08:01.586 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:15:03.778 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 114, train_steps: 4560 +2026-06-20 04:15:04.473 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 04:15:04.619 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:15:04.671 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:15:04.770 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:16:10.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000114/group_7/index_1742_seed_29550_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 04:16:14.418 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 04:16:31.535 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:16:31.546 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:16:31.546 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:17:40.419 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000114/group_7/index_1742_seed_29551_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 04:17:43.499 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 04:18:00.846 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:18:04.152 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:18:04.180 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 04:18:05.224 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:20:00.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.72s (1.91min) +2026-06-20 04:20:00.895 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:22:00.099 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:29:01.473 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 115, train_steps: 4600 +2026-06-20 04:29:02.193 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 04:29:02.402 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:29:02.520 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:29:02.599 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:30:08.296 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_7/index_4321_seed_30348_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 04:30:11.348 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 04:30:28.131 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:30:28.142 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:30:28.142 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:31:39.025 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_7/index_4321_seed_30349_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 04:31:42.056 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 04:31:59.183 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:32:02.494 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:32:02.506 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 04:32:03.402 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:33:58.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.33s (1.91min) +2026-06-20 04:33:58.775 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:35:55.625 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:42:56.473 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 116, train_steps: 4640 +2026-06-20 04:42:57.159 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 04:42:57.302 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:42:57.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:42:57.449 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:44:03.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_7/index_530_seed_9290_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 04:44:06.671 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 04:44:23.433 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:44:23.442 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:44:23.443 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:45:34.202 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_7/index_530_seed_9291_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 04:45:39.968 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 04:45:57.980 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:45:58.381 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:45:58.392 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 04:45:59.416 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:47:53.698 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.25s (1.90min) +2026-06-20 04:47:54.407 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:49:50.888 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:56:50.226 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 117, train_steps: 4680 +2026-06-20 04:56:50.953 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 04:56:51.135 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:56:51.220 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:56:51.296 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:57:57.109 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_7/index_1458_seed_1167_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 04:58:01.712 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 04:58:19.311 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:58:19.321 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:58:19.321 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:59:28.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_7/index_1458_seed_1168_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 04:59:32.706 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 04:59:50.655 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:59:51.798 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:59:51.851 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 04:59:52.811 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:01:46.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.49s (1.89min) +2026-06-20 05:01:47.078 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:03:44.145 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:10:42.586 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 118, train_steps: 4720 +2026-06-20 05:10:43.296 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 05:10:43.459 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:10:43.532 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:10:43.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:11:49.617 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000118/group_7/index_7576_seed_18065_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 05:11:53.284 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 05:12:10.667 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:12:10.677 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:12:10.677 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:13:22.945 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000118/group_7/index_7576_seed_18066_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 05:13:28.070 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 05:13:45.846 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:13:46.957 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:13:47.104 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 05:13:47.992 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:15:41.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.26s (1.89min) +2026-06-20 05:15:42.001 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:17:38.993 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:24:37.906 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 119, train_steps: 4760 +2026-06-20 05:24:38.623 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 05:24:38.793 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:24:38.871 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:24:38.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:25:44.832 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_7/index_3062_seed_19644_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 05:25:49.360 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 05:26:07.259 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:26:07.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:26:07.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:27:14.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_7/index_3062_seed_19645_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 05:27:19.657 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 05:27:37.598 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:27:40.119 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:27:40.154 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 05:27:41.302 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:29:34.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-20 05:29:35.559 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:31:33.104 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:38:32.154 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 120, train_steps: 4800 +2026-06-20 05:39:09.252 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=120 (missing=0, unexpected=0) +2026-06-20 05:39:10.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 05:39:11.088 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:39:11.190 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:39:11.288 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:40:19.046 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000120/group_7/index_7070_seed_1839_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 05:40:23.031 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 05:40:40.824 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:40:40.834 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:40:40.836 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:41:48.625 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000120/group_7/index_7070_seed_1840_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 05:41:53.683 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 05:42:11.705 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:42:12.566 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:42:12.578 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 05:42:13.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:44:08.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.96s (1.92min) +2026-06-20 05:44:09.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:46:10.085 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:53:12.059 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 121, train_steps: 4840 +2026-06-20 05:53:12.775 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 05:53:12.914 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:53:12.950 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:53:13.037 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:54:19.359 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_7/index_6007_seed_32259_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 05:54:24.044 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 05:54:41.625 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:54:41.635 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:54:41.635 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:55:48.795 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_7/index_6007_seed_32260_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 05:55:53.906 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 05:56:12.082 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:56:13.156 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:56:13.174 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 05:56:14.191 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:58:09.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.97s (1.92min) +2026-06-20 05:58:10.076 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:00:10.574 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:07:11.953 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 122, train_steps: 4880 +2026-06-20 06:07:12.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 06:07:12.752 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:07:12.792 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:07:12.824 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:08:18.771 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_7/index_6530_seed_20319_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 06:08:22.764 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 06:08:40.868 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:08:40.874 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:08:40.875 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:09:48.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_7/index_6530_seed_20320_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 06:09:51.977 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 06:10:09.515 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:10:13.195 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:10:13.322 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 06:10:14.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:12:08.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.25s (1.90min) +2026-06-20 06:12:09.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:14:10.578 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:21:11.867 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 123, train_steps: 4920 +2026-06-20 06:21:12.561 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 06:21:12.744 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:21:12.841 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:21:12.920 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:22:18.756 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_7/index_8555_seed_33389_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 06:22:21.802 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 06:22:40.806 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:22:40.813 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:22:40.814 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:23:48.326 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_7/index_8555_seed_33390_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 06:23:51.563 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 06:24:09.054 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:24:12.071 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:24:12.124 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 06:24:13.171 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:26:08.788 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.47s (1.92min) +2026-06-20 06:26:09.511 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:28:09.832 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:35:11.849 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 124, train_steps: 4960 +2026-06-20 06:35:12.541 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 06:35:12.734 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:35:12.826 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:35:12.907 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:36:18.594 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_7/index_7531_seed_30708_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 06:36:23.548 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 06:36:41.266 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:36:41.275 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:36:41.275 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:37:48.250 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_7/index_7531_seed_30709_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 06:37:52.095 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 06:38:09.403 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:38:12.129 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:38:12.141 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 06:38:13.153 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:40:08.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.44s (1.92min) +2026-06-20 06:40:09.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:42:09.771 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:49:11.649 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 125, train_steps: 5000 +2026-06-20 06:50:15.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 06:50:15.347 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:50:15.393 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:50:15.501 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:51:21.490 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_7/index_1262_seed_33350_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 06:51:26.788 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 06:51:44.458 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:51:44.464 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:51:44.464 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:52:51.110 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_7/index_1262_seed_33351_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 06:52:54.913 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 06:53:12.387 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:53:15.926 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:53:16.024 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 06:53:16.849 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:55:11.654 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.68s (1.91min) +2026-06-20 06:55:12.372 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:57:11.762 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:04:11.922 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 126, train_steps: 5040 +2026-06-20 07:04:12.630 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 07:04:12.738 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:04:12.770 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:04:12.798 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:05:19.206 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_7/index_3243_seed_10880_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 07:05:24.241 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 07:05:41.787 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:05:41.792 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:05:41.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:06:49.452 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_7/index_3243_seed_10881_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 07:06:53.279 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 07:07:10.725 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:07:13.988 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:07:14.119 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 07:07:15.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:09:10.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.08s (1.92min) +2026-06-20 07:09:11.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:11:11.165 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:18:11.927 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 127, train_steps: 5080 +2026-06-20 07:18:12.629 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 07:18:12.737 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:18:12.777 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:18:12.823 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:19:18.851 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_7/index_2423_seed_21673_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 07:19:23.924 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 07:19:41.655 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:19:41.664 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:19:41.664 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:20:48.779 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_7/index_2423_seed_21674_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 07:20:52.688 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 07:21:10.586 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:21:13.050 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:21:13.116 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 07:21:14.015 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:23:09.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.50s (1.93min) +2026-06-20 07:23:10.327 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:25:10.157 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:32:11.179 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 128, train_steps: 5120 +2026-06-20 07:32:11.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 07:32:12.036 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:32:12.093 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:32:12.196 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:33:18.074 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_7/index_9709_seed_25087_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 07:33:23.275 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 07:33:40.917 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:33:40.923 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:33:40.924 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:34:48.784 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_7/index_9709_seed_25088_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 07:34:52.791 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 07:35:10.391 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:35:13.016 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:35:13.158 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 07:35:13.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:37:08.641 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.56s (1.91min) +2026-06-20 07:37:09.373 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:39:09.930 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:46:11.986 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 129, train_steps: 5160 +2026-06-20 07:46:12.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 07:46:12.894 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:46:13.006 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:46:13.105 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:47:19.402 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_7/index_3016_seed_11054_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 07:47:24.483 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 07:47:42.202 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:47:42.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:47:42.212 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:48:48.865 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_7/index_3016_seed_11055_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 07:48:53.598 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 07:49:11.576 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:49:13.070 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:49:13.182 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 07:49:14.082 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:51:09.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-20 07:51:09.758 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:53:10.302 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:00:11.786 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 130, train_steps: 5200 +2026-06-20 08:00:49.562 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=130 (missing=0, unexpected=0) +2026-06-20 08:00:51.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 08:00:51.415 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:00:51.517 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:00:51.615 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:02:00.305 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_7/index_6050_seed_883_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 08:02:05.004 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 08:02:23.344 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:02:23.354 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:02:23.354 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:03:28.874 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_7/index_6050_seed_884_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 08:03:32.742 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 08:03:50.307 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:03:52.509 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:03:52.521 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 08:03:53.456 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:05:47.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.72s (1.90min) +2026-06-20 08:05:48.018 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:07:44.843 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:14:45.315 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 131, train_steps: 5240 +2026-06-20 08:14:46.026 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 08:14:46.196 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:14:46.273 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:14:46.389 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:15:52.533 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_7/index_7621_seed_16736_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 08:15:57.745 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 08:16:15.316 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:16:15.325 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:16:15.325 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:17:23.319 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_7/index_7621_seed_16737_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 08:17:28.114 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 08:17:45.889 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:17:46.990 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:17:47.040 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 08:17:48.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:19:41.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.38s (1.89min) +2026-06-20 08:19:42.139 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:21:39.465 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:28:37.824 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 132, train_steps: 5280 +2026-06-20 08:28:38.518 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 08:28:38.614 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:28:38.657 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:28:38.704 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:29:45.021 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000132/group_7/index_6699_seed_26356_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 08:29:49.051 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 08:30:06.454 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:30:06.465 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:30:06.465 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:31:15.576 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000132/group_7/index_6699_seed_26357_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 08:31:20.713 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 08:31:40.465 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:31:40.849 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:31:40.861 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 08:31:41.888 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:33:35.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.44s (1.89min) +2026-06-20 08:33:36.119 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:35:35.261 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:42:34.338 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 133, train_steps: 5320 +2026-06-20 08:42:35.020 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 08:42:35.234 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:42:35.347 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:42:35.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:43:41.903 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_7/index_9380_seed_19653_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 08:43:46.689 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 08:44:04.275 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:44:04.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:44:04.285 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:45:11.977 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_7/index_9380_seed_19654_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 08:45:15.960 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 08:45:33.200 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:45:35.888 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:45:36.012 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 08:45:36.841 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:47:32.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.77s (1.93min) +2026-06-20 08:47:33.450 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:49:30.691 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:56:29.436 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 134, train_steps: 5360 +2026-06-20 08:56:30.116 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 08:56:30.292 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:56:30.371 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:56:30.473 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:57:36.565 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000134/group_7/index_6809_seed_15950_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 08:57:40.069 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 08:57:57.969 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:57:57.978 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:57:57.980 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:59:07.641 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000134/group_7/index_6809_seed_15951_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 08:59:10.887 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 08:59:28.286 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:59:32.041 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:59:32.176 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 08:59:32.938 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:01:26.319 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 09:01:27.048 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:03:23.804 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:10:22.986 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 135, train_steps: 5400 +2026-06-20 09:10:23.677 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 09:10:23.876 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:10:23.982 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:10:24.077 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:11:30.131 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000135/group_7/index_1261_seed_33364_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 09:11:35.224 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 09:11:53.105 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:11:53.116 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:11:53.116 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:13:00.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000135/group_7/index_1261_seed_33365_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 09:13:05.198 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 09:13:23.123 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:13:23.932 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:13:24.055 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 09:13:24.855 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:15:18.588 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.61s (1.89min) +2026-06-20 09:15:19.310 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:17:17.098 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:24:16.804 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 136, train_steps: 5440 +2026-06-20 09:24:17.534 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 09:24:17.759 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:24:17.872 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:24:17.951 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:25:23.738 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_7/index_8329_seed_24158_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 09:25:28.544 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 09:25:46.272 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:25:46.282 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:25:46.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:26:54.193 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_7/index_8329_seed_24159_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 09:26:58.956 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 09:27:16.918 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:27:18.656 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:27:18.696 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 09:27:19.694 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:29:12.959 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.15s (1.89min) +2026-06-20 09:29:13.686 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:31:13.176 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:38:13.218 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 137, train_steps: 5480 +2026-06-20 09:38:13.938 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 09:38:14.146 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:38:14.241 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:38:14.336 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:39:20.089 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000137/group_7/index_861_seed_19486_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 09:39:23.950 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 09:39:41.554 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:39:41.563 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:39:41.564 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:40:50.517 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000137/group_7/index_861_seed_19487_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 09:40:53.722 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 09:41:11.359 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:41:14.371 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:41:14.457 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 09:41:15.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:43:09.522 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-20 09:43:10.251 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:45:10.502 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:52:12.149 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 138, train_steps: 5520 +2026-06-20 09:52:12.869 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 09:52:13.086 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:52:13.171 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:52:13.185 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:53:19.392 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_7/index_604_seed_3408_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 09:53:24.552 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 09:53:42.223 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:53:42.235 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:53:42.236 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:54:49.425 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_7/index_604_seed_3409_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 09:54:53.283 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 09:55:11.001 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:55:13.383 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:55:13.396 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 09:55:14.547 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:57:10.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.46s (1.92min) +2026-06-20 09:57:10.921 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:59:11.151 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:06:12.670 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 139, train_steps: 5560 +2026-06-20 10:06:13.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 10:06:13.579 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:06:13.688 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:06:13.767 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:07:19.738 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_7/index_4896_seed_32999_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 10:07:24.961 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 10:07:42.501 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:07:42.510 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:07:42.511 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:08:50.009 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_7/index_4896_seed_33000_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 10:08:54.724 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 10:09:12.411 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:09:13.264 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:09:13.323 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 10:09:14.282 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:11:08.675 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.20s (1.90min) +2026-06-20 10:11:09.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:13:09.780 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:20:12.679 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 140, train_steps: 5600 +2026-06-20 10:20:49.792 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=140 (missing=0, unexpected=0) +2026-06-20 10:20:51.813 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 10:20:51.960 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:20:52.013 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:20:52.116 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:21:59.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000140/group_7/index_3735_seed_24710_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 10:22:03.578 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 10:22:20.475 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:22:20.484 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:22:20.485 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:23:29.074 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000140/group_7/index_3735_seed_24711_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 10:23:34.313 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 10:23:52.105 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:23:53.040 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:23:53.112 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 10:23:54.062 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:25:47.600 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-20 10:25:48.313 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:27:45.362 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:34:44.003 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 141, train_steps: 5640 +2026-06-20 10:34:44.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 10:34:44.801 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:34:44.845 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:34:44.911 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:35:50.785 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_7/index_6734_seed_5080_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 10:35:54.755 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 10:36:12.062 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:36:12.071 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:36:12.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:37:22.971 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_7/index_6734_seed_5081_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 10:37:26.232 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 10:37:43.533 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:37:46.846 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:37:46.974 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 10:37:47.765 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:39:41.415 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.51s (1.89min) +2026-06-20 10:39:42.148 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:41:39.463 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:48:38.462 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 142, train_steps: 5680 +2026-06-20 10:48:39.159 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 10:48:39.352 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:48:39.444 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:48:39.524 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:49:46.305 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000142/group_7/index_3542_seed_24095_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 10:49:49.479 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 10:50:07.643 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:50:07.653 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:50:07.654 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:51:19.275 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000142/group_7/index_3542_seed_24096_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 10:51:23.320 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 10:51:40.836 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:51:43.153 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:51:43.199 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 10:51:44.050 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:53:37.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-20 10:53:37.851 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:55:37.992 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:02:37.921 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 143, train_steps: 5720 +2026-06-20 11:02:38.633 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 11:02:38.813 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:02:38.904 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:02:38.983 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:03:45.076 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000143/group_7/index_6394_seed_10720_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 11:03:48.276 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 11:04:05.530 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:04:05.537 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:04:05.537 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:05:18.241 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000143/group_7/index_6394_seed_10721_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 11:05:21.456 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 11:05:40.442 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:05:42.169 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:05:42.300 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 11:05:43.222 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:07:36.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 11:07:37.313 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:09:36.842 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:16:37.102 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 144, train_steps: 5760 +2026-06-20 11:16:37.800 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 11:16:38.001 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:16:38.102 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:16:38.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:17:44.057 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_7/index_6114_seed_11110_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 11:17:49.253 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 11:18:06.760 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:18:06.768 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:18:06.769 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:19:13.342 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_7/index_6114_seed_11111_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 11:19:17.302 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 11:19:34.961 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:19:38.707 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:19:38.862 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 11:19:39.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:21:33.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.47s (1.89min) +2026-06-20 11:21:34.087 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:23:30.883 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:30:30.397 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 145, train_steps: 5800 +2026-06-20 11:30:31.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 11:30:31.299 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:30:31.404 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:30:31.484 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:31:37.298 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000145/group_7/index_3975_seed_5501_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 11:31:40.359 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 11:31:58.234 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:31:58.244 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:31:58.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:33:08.986 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000145/group_7/index_3975_seed_5502_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 11:33:12.872 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 11:33:30.720 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:33:33.207 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:33:33.332 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 11:33:34.238 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:35:28.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.89s (1.90min) +2026-06-20 11:35:28.900 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:37:25.685 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:44:23.916 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 146, train_steps: 5840 +2026-06-20 11:44:24.629 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 11:44:24.796 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:44:24.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:44:24.978 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:45:30.924 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_7/index_2947_seed_2985_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 11:45:35.687 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 11:45:53.278 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:45:53.289 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:45:53.290 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:46:59.508 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_7/index_2947_seed_2986_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 11:47:04.561 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 11:47:22.095 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:47:23.247 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:47:23.315 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 11:47:24.273 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:49:19.133 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-20 11:49:19.856 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:51:17.039 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:58:15.923 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 147, train_steps: 5880 +2026-06-20 11:58:16.619 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 11:58:16.831 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:58:16.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:58:17.034 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:59:23.174 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_7/index_5007_seed_10633_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 11:59:28.015 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 11:59:45.861 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:59:45.872 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:59:45.872 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:00:52.664 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_7/index_5007_seed_10634_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 12:00:56.559 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 12:01:14.104 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:01:16.232 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:01:16.363 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 12:01:17.214 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:03:12.586 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.29s (1.92min) +2026-06-20 12:03:13.294 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:05:12.904 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:12:12.700 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 148, train_steps: 5920 +2026-06-20 12:12:13.407 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 12:12:13.608 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:12:13.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:12:13.785 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:13:19.680 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000148/group_7/index_1344_seed_2328_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 12:13:22.565 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 12:13:40.400 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:13:40.410 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:13:40.411 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:14:49.460 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000148/group_7/index_1344_seed_2329_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 12:14:54.378 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 12:15:12.013 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:15:13.144 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:15:13.215 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 12:15:14.176 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:17:09.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.69s (1.91min) +2026-06-20 12:17:09.789 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:19:10.586 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:26:11.393 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 149, train_steps: 5960 +2026-06-20 12:26:12.136 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 12:26:12.316 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:26:12.407 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:26:12.488 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:27:18.407 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000149/group_7/index_9979_seed_534_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 12:27:22.408 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 12:27:40.842 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:27:40.851 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:27:40.852 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:28:48.701 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000149/group_7/index_9979_seed_535_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 12:28:53.617 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 12:29:11.407 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:29:13.029 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:29:13.119 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 12:29:14.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:31:08.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.72s (1.91min) +2026-06-20 12:31:09.533 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:33:09.829 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:40:10.990 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 150, train_steps: 6000 +2026-06-20 12:41:53.913 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=150 (missing=0, unexpected=0) +2026-06-20 12:41:56.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 12:41:56.165 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:41:56.221 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:41:56.270 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:43:04.019 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_7/index_9935_seed_20263_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 12:43:08.079 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 12:43:25.007 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:43:25.017 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:43:25.017 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:44:32.597 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_7/index_9935_seed_20264_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 12:44:37.629 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 12:44:57.731 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:44:59.744 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:44:59.803 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 12:45:00.805 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:46:55.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.70s (1.91min) +2026-06-20 12:46:56.236 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:48:52.988 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:55:51.598 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 151, train_steps: 6040 +2026-06-20 12:55:52.307 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 12:55:52.479 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:55:52.559 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:55:52.671 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:56:58.515 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_7/index_6917_seed_31836_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 12:57:02.590 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 12:57:20.043 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:57:20.054 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:57:20.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:58:27.202 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_7/index_6917_seed_31837_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 12:58:32.145 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 12:58:49.881 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:58:50.721 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:58:50.741 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 12:58:51.755 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:00:46.190 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.41s (1.91min) +2026-06-20 13:00:46.908 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:02:43.651 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:09:41.903 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 152, train_steps: 6080 +2026-06-20 13:09:42.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 13:09:42.789 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:09:42.887 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:09:42.983 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:10:49.097 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_7/index_6286_seed_11493_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 13:10:53.839 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 13:11:11.335 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:11:11.345 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:11:11.346 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:12:20.373 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_7/index_6286_seed_11494_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 13:12:24.264 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 13:12:41.750 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:12:43.910 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:12:43.932 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 13:12:44.827 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:14:38.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-20 13:14:39.492 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:16:38.549 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:23:37.225 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 153, train_steps: 6120 +2026-06-20 13:23:38.091 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 13:23:38.245 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:23:38.298 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:23:38.419 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:24:44.998 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_7/index_3682_seed_32177_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 13:24:48.154 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 13:25:05.412 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:25:05.422 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:25:05.423 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:26:15.100 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_7/index_3682_seed_32178_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 13:26:18.147 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 13:26:35.575 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:26:40.127 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:26:40.217 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 13:26:41.209 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:28:34.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.39s (1.89min) +2026-06-20 13:28:35.473 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:30:32.907 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:37:31.696 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 154, train_steps: 6160 +2026-06-20 13:37:32.392 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 13:37:32.567 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:37:32.645 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:37:32.742 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:38:38.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_7/index_6840_seed_16366_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 13:38:43.845 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 13:39:01.711 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:39:01.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:39:01.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:40:09.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_7/index_6840_seed_16367_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 13:40:14.210 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 13:40:32.235 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:40:33.788 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:40:33.895 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 13:40:34.811 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:42:29.329 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.47s (1.91min) +2026-06-20 13:42:30.062 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:44:27.159 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:51:24.970 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 155, train_steps: 6200 +2026-06-20 13:51:25.701 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 13:51:25.868 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:51:25.946 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:51:26.064 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:52:31.813 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_7/index_5812_seed_2512_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 13:52:35.641 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 13:52:52.852 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:52:52.862 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:52:52.862 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:54:01.140 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_7/index_5812_seed_2513_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 13:54:04.355 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 13:54:21.960 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:54:24.447 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:54:24.547 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 13:54:25.418 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:56:19.129 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.63s (1.89min) +2026-06-20 13:56:19.859 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:58:16.748 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:05:15.412 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 156, train_steps: 6240 +2026-06-20 14:05:16.040 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 14:05:16.267 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:05:16.407 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:05:16.488 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:06:22.515 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_7/index_7835_seed_17672_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 14:06:25.594 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 14:06:43.009 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:06:43.016 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:06:43.016 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:07:51.605 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_7/index_7835_seed_17673_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 14:07:54.702 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 14:08:12.286 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:08:15.931 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:08:16.067 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 14:08:16.832 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:10:11.722 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.74s (1.91min) +2026-06-20 14:10:12.457 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:12:11.102 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:19:12.239 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 157, train_steps: 6280 +2026-06-20 14:19:12.925 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 14:19:13.070 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:19:13.131 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:19:13.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:20:19.104 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000157/group_7/index_4500_seed_10808_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 14:20:24.127 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 14:20:41.833 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:20:41.843 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:20:41.843 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:21:48.118 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000157/group_7/index_4500_seed_10809_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 14:21:52.956 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 14:22:10.695 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:22:12.398 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:22:12.499 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 14:22:13.544 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:24:08.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.22s (1.92min) +2026-06-20 14:24:09.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:26:09.905 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:33:10.828 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 158, train_steps: 6320 +2026-06-20 14:33:11.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 14:33:11.719 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:33:11.820 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:33:11.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:34:17.564 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_7/index_9167_seed_20353_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 14:34:22.307 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 14:34:40.789 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:34:40.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:34:40.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:35:47.377 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_7/index_9167_seed_20354_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 14:35:52.125 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 14:36:09.989 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:36:10.935 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:36:11.032 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 14:36:12.053 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:38:07.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.88s (1.91min) +2026-06-20 14:38:07.784 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:40:05.257 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:47:04.563 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 159, train_steps: 6360 +2026-06-20 14:47:05.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 14:47:05.485 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:47:05.605 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:47:05.702 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:48:11.339 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_7/index_977_seed_186_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 14:48:16.165 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 14:48:33.907 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:48:33.915 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:48:33.915 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:49:40.142 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_7/index_977_seed_187_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 14:49:44.973 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 14:50:03.008 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:50:04.049 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:50:04.199 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 14:50:05.249 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:51:59.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.86s (1.90min) +2026-06-20 14:51:59.929 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:53:57.323 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:00:59.177 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 160, train_steps: 6400 +2026-06-20 15:01:36.921 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=160 (missing=0, unexpected=0) +2026-06-20 15:01:39.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 15:01:39.453 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:01:39.526 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:01:39.675 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:02:48.039 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_7/index_8595_seed_20937_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 15:02:52.836 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 15:03:10.569 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:03:10.578 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:03:10.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:04:19.100 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_7/index_8595_seed_20938_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 15:04:23.988 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 15:04:41.999 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:04:42.879 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:04:42.940 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 15:04:43.902 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:06:37.409 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.48s (1.89min) +2026-06-20 15:06:38.136 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:08:38.301 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:15:37.522 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 161, train_steps: 6440 +2026-06-20 15:15:38.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 15:15:38.422 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:15:38.547 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:15:38.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:16:44.662 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_7/index_3558_seed_20761_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 15:16:49.522 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 15:17:07.327 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:17:07.337 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:17:07.337 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:18:16.689 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_7/index_3558_seed_20762_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 15:18:21.587 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 15:18:40.335 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:18:40.375 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:18:40.518 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 15:18:41.586 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:20:35.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-20 15:20:36.332 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:22:36.155 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:29:36.899 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 162, train_steps: 6480 +2026-06-20 15:29:37.568 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 15:29:37.699 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:29:37.733 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:29:37.813 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:30:43.807 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_7/index_5912_seed_4947_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 15:30:48.610 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 15:31:06.118 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:31:06.127 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:31:06.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:32:11.956 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_7/index_5912_seed_4948_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 15:32:16.790 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 15:32:34.585 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:32:37.444 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:32:37.580 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 15:32:38.620 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:34:33.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.48s (1.91min) +2026-06-20 15:34:33.827 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:36:31.315 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:43:31.482 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 163, train_steps: 6520 +2026-06-20 15:43:32.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 15:43:32.392 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:43:32.521 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:43:32.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:44:38.640 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000163/group_7/index_9140_seed_25099_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 15:44:42.545 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 15:45:00.071 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:45:00.080 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:45:00.081 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:46:09.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000163/group_7/index_9140_seed_25100_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 15:46:12.455 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 15:46:29.979 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:46:33.676 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:46:33.803 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 15:46:34.580 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:48:29.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.47s (1.91min) +2026-06-20 15:48:29.934 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:50:26.537 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:57:25.506 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 164, train_steps: 6560 +2026-06-20 15:57:26.229 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 15:57:26.390 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:57:26.448 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:57:26.557 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:58:32.403 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_7/index_4340_seed_28843_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 15:58:37.516 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 15:58:56.317 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:58:56.325 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:58:56.325 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:00:02.635 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_7/index_4340_seed_28844_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 16:00:06.481 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 16:00:24.163 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:00:26.402 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:00:26.535 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 16:00:27.325 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:02:21.064 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.61s (1.89min) +2026-06-20 16:02:21.793 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:04:18.346 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:11:18.370 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 165, train_steps: 6600 +2026-06-20 16:11:19.087 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 16:11:19.255 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:11:19.328 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:11:19.438 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:12:25.662 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_7/index_7523_seed_26468_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 16:12:30.761 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 16:12:48.634 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:12:48.643 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:12:48.643 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:13:54.690 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_7/index_7523_seed_26469_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 16:13:59.600 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 16:14:17.886 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:14:19.106 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:14:19.216 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 16:14:20.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:16:14.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.75s (1.90min) +2026-06-20 16:16:14.870 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:18:14.124 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:25:13.929 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 166, train_steps: 6640 +2026-06-20 16:25:14.621 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 16:25:14.835 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:25:14.961 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:25:15.061 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:26:21.286 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_7/index_3807_seed_14229_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 16:26:26.121 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 16:26:43.697 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:26:43.707 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:26:43.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:27:50.278 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_7/index_3807_seed_14230_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 16:27:53.417 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 16:28:10.742 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:28:13.831 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:28:13.976 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 16:28:14.897 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:30:08.844 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.78s (1.90min) +2026-06-20 16:30:09.574 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:32:10.305 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:39:12.151 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 167, train_steps: 6680 +2026-06-20 16:39:12.849 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 16:39:13.030 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:39:13.120 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:39:13.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:40:19.337 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_7/index_3069_seed_10994_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 16:40:22.456 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 16:40:41.116 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:40:41.130 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:40:41.135 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:41:48.425 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_7/index_3069_seed_10995_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 16:41:53.625 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 16:42:11.612 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:42:12.095 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:42:12.206 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 16:42:13.364 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:44:08.745 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.36s (1.92min) +2026-06-20 16:44:09.477 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:46:10.018 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:53:11.376 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 168, train_steps: 6720 +2026-06-20 16:53:12.088 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 16:53:12.179 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:53:12.232 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:53:12.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:54:18.528 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_7/index_35_seed_19315_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 16:54:23.107 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 16:54:40.872 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:54:40.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:54:40.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:55:47.517 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_7/index_35_seed_19316_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 16:55:52.367 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 16:56:10.475 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:56:11.767 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:56:11.865 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 16:56:12.810 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:58:07.736 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-20 16:58:08.464 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:00:07.578 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:07:09.177 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 169, train_steps: 6760 +2026-06-20 17:07:09.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 17:07:10.101 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 17:07:10.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:07:10.364 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:08:16.397 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_7/index_5483_seed_8178_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 17:08:20.487 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 17:08:37.613 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:08:37.623 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:08:37.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:09:47.956 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_7/index_5483_seed_8179_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 17:09:51.083 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 17:10:08.517 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:10:11.904 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 17:10:12.000 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 17:10:12.863 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:12:08.255 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.20s (1.92min) +2026-06-20 17:12:08.969 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:14:09.318 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:21:11.119 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 170, train_steps: 6800 +2026-06-20 17:21:47.249 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=170 (missing=0, unexpected=0) +2026-06-20 17:21:49.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 17:21:49.356 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 17:21:49.389 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:21:49.479 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:22:56.949 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_7/index_8789_seed_4146_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 17:23:02.060 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 17:23:19.526 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:23:19.535 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:23:19.535 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:24:25.736 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_7/index_8789_seed_4147_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 17:24:30.525 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 17:24:48.256 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:24:49.925 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 17:24:50.004 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 17:24:50.963 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:26:44.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.05s (1.88min) +2026-06-20 17:26:44.743 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:28:41.425 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:35:39.789 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 28, sp_rank 0: update step: 171, train_steps: 6840 +2026-06-20 17:35:40.432 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 28, dp_rank 14: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=0 +2026-06-20 17:35:40.606 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 17:35:40.835 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:35:40.943 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:36:48.446 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000171/group_7/index_6228_seed_10043_dp_14_rank_28_sample_0_slot_0.mp4 +2026-06-20 17:36:52.708 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 17:37:10.227 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:37:10.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 28 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:37:10.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 28 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:38:19.343 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 28: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000171/group_7/index_6228_seed_10044_dp_14_rank_28_sample_1_slot_1.mp4 +2026-06-20 17:38:24.438 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:4 for reward scoring. +2026-06-20 17:38:42.466 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:38:42.537 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 28: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 17:38:42.598 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 28 saved 4 entries for group 7 +2026-06-20 17:38:43.757 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 28: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:40:36.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 28: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.89s (1.88min) +2026-06-20 17:40:37.378 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 28, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-29 01:30:50.943 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 01:30:51.613 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 01:30:51.615 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 01:53:15.070 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 01:53:15.733 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 01:53:15.733 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 02:22:43.034 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 02:22:43.700 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 02:22:43.701 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 02:48:38.424 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 02:48:39.100 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 02:48:39.101 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:09:53.532 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 03:09:54.194 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:09:54.197 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:33:45.973 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 03:33:46.631 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:33:46.633 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:00:51.590 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 04:00:52.251 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:00:52.252 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:27:42.693 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 04:27:43.368 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:27:43.368 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:53:59.919 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 04:54:00.599 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:54:00.601 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 05:18:10.282 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 05:18:10.963 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 05:18:10.964 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 05:41:14.179 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 05:41:14.853 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 05:41:14.856 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:04:28.641 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 06:04:29.322 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:04:29.323 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:26:59.352 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 06:27:00.022 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:27:00.022 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:51:34.380 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 06:51:35.040 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:51:35.041 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 07:15:17.463 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 07:15:18.167 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 07:15:18.168 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 07:36:26.947 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 07:36:27.604 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 07:36:27.604 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:08:49.859 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 08:08:50.529 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:08:50.530 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:33:27.295 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 08:33:27.958 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:33:27.959 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:01:16.359 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 09:01:17.013 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:01:17.013 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:24:41.710 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 09:24:42.366 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:24:42.371 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:49:41.128 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 09:49:41.808 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:49:41.810 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 10:12:10.779 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 10:12:11.437 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 10:12:11.438 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 10:38:25.205 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 10:38:25.885 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 10:38:25.886 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:07:04.316 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 11:07:04.972 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:07:04.972 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:29:17.204 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 11:29:17.876 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:29:17.877 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:57:16.388 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 11:57:17.054 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:57:17.055 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 12:21:48.568 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 12:21:49.242 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 12:21:49.243 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 12:48:43.192 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 12:48:43.852 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 12:48:43.852 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 13:14:42.915 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 13:14:43.588 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 13:14:43.591 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 13:39:09.858 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 13:39:10.512 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 13:39:10.516 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:05:47.441 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 14:05:48.110 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:05:48.112 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:30:31.198 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 14:30:31.870 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:30:31.870 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:51:45.123 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 14:51:45.804 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:51:45.804 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 15:23:01.697 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 15:23:02.366 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 15:23:02.369 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 15:48:53.347 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 15:48:54.034 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 15:48:54.037 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 16:17:44.756 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 16:17:45.419 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 16:17:45.422 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 16:43:02.854 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 16:43:03.515 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 16:43:03.516 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:11:39.294 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 17:11:39.975 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:11:39.978 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:34:38.509 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 17:34:39.177 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:34:39.177 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:56:06.182 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 17:56:06.849 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:56:06.849 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 18:18:46.100 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 18:18:46.773 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 18:18:46.773 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 18:40:43.161 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 18:40:43.844 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 18:40:43.846 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_30.log b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_30.log new file mode 100644 index 0000000000000000000000000000000000000000..545addf208bfd29cca2e6b6e1d9843cf89e48d9d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_30.log @@ -0,0 +1,3241 @@ +2026-06-19 01:22:18.212 | INFO | __main__:__init__:933 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-19 01:22:18.891 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-19 01:22:18.892 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-19 01:30:17.294 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 01:30:31.778 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 01:30:31.781 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:30:31.782 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:31:52.636 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_7/index_7599_seed_8268_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 01:31:56.458 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 01:32:13.780 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:32:13.789 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:32:13.789 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:33:22.226 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_7/index_7599_seed_8269_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 01:33:25.425 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 01:33:42.553 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:33:45.515 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 01:34:00.969 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 01:35:56.997 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.94s (1.93min) +2026-06-19 01:35:57.695 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 01:37:54.182 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 01:37:57.220 | INFO | hyvideo.utils.grpo_utils:train_one_step:3127 - Rank 30, sp_rank 0: init adaptive KL controller (init=1.000e-07, min=1.000e-07, max=1.000e-05) +2026-06-19 01:45:11.875 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 1, train_steps: 40 +2026-06-19 01:45:12.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 01:45:12.709 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 01:45:12.742 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:45:12.817 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:46:18.387 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000001/group_7/index_9764_seed_19709_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 01:46:23.283 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 01:46:40.771 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:46:40.780 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:46:40.780 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:47:47.524 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000001/group_7/index_9764_seed_19710_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 01:47:52.544 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 01:48:10.935 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:48:10.944 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 01:48:11.915 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 01:50:07.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.70s (1.93min) +2026-06-19 01:50:08.466 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 01:52:06.953 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 01:59:06.762 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 2, train_steps: 80 +2026-06-19 01:59:07.381 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 01:59:07.544 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 01:59:07.610 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:59:07.646 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:00:12.728 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000002/group_7/index_4618_seed_27259_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 02:00:17.770 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:00:35.409 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:00:35.416 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:00:35.416 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:01:46.181 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000002/group_7/index_4618_seed_27260_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 02:01:50.997 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:02:09.026 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:02:10.528 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:02:11.719 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:04:06.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.88s (1.91min) +2026-06-19 02:04:07.374 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:06:03.754 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:13:04.860 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 3, train_steps: 120 +2026-06-19 02:13:05.466 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 02:13:05.622 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:13:05.686 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:13:05.732 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:14:11.036 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_7/index_9149_seed_30245_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 02:14:15.754 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:14:33.570 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:14:33.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:14:33.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:15:40.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_7/index_9149_seed_30246_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 02:15:45.550 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:16:03.678 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:16:04.655 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:16:05.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:17:59.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.77s (1.90min) +2026-06-19 02:18:00.448 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:19:59.350 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:26:59.883 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 4, train_steps: 160 +2026-06-19 02:27:00.546 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 02:27:00.704 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:27:00.767 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:27:00.895 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:28:06.386 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000004/group_7/index_8338_seed_31413_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 02:28:11.102 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:28:28.987 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:28:28.997 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:28:28.997 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:29:35.046 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000004/group_7/index_8338_seed_31414_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 02:29:39.543 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:29:58.022 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:30:00.316 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:30:01.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:31:55.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.62s (1.91min) +2026-06-19 02:31:56.651 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:33:53.222 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:40:53.219 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 5, train_steps: 200 +2026-06-19 02:40:53.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 02:40:54.001 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:40:54.039 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:40:54.086 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:41:59.649 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_7/index_3286_seed_20133_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 02:42:04.618 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:42:22.407 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:42:22.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:42:22.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:43:28.518 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_7/index_3286_seed_20134_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 02:43:32.534 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:43:49.986 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:43:51.812 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:43:52.692 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:45:46.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.50s (1.89min) +2026-06-19 02:45:46.993 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:47:43.351 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:54:40.664 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 6, train_steps: 240 +2026-06-19 02:54:41.360 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 02:54:41.506 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:54:41.569 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:54:41.905 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:55:48.318 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_7/index_2712_seed_22677_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 02:55:53.317 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:56:11.205 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:56:11.216 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:56:11.216 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:57:19.257 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_7/index_2712_seed_22678_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 02:57:24.096 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:57:42.113 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:57:42.949 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:57:43.932 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:59:37.354 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.38s (1.89min) +2026-06-19 02:59:38.043 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:01:37.729 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:08:37.491 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 7, train_steps: 280 +2026-06-19 03:08:38.147 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 03:08:38.298 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:08:38.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:08:38.477 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:09:43.940 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_7/index_9335_seed_32633_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 03:09:48.705 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:10:06.507 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:10:06.515 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:10:06.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:11:14.612 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_7/index_9335_seed_32634_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 03:11:19.397 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:11:37.403 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:11:38.337 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:11:39.622 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:13:34.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.68s (1.91min) +2026-06-19 03:13:34.977 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:15:32.042 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:22:31.154 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 8, train_steps: 320 +2026-06-19 03:22:31.816 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 03:22:32.020 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:22:32.141 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:22:32.235 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:23:38.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000008/group_7/index_9763_seed_21218_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 03:23:42.680 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:24:00.309 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:24:00.318 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:24:00.318 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:25:08.951 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000008/group_7/index_9763_seed_21219_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 03:25:11.975 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:25:29.556 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:25:33.100 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:25:33.983 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:27:28.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-19 03:27:29.431 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:29:26.903 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:36:25.654 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 9, train_steps: 360 +2026-06-19 03:36:26.346 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 03:36:26.470 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:36:26.507 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:36:26.576 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:37:32.141 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_7/index_9813_seed_26729_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 03:37:37.141 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:37:54.919 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:37:54.925 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:37:54.925 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:39:01.921 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_7/index_9813_seed_26730_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 03:39:05.832 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:39:23.427 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:39:26.010 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:39:26.895 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:41:20.400 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.36s (1.89min) +2026-06-19 03:41:21.105 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:43:17.560 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:50:16.959 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 10, train_steps: 400 +2026-06-19 03:50:52.827 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=10 (missing=0, unexpected=0) +2026-06-19 03:50:56.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 03:50:56.338 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:50:56.420 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:50:56.548 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:52:03.411 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_7/index_1435_seed_11864_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 03:52:06.394 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:52:23.617 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:52:23.626 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:52:23.626 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:53:32.466 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_7/index_1435_seed_11865_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 03:53:37.520 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:53:57.519 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:53:57.828 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:53:58.828 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:55:52.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.65s (1.89min) +2026-06-19 03:55:53.210 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:57:49.763 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:04:48.379 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 11, train_steps: 440 +2026-06-19 04:04:49.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 04:04:49.227 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:04:49.279 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:04:49.380 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:05:55.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_7/index_3839_seed_16054_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 04:05:59.182 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:06:16.778 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:06:16.786 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:06:16.787 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:07:24.556 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_7/index_3839_seed_16055_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 04:07:29.575 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:07:47.966 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:07:48.853 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:07:49.856 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:09:43.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.25s (1.89min) +2026-06-19 04:09:43.797 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:11:40.268 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:18:37.293 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 12, train_steps: 480 +2026-06-19 04:18:37.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 04:18:38.071 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:18:38.116 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:18:38.164 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:19:44.050 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_7/index_8397_seed_23663_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 04:19:48.869 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:20:06.402 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:20:06.411 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:20:06.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:21:12.663 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_7/index_8397_seed_23664_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 04:21:16.686 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:21:34.659 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:21:37.180 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:21:38.104 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:23:33.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.19s (1.92min) +2026-06-19 04:23:34.070 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:25:31.242 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:32:29.892 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 13, train_steps: 520 +2026-06-19 04:32:30.514 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 04:32:30.751 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:32:30.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:32:30.958 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:33:36.844 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_7/index_4505_seed_26750_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 04:33:40.318 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:33:57.845 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:33:57.854 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:33:57.854 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:35:06.069 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_7/index_4505_seed_26751_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 04:35:11.237 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:35:29.646 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:35:30.313 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:35:31.310 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:37:24.102 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.77s (1.88min) +2026-06-19 04:37:24.791 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:39:21.077 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:46:18.587 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 14, train_steps: 560 +2026-06-19 04:46:19.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 04:46:19.415 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:46:19.496 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:46:19.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:47:25.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_7/index_6678_seed_31786_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 04:47:30.187 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:47:48.006 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:47:48.013 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:47:48.013 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:48:54.129 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_7/index_6678_seed_31787_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 04:48:59.092 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:49:17.234 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:49:17.445 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:49:18.432 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:51:12.963 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.49s (1.91min) +2026-06-19 04:51:13.650 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:53:12.995 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:00:11.831 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 15, train_steps: 600 +2026-06-19 05:00:12.526 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 05:00:12.734 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:00:12.868 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:00:12.944 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:01:18.572 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000015/group_7/index_9584_seed_9623_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 05:01:22.619 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:01:40.870 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:01:40.876 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:01:40.876 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:02:47.726 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000015/group_7/index_9584_seed_9624_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 05:02:52.891 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:03:11.001 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:03:11.349 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:03:12.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:05:07.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.64s (1.91min) +2026-06-19 05:05:08.061 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:07:05.976 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:14:04.909 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 16, train_steps: 640 +2026-06-19 05:14:05.573 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 05:14:05.724 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:14:05.785 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:14:05.916 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:15:11.081 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_7/index_8728_seed_3855_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 05:15:15.870 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:15:33.432 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:15:33.441 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:15:33.441 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:16:42.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_7/index_8728_seed_3856_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 05:16:47.769 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:17:05.625 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:17:06.988 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:17:07.959 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:19:03.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.02s (1.92min) +2026-06-19 05:19:03.741 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:21:01.960 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:28:03.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 17, train_steps: 680 +2026-06-19 05:28:04.006 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 05:28:04.148 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:28:04.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:28:04.277 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:29:09.897 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_7/index_8287_seed_15156_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 05:29:14.748 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:29:32.489 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:29:32.498 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:29:32.499 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:30:39.421 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_7/index_8287_seed_15157_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 05:30:44.250 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:31:02.102 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:31:03.926 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:31:05.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:32:58.689 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.48s (1.89min) +2026-06-19 05:32:59.387 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:34:56.266 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:41:58.101 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 18, train_steps: 720 +2026-06-19 05:41:58.777 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 05:41:58.975 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:41:59.110 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:41:59.213 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:43:04.395 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_7/index_2825_seed_30780_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 05:43:08.414 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:43:26.035 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:43:26.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:43:26.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:44:35.236 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_7/index_2825_seed_30781_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 05:44:38.638 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:44:57.916 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:45:00.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:45:01.124 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:46:55.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.88s (1.90min) +2026-06-19 05:46:55.797 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:48:52.329 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:55:50.499 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 19, train_steps: 760 +2026-06-19 05:55:51.178 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 05:55:51.355 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:55:51.455 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:55:51.547 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:56:57.045 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_7/index_3009_seed_26887_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 05:57:00.123 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:57:17.577 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:57:17.583 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:57:17.583 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:58:25.527 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_7/index_3009_seed_26888_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 05:58:28.750 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:58:46.266 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:58:49.068 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:58:49.958 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:00:43.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.10s (1.89min) +2026-06-19 06:00:43.876 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:02:40.340 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:09:39.759 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 20, train_steps: 800 +2026-06-19 06:10:16.272 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=20 (missing=0, unexpected=0) +2026-06-19 06:10:18.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 06:10:18.951 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:10:19.040 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:10:19.120 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:11:25.923 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_7/index_8665_seed_15913_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 06:11:29.717 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:11:47.092 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:11:47.102 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:11:47.102 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:12:54.609 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_7/index_8665_seed_15914_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 06:12:59.705 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:13:17.537 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:13:19.050 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:13:20.161 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:15:13.224 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.90s (1.88min) +2026-06-19 06:15:13.910 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:17:12.688 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:24:12.567 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 21, train_steps: 840 +2026-06-19 06:24:13.271 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 06:24:13.427 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:24:13.479 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:24:13.581 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:25:18.779 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_7/index_9748_seed_6102_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 06:25:23.580 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:25:41.526 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:25:41.531 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:25:41.531 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:26:47.693 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_7/index_9748_seed_6103_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 06:26:51.831 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:27:09.887 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:27:11.261 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:27:12.267 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:29:06.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.51s (1.91min) +2026-06-19 06:29:07.667 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:31:04.578 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:38:04.282 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 22, train_steps: 880 +2026-06-19 06:38:04.947 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 06:38:05.101 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:38:05.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:38:05.269 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:39:10.759 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_7/index_980_seed_30195_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 06:39:13.905 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:39:31.660 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:39:31.668 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:39:31.669 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:40:39.798 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_7/index_980_seed_30196_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 06:40:44.964 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:41:03.073 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:41:03.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:41:04.328 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:42:58.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.99s (1.90min) +2026-06-19 06:42:59.165 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:44:56.505 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:51:58.171 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 23, train_steps: 920 +2026-06-19 06:51:58.842 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 06:51:58.958 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:51:59.009 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:51:59.114 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:53:04.479 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000023/group_7/index_6384_seed_29770_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 06:53:09.289 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:53:27.250 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:53:27.257 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:53:27.257 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:54:34.908 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000023/group_7/index_6384_seed_29771_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 06:54:39.206 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:54:58.263 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:54:59.789 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:55:00.742 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:56:54.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.11s (1.90min) +2026-06-19 06:56:55.622 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:58:52.319 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:05:51.323 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 24, train_steps: 960 +2026-06-19 07:05:52.005 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 07:05:52.163 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:05:52.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:05:52.388 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:06:58.066 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_7/index_3392_seed_29720_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 07:07:01.165 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:07:18.811 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:07:18.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:07:18.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:08:27.341 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_7/index_3392_seed_29721_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 07:08:32.629 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:08:50.998 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:08:52.217 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:08:53.205 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:10:46.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-19 07:10:46.871 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:12:44.153 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:19:41.971 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 25, train_steps: 1000 +2026-06-19 07:20:52.105 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 07:20:52.313 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:20:52.448 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:20:52.527 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:21:58.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_7/index_6896_seed_8722_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 07:22:02.391 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:22:20.462 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:22:20.472 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:22:20.472 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:23:28.733 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_7/index_6896_seed_8723_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 07:23:33.924 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:23:52.473 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:23:52.666 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:23:53.664 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:25:46.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.81s (1.88min) +2026-06-19 07:25:47.206 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:27:44.380 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:34:43.680 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 26, train_steps: 1040 +2026-06-19 07:34:44.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 07:34:44.504 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:34:44.556 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:34:44.680 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:35:49.996 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_7/index_4002_seed_18278_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 07:35:53.884 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:36:12.144 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:36:12.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:36:12.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:37:20.983 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_7/index_4002_seed_18279_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 07:37:24.118 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:37:42.381 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:37:45.443 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:37:46.387 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:39:39.639 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.14s (1.89min) +2026-06-19 07:39:40.348 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:41:39.071 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:48:39.676 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 27, train_steps: 1080 +2026-06-19 07:48:40.347 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 07:48:40.443 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:48:40.475 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:48:40.506 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:49:47.451 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_7/index_7479_seed_26138_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 07:49:52.749 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:50:10.997 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:50:11.003 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:50:11.003 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:51:19.139 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_7/index_7479_seed_26139_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 07:51:23.923 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:51:42.209 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:51:43.285 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:51:44.276 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:53:38.298 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.99s (1.90min) +2026-06-19 07:53:39.007 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:55:38.422 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:02:38.172 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 28, train_steps: 1120 +2026-06-19 08:02:38.880 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 08:02:39.075 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:02:39.184 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:02:39.282 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:03:44.863 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_7/index_5718_seed_11902_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 08:03:49.091 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:04:07.292 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:04:07.302 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:04:07.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:05:17.278 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_7/index_5718_seed_11903_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 08:05:20.369 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:05:39.831 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:05:41.901 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:05:43.028 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:07:37.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.24s (1.90min) +2026-06-19 08:07:38.092 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:09:37.933 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:16:37.863 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 29, train_steps: 1160 +2026-06-19 08:16:38.691 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 08:16:38.809 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:16:38.845 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:16:38.958 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:17:44.642 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_7/index_4247_seed_11174_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 08:17:47.723 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:18:05.584 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:18:05.594 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:18:05.594 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:19:17.921 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_7/index_4247_seed_11175_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 08:19:23.076 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:19:41.496 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:19:42.881 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:19:43.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:21:37.264 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 08:21:37.961 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:23:37.406 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:30:39.546 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 30, train_steps: 1200 +2026-06-19 08:31:15.681 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=30 (missing=0, unexpected=0) +2026-06-19 08:31:18.259 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 08:31:18.472 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:31:18.588 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:31:18.667 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:32:26.323 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000030/group_7/index_1115_seed_27294_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 08:32:30.535 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:32:48.511 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:32:48.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:32:48.520 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:33:55.611 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000030/group_7/index_1115_seed_27295_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 08:33:58.733 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:34:16.922 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:34:19.980 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:34:21.049 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:36:14.510 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-19 08:36:15.227 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:38:14.171 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:45:14.285 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 31, train_steps: 1240 +2026-06-19 08:45:14.992 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 08:45:15.153 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:45:15.210 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:45:15.345 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:46:20.972 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_7/index_5156_seed_19056_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 08:46:24.098 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:46:42.039 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:46:42.045 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:46:42.045 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:47:50.576 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_7/index_5156_seed_19057_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 08:47:53.771 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:48:12.039 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:48:15.288 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:48:16.383 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:50:10.323 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.84s (1.90min) +2026-06-19 08:50:11.023 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:52:11.148 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:59:13.102 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 32, train_steps: 1280 +2026-06-19 08:59:13.777 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 08:59:13.968 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:59:14.071 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:59:14.180 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:00:19.825 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000032/group_7/index_8255_seed_16742_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 09:00:22.943 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:00:40.803 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:00:40.809 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:00:40.809 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:01:48.637 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000032/group_7/index_8255_seed_16743_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 09:01:51.837 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:02:09.927 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:02:13.436 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:02:14.539 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:04:08.713 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.91s (1.90min) +2026-06-19 09:04:09.420 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:06:10.406 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:13:13.282 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 33, train_steps: 1320 +2026-06-19 09:13:13.953 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 09:13:14.146 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:13:14.224 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:13:14.332 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:14:20.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000033/group_7/index_1567_seed_29390_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 09:14:23.315 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:14:41.296 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:14:41.306 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:14:41.306 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:15:50.069 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000033/group_7/index_1567_seed_29391_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 09:15:54.769 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:16:12.820 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:16:14.822 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:16:15.774 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:18:11.066 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.06s (1.92min) +2026-06-19 09:18:11.785 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:20:11.626 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:27:13.732 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 34, train_steps: 1360 +2026-06-19 09:27:14.445 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 09:27:14.596 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:27:14.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:27:14.792 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:28:20.201 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_7/index_4110_seed_23551_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 09:28:23.214 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:28:41.288 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:28:41.297 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:28:41.300 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:29:49.801 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_7/index_4110_seed_23552_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 09:29:54.956 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:30:13.685 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:30:13.695 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:30:14.977 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:32:08.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.81s (1.90min) +2026-06-19 09:32:09.592 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:34:09.778 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:41:12.395 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 35, train_steps: 1400 +2026-06-19 09:41:13.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 09:41:13.237 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:41:13.328 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:41:13.422 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:42:18.885 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000035/group_7/index_9612_seed_1511_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 09:42:23.582 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:42:41.928 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:42:41.937 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:42:41.937 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:43:48.411 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000035/group_7/index_9612_seed_1512_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 09:43:53.221 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:44:11.894 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:44:13.080 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:44:14.041 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:46:10.686 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 116.56s (1.94min) +2026-06-19 09:46:11.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:48:10.894 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:55:11.965 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 36, train_steps: 1440 +2026-06-19 09:55:12.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 09:55:12.778 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:55:12.827 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:55:12.938 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:56:18.742 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_7/index_9993_seed_20996_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 09:56:23.554 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:56:41.961 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:56:41.970 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:56:41.970 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:57:48.026 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_7/index_9993_seed_20997_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 09:57:52.916 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:58:11.482 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:58:12.646 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:58:13.868 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:00:08.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-19 10:00:09.672 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:02:10.007 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:09:11.706 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 37, train_steps: 1480 +2026-06-19 10:09:12.395 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 10:09:12.498 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:09:12.552 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:09:12.602 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:10:18.309 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_7/index_883_seed_4567_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 10:10:22.194 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:10:40.783 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:10:40.792 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:10:40.794 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:11:47.751 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_7/index_883_seed_4568_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 10:11:52.967 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:12:11.519 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:12:13.042 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:12:14.025 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:14:09.030 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.94s (1.92min) +2026-06-19 10:14:09.737 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:16:10.368 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:23:12.357 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 38, train_steps: 1520 +2026-06-19 10:23:13.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 10:23:13.248 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:23:13.373 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:23:13.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:24:18.925 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000038/group_7/index_2559_seed_1509_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 10:24:22.985 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:24:41.158 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:24:41.166 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:24:41.167 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:25:48.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000038/group_7/index_2559_seed_1510_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 10:25:53.778 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:26:12.657 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:26:12.667 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:26:13.876 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:28:09.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.16s (1.92min) +2026-06-19 10:28:09.954 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:30:10.745 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:37:11.953 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 39, train_steps: 1560 +2026-06-19 10:37:12.646 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 10:37:12.765 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:37:12.806 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:37:12.872 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:38:18.783 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000039/group_7/index_3930_seed_13382_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 10:38:23.650 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:38:42.242 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:38:42.250 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:38:42.253 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:41:17.386 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000039/group_7/index_3930_seed_13383_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 10:41:22.178 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:41:40.943 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:41:42.526 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:41:43.713 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:43:36.471 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.70s (1.88min) +2026-06-19 10:43:37.180 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:45:37.176 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:52:37.930 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 40, train_steps: 1600 +2026-06-19 10:53:14.060 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=40 (missing=0, unexpected=0) +2026-06-19 10:53:15.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 10:53:15.923 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:53:15.998 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:53:16.089 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:54:23.368 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_7/index_6349_seed_21467_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 10:54:28.175 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:54:46.274 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:54:46.282 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:54:46.282 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:55:52.424 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_7/index_6349_seed_21468_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 10:55:57.175 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:56:15.646 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:56:16.410 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:56:17.324 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:58:11.717 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.27s (1.90min) +2026-06-19 10:58:12.417 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:00:11.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:07:12.584 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 41, train_steps: 1640 +2026-06-19 11:07:13.291 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 11:07:13.429 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:07:13.472 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:07:13.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:08:19.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_7/index_1241_seed_13761_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 11:08:24.543 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:08:42.980 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:08:42.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:08:42.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:09:48.676 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_7/index_1241_seed_13762_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 11:09:53.677 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:10:12.470 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:10:12.668 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:10:13.885 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:12:08.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.28s (1.90min) +2026-06-19 11:12:09.087 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:14:08.757 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:21:10.893 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 42, train_steps: 1680 +2026-06-19 11:21:11.804 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 11:21:11.952 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:21:11.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:21:12.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:22:18.022 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000042/group_7/index_1454_seed_29056_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 11:22:22.950 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:22:41.433 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:22:41.442 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:22:41.442 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:23:47.282 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000042/group_7/index_1454_seed_29057_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 11:23:52.196 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:24:10.990 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:24:12.073 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:24:13.234 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:26:07.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.32s (1.91min) +2026-06-19 11:26:08.279 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:28:06.795 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:35:06.777 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 43, train_steps: 1720 +2026-06-19 11:35:07.458 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 11:35:07.589 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:35:07.631 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:35:07.702 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:36:13.254 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_7/index_3572_seed_3091_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 11:36:17.278 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:36:35.244 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:36:35.253 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:36:35.255 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:37:46.336 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_7/index_3572_seed_3092_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 11:37:50.761 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:38:09.587 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:38:10.299 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:38:11.333 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:40:05.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.38s (1.91min) +2026-06-19 11:40:06.549 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:42:03.524 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:49:04.436 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 44, train_steps: 1760 +2026-06-19 11:49:05.114 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 11:49:05.302 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:49:05.411 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:49:05.514 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:50:11.158 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000044/group_7/index_8774_seed_17449_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 11:50:15.974 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:50:35.448 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:50:35.458 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:50:35.458 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:51:41.581 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000044/group_7/index_8774_seed_17450_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 11:51:46.569 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:52:05.093 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:52:06.491 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:52:07.487 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:54:01.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.69s (1.89min) +2026-06-19 11:54:01.910 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:56:01.233 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:03:03.872 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 45, train_steps: 1800 +2026-06-19 12:03:04.557 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 12:03:04.769 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:03:04.907 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:03:05.008 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:04:10.585 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000045/group_7/index_8755_seed_21606_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 12:04:15.354 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:04:33.923 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:04:33.933 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:04:33.933 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:05:40.335 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000045/group_7/index_8755_seed_21607_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 12:05:44.608 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:06:02.984 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:06:05.347 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:06:06.277 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:08:01.417 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.05s (1.92min) +2026-06-19 12:08:02.101 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:10:01.380 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:17:03.596 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 46, train_steps: 1840 +2026-06-19 12:17:04.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 12:17:04.498 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:17:04.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:17:04.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:18:10.244 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_7/index_220_seed_12922_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 12:18:15.469 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:18:33.959 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:18:33.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:18:33.969 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:19:40.469 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_7/index_220_seed_12923_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 12:19:44.651 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:20:03.400 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:20:05.333 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:20:06.443 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:22:01.733 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.17s (1.92min) +2026-06-19 12:22:02.442 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:24:01.449 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:31:02.925 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 47, train_steps: 1880 +2026-06-19 12:31:03.606 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 12:31:03.711 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:31:03.752 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:31:03.800 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:32:09.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000047/group_7/index_8301_seed_4156_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 12:32:14.293 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:32:32.744 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:32:32.753 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:32:32.753 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:33:39.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000047/group_7/index_8301_seed_4157_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 12:33:43.326 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:34:02.074 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:34:03.182 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:34:04.253 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:35:58.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-19 12:35:58.982 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:37:55.889 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:44:57.758 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 48, train_steps: 1920 +2026-06-19 12:44:58.355 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 12:44:58.545 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:44:58.666 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:44:58.760 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:46:04.242 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_7/index_319_seed_10821_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 12:46:09.538 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:46:27.912 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:46:27.921 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:46:27.921 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:47:34.684 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_7/index_319_seed_10822_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 12:47:39.558 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:47:58.770 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:47:58.931 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:47:59.970 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:49:54.620 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-19 12:49:55.343 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:51:52.288 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:58:52.086 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 49, train_steps: 1960 +2026-06-19 12:58:52.783 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 12:58:52.989 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:58:53.092 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:58:53.210 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:59:58.583 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_7/index_9039_seed_23559_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 13:00:02.703 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:00:20.871 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:00:20.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:00:20.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:01:30.175 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_7/index_9039_seed_23560_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 13:01:35.407 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:01:54.054 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:01:54.369 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:01:55.371 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:03:48.446 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.04s (1.88min) +2026-06-19 13:03:49.149 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:05:45.466 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:12:45.271 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 50, train_steps: 2000 +2026-06-19 13:14:21.488 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=50 (missing=0, unexpected=0) +2026-06-19 13:14:23.203 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 13:14:23.371 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:14:23.469 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:14:23.551 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:15:31.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_7/index_5162_seed_10640_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 13:15:35.220 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:15:53.246 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:15:53.252 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:15:53.252 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:17:00.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_7/index_5162_seed_10641_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 13:17:06.182 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:17:24.315 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:17:25.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:17:26.126 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:19:20.081 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.84s (1.90min) +2026-06-19 13:19:20.808 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:21:17.522 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:28:18.472 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 51, train_steps: 2040 +2026-06-19 13:28:19.184 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 13:28:19.287 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:28:19.329 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:28:19.375 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:29:24.869 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000051/group_7/index_461_seed_27681_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 13:29:28.698 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:29:46.690 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:29:46.699 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:29:46.699 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:30:54.019 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000051/group_7/index_461_seed_27682_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 13:30:57.190 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:31:15.669 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:31:18.029 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:31:18.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:33:13.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.83s (1.90min) +2026-06-19 13:33:13.742 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:35:13.227 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:42:13.619 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 52, train_steps: 2080 +2026-06-19 13:42:14.276 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 13:42:14.407 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:42:14.441 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:42:14.526 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:43:20.635 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_7/index_885_seed_21558_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 13:43:23.885 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:43:43.307 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:43:43.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:43:43.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:44:50.206 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_7/index_885_seed_21559_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 13:44:55.574 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:45:15.276 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:45:15.287 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:45:16.404 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:47:10.481 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.03s (1.90min) +2026-06-19 13:47:11.197 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:49:11.264 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:56:12.600 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 53, train_steps: 2120 +2026-06-19 13:56:13.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 13:56:13.487 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:56:13.582 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:56:13.675 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:57:19.117 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_7/index_4701_seed_3971_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 13:57:23.360 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:57:41.584 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:57:41.592 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:57:41.593 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:58:48.639 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_7/index_4701_seed_3972_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 13:58:53.919 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:59:12.704 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:59:13.130 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:59:14.133 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:01:08.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.41s (1.91min) +2026-06-19 14:01:09.529 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:03:10.474 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:10:12.164 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 54, train_steps: 2160 +2026-06-19 14:10:12.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 14:10:13.050 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:10:13.135 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:10:13.260 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:11:18.474 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_7/index_6946_seed_27634_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 14:11:23.251 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:11:41.382 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:11:41.391 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:11:41.394 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:12:47.894 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_7/index_6946_seed_27635_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 14:12:53.072 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:13:11.605 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:13:12.866 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:13:13.883 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:15:08.493 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.42s (1.91min) +2026-06-19 14:15:09.218 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:17:09.305 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:24:11.910 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 55, train_steps: 2200 +2026-06-19 14:24:12.602 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 14:24:12.698 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:24:12.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:24:12.787 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:25:18.599 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_7/index_6874_seed_29260_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 14:25:22.345 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:25:40.865 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:25:40.874 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:25:40.875 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:26:48.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_7/index_6874_seed_29261_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 14:26:53.459 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:27:12.457 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:27:12.754 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:27:13.933 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:29:09.888 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.82s (1.93min) +2026-06-19 14:29:10.604 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:31:10.762 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:38:12.118 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 56, train_steps: 2240 +2026-06-19 14:38:12.810 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 14:38:12.903 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:38:12.945 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:38:12.993 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:39:18.736 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_7/index_5754_seed_25261_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 14:39:22.746 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:39:40.848 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:39:40.855 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:39:40.855 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:40:48.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_7/index_5754_seed_25262_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 14:40:53.916 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:41:12.485 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:41:13.092 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:41:14.103 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:43:08.781 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.66s (1.91min) +2026-06-19 14:43:09.475 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:45:09.716 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:52:11.847 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 57, train_steps: 2280 +2026-06-19 14:52:12.526 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 14:52:12.680 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:52:12.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:52:12.867 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:53:18.373 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_7/index_9633_seed_23393_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 14:53:22.447 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:53:40.783 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:53:40.792 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:53:40.794 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:54:48.701 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_7/index_9633_seed_23394_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 14:54:51.847 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:55:09.989 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:55:13.057 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:55:13.997 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:57:08.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.79s (1.91min) +2026-06-19 14:57:09.597 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:59:09.879 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:06:12.292 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 58, train_steps: 2320 +2026-06-19 15:06:12.985 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 15:06:13.161 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:06:13.262 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:06:13.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:07:18.994 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_7/index_1219_seed_3085_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 15:07:24.224 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:07:42.762 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:07:42.771 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:07:42.771 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:08:48.871 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_7/index_1219_seed_3086_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 15:08:53.791 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:09:11.821 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:09:13.062 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:09:13.978 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:11:08.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.71s (1.91min) +2026-06-19 15:11:09.498 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:13:09.809 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:20:12.055 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 59, train_steps: 2360 +2026-06-19 15:20:12.738 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 15:20:12.870 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:20:12.905 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:20:12.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:21:18.596 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_7/index_958_seed_29500_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 15:21:21.569 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:21:40.121 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:21:40.131 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:21:40.131 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:22:47.927 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_7/index_958_seed_29501_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 15:22:51.029 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:23:09.064 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:23:12.372 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:23:13.463 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:25:09.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.28s (1.92min) +2026-06-19 15:25:09.742 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:27:10.365 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:34:12.072 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 60, train_steps: 2400 +2026-06-19 15:34:47.737 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=60 (missing=0, unexpected=0) +2026-06-19 15:34:50.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 15:34:50.750 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:34:50.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:34:50.911 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:35:58.140 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_7/index_6082_seed_11622_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 15:36:02.705 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:36:21.274 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:36:21.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:36:21.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:37:28.426 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_7/index_6082_seed_11623_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 15:37:32.497 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:37:51.008 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:37:53.529 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:37:54.646 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:39:47.707 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-19 15:39:48.403 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:41:45.232 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:48:44.596 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 61, train_steps: 2440 +2026-06-19 15:48:45.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 15:48:45.442 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:48:45.499 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:48:45.653 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:49:51.360 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_7/index_3018_seed_25251_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 15:49:54.533 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:50:12.487 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:50:12.496 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:50:12.497 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:51:23.235 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_7/index_3018_seed_25252_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 15:51:26.493 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:51:44.508 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:51:47.610 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:51:48.503 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:53:41.681 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.04s (1.88min) +2026-06-19 15:53:42.399 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:55:39.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:02:39.007 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 62, train_steps: 2480 +2026-06-19 16:02:39.688 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 16:02:39.917 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:02:40.043 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:02:40.121 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:03:46.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000062/group_7/index_9347_seed_1920_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 16:03:49.909 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:04:07.970 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:04:07.978 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:04:07.979 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:05:18.709 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000062/group_7/index_9347_seed_1921_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 16:05:21.887 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:05:40.982 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:05:44.132 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:05:45.141 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:07:38.761 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.59s (1.89min) +2026-06-19 16:07:39.471 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:09:38.852 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:16:39.319 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 63, train_steps: 2520 +2026-06-19 16:16:40.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 16:16:40.210 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:16:40.287 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:16:40.392 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:17:47.055 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_7/index_6248_seed_27866_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 16:17:51.954 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:18:10.332 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:18:10.342 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:18:10.342 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:19:19.847 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_7/index_6248_seed_27867_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 16:19:24.766 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:19:43.388 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:19:45.197 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:19:46.207 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:21:39.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.36s (1.89min) +2026-06-19 16:21:40.303 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:23:38.880 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:30:40.934 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 64, train_steps: 2560 +2026-06-19 16:30:41.626 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 16:30:41.809 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:30:41.892 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:30:41.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:31:48.285 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_7/index_9334_seed_26019_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 16:31:52.360 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:32:10.240 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:32:10.251 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:32:10.251 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:33:20.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_7/index_9334_seed_26020_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 16:33:26.050 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:33:44.406 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:33:45.471 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:33:46.476 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:35:40.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.63s (1.89min) +2026-06-19 16:35:40.844 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:37:39.397 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:44:41.296 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 65, train_steps: 2600 +2026-06-19 16:44:41.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 16:44:42.171 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:44:42.301 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:44:42.376 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:45:49.001 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_7/index_9382_seed_28304_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 16:45:52.920 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:46:11.042 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:46:11.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:46:11.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:47:22.838 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_7/index_9382_seed_28305_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 16:47:28.186 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:47:46.855 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:47:47.482 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:47:48.489 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:49:42.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.72s (1.90min) +2026-06-19 16:49:42.958 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:51:40.035 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:58:41.277 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 66, train_steps: 2640 +2026-06-19 16:58:41.989 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 16:58:42.137 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:58:42.207 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:58:42.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:59:48.485 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_7/index_2001_seed_17928_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 16:59:53.322 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:00:11.519 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:00:11.529 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:00:11.532 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:01:19.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_7/index_2001_seed_17929_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 17:01:24.285 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:01:42.824 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:01:43.599 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:01:44.586 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:03:37.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.34s (1.89min) +2026-06-19 17:03:38.682 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:05:38.381 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:12:40.126 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 67, train_steps: 2680 +2026-06-19 17:12:40.910 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 17:12:41.312 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:12:41.489 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:12:41.605 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:13:48.181 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_7/index_6448_seed_14524_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 17:13:52.530 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:14:10.558 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:14:10.567 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:14:10.567 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:15:20.344 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_7/index_6448_seed_14525_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 17:15:25.476 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:15:44.064 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:15:44.529 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:15:45.538 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:17:38.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 17:17:39.679 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:19:39.111 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:26:39.545 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 68, train_steps: 2720 +2026-06-19 17:26:40.215 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 17:26:40.366 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:26:40.422 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:26:40.534 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:27:47.654 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_7/index_8843_seed_24265_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 17:27:52.504 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:28:10.840 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:28:10.850 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:28:10.850 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:29:19.959 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_7/index_8843_seed_24266_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 17:29:24.831 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:29:43.297 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:29:43.886 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:29:44.907 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:31:38.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.76s (1.90min) +2026-06-19 17:31:39.420 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:33:38.460 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:40:38.376 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 69, train_steps: 2760 +2026-06-19 17:40:39.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 17:40:39.167 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:40:39.185 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:40:39.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:41:45.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000069/group_7/index_6290_seed_16748_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 17:41:49.538 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:42:07.476 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:42:07.485 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:42:07.486 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:43:16.763 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000069/group_7/index_6290_seed_16749_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 17:43:21.956 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:43:40.832 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:43:41.069 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:43:42.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:45:35.854 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.55s (1.89min) +2026-06-19 17:45:36.569 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:47:37.115 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:54:38.580 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 70, train_steps: 2800 +2026-06-19 17:55:14.957 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=70 (missing=0, unexpected=0) +2026-06-19 17:55:16.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 17:55:17.185 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:55:17.294 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:55:17.375 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:56:24.974 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_7/index_1640_seed_923_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 17:56:29.799 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:56:47.970 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:56:47.980 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:56:47.980 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:57:54.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_7/index_1640_seed_924_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 17:57:59.192 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:58:17.366 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:58:18.228 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:58:19.448 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:00:13.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.58s (1.89min) +2026-06-19 18:00:13.838 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:02:13.851 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:09:13.900 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 71, train_steps: 2840 +2026-06-19 18:09:14.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 18:09:14.810 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:09:14.928 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:09:15.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:10:20.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_7/index_9153_seed_3718_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 18:10:24.822 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:10:43.044 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:10:43.053 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:10:43.053 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:11:51.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_7/index_9153_seed_3719_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 18:11:56.842 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:12:15.601 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:12:16.021 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:12:17.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:14:12.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.67s (1.93min) +2026-06-19 18:14:13.460 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:16:12.691 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:23:13.828 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 72, train_steps: 2880 +2026-06-19 18:23:14.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 18:23:14.638 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:23:14.685 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:23:14.763 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:24:20.865 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_7/index_5665_seed_22732_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 18:24:25.118 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:24:43.217 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:24:43.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:24:43.226 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:25:51.028 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_7/index_5665_seed_22733_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 18:25:54.224 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:26:12.408 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:26:14.711 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:26:15.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:28:10.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.45s (1.91min) +2026-06-19 18:28:11.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:30:10.986 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:37:12.796 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 73, train_steps: 2920 +2026-06-19 18:37:13.456 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 18:37:13.639 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:37:13.750 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:37:13.858 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:38:19.651 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_7/index_6858_seed_6954_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 18:38:22.887 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:38:41.116 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:38:41.122 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:38:41.122 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:39:50.029 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_7/index_6858_seed_6955_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 18:39:53.198 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:40:11.442 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:40:13.250 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:40:14.162 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:42:09.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.16s (1.92min) +2026-06-19 18:42:10.284 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:44:10.446 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:51:12.760 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 74, train_steps: 2960 +2026-06-19 18:51:13.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 18:51:13.604 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:51:13.638 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:51:13.731 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:52:19.689 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_7/index_5502_seed_3132_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 18:52:22.834 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:52:40.866 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:52:40.875 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:52:40.875 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:53:49.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_7/index_5502_seed_3133_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 18:53:54.957 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:54:13.500 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:54:13.982 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:54:14.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:56:10.312 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.21s (1.92min) +2026-06-19 18:56:11.031 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:58:11.441 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:05:14.025 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 75, train_steps: 3000 +2026-06-19 19:06:23.231 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 19:06:23.436 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:06:23.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:06:23.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:07:29.432 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000075/group_7/index_6863_seed_12044_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 19:07:33.524 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:07:51.786 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:07:51.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:07:51.800 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:08:58.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000075/group_7/index_6863_seed_12045_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 19:09:01.893 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:09:20.218 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:09:24.209 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:09:25.161 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:11:18.397 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.14s (1.89min) +2026-06-19 19:11:19.100 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:13:16.252 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:20:16.391 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 76, train_steps: 3040 +2026-06-19 19:20:16.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 19:20:17.234 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:20:17.331 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:20:17.411 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:21:22.596 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000076/group_7/index_4184_seed_3113_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 19:21:27.819 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:21:46.152 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:21:46.161 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:21:46.161 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:22:53.546 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000076/group_7/index_4184_seed_3114_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 19:22:58.386 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:23:16.709 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:23:18.455 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:23:19.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:25:13.441 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.66s (1.89min) +2026-06-19 19:25:14.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:27:12.732 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:34:13.268 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 77, train_steps: 3080 +2026-06-19 19:34:13.952 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 19:34:14.168 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:34:14.312 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:34:14.391 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:35:19.851 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_7/index_8938_seed_15553_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 19:35:24.791 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:35:43.298 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:35:43.307 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:35:43.307 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:36:49.353 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_7/index_8938_seed_15554_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 19:36:54.315 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:37:12.917 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:37:13.073 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:37:14.085 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:39:08.734 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.60s (1.91min) +2026-06-19 19:39:09.441 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:41:09.768 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:48:12.647 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 78, train_steps: 3120 +2026-06-19 19:48:13.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 19:48:13.470 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:48:13.508 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:48:13.556 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:49:19.626 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_7/index_9727_seed_18008_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 19:49:24.476 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:49:42.921 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:49:42.932 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:49:42.932 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:50:49.852 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_7/index_9727_seed_18009_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 19:50:54.805 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:51:13.313 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:51:13.323 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:51:14.353 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:53:09.267 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-19 19:53:09.975 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:55:09.984 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:02:13.175 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 79, train_steps: 3160 +2026-06-19 20:02:13.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 20:02:14.008 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:02:14.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:02:14.116 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:03:19.831 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000079/group_7/index_3920_seed_15938_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 20:03:24.669 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:03:44.032 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:03:44.042 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:03:44.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:04:49.644 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000079/group_7/index_3920_seed_15939_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 20:04:53.576 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:05:12.012 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:05:13.093 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:05:14.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:07:09.660 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.53s (1.93min) +2026-06-19 20:07:10.385 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:09:11.032 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:16:13.968 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 80, train_steps: 3200 +2026-06-19 20:16:51.126 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=80 (missing=0, unexpected=0) +2026-06-19 20:16:53.093 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 20:16:53.314 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:16:53.461 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:16:53.531 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:18:01.747 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_7/index_2621_seed_10019_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 20:18:04.950 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:18:23.315 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:18:23.326 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:18:23.326 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:19:31.067 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_7/index_2621_seed_10020_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 20:19:34.332 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:19:52.789 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:19:54.752 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:19:55.678 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:21:49.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-19 20:21:50.363 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:23:47.271 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:30:47.438 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 81, train_steps: 3240 +2026-06-19 20:30:48.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 20:30:48.284 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:30:48.322 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:30:48.424 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:31:54.580 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_7/index_8865_seed_16006_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 20:31:59.741 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:32:17.950 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:32:17.960 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:32:17.960 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:33:25.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_7/index_8865_seed_16007_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 20:33:29.962 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:33:48.550 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:33:49.273 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:33:50.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:35:44.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.70s (1.89min) +2026-06-19 20:35:44.752 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:37:41.676 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:44:40.150 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 82, train_steps: 3280 +2026-06-19 20:44:40.696 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 20:44:40.804 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:44:40.846 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:44:40.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:45:47.961 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_7/index_940_seed_6157_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 20:45:52.870 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:46:11.334 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:46:11.343 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:46:11.343 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:47:19.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_7/index_940_seed_6158_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 20:47:24.289 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:47:42.399 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:47:44.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:47:45.066 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:49:38.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.41s (1.89min) +2026-06-19 20:49:39.275 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:51:38.432 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:58:38.997 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 83, train_steps: 3320 +2026-06-19 20:58:39.739 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 20:58:39.947 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:58:39.997 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:58:40.068 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:59:46.692 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_7/index_4977_seed_16400_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 20:59:51.094 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:00:10.014 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:00:10.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:00:10.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:01:19.784 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_7/index_4977_seed_16401_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 21:01:24.752 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:01:43.510 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:01:44.304 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:01:45.315 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:03:38.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.28s (1.89min) +2026-06-19 21:03:39.339 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:05:38.817 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:12:38.999 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 84, train_steps: 3360 +2026-06-19 21:12:39.649 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 21:12:39.841 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:12:39.960 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:12:40.053 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:13:46.422 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_7/index_6200_seed_4439_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 21:13:51.251 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:14:09.825 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:14:09.835 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:14:09.836 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:15:19.346 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_7/index_6200_seed_4440_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 21:15:24.345 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:15:42.964 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:15:43.791 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:15:44.794 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:17:38.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.31s (1.89min) +2026-06-19 21:17:38.863 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:19:38.522 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:26:39.497 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 85, train_steps: 3400 +2026-06-19 21:26:40.168 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 21:26:40.272 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:26:40.314 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:26:40.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:27:47.548 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_7/index_1316_seed_392_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 21:27:52.417 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:28:10.894 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:28:10.904 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:28:10.904 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:29:19.860 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_7/index_1316_seed_393_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 21:29:24.670 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:29:43.231 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:29:43.697 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:29:44.706 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:31:38.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-19 21:31:38.988 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:33:38.755 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:40:38.652 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 86, train_steps: 3440 +2026-06-19 21:40:39.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 21:40:39.506 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:40:39.547 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:40:39.643 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:41:46.788 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_7/index_2634_seed_11412_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 21:41:50.918 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:42:08.943 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:42:08.949 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:42:08.949 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:43:18.635 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_7/index_2634_seed_11413_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 21:43:23.937 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:43:42.526 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:43:42.867 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:43:43.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:45:37.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 21:45:37.987 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:47:37.708 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:54:39.047 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 87, train_steps: 3480 +2026-06-19 21:54:39.744 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 21:54:39.961 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:54:40.085 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:54:40.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:55:46.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000087/group_7/index_8609_seed_32345_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 21:55:51.434 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:56:09.667 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:56:09.676 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:56:09.677 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:57:18.886 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000087/group_7/index_8609_seed_32346_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 21:57:22.932 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:57:41.345 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:57:43.169 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:57:44.119 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:59:37.604 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.39s (1.89min) +2026-06-19 21:59:38.318 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:01:38.553 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:08:39.997 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 88, train_steps: 3520 +2026-06-19 22:08:40.671 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 22:08:40.887 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:08:40.995 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:08:41.092 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:09:47.858 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_7/index_4075_seed_22436_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 22:09:51.087 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:10:09.171 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:10:09.181 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:10:09.181 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:11:19.451 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_7/index_4075_seed_22437_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 22:11:22.759 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:11:41.170 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:11:43.755 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:11:44.679 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:13:38.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.57s (1.89min) +2026-06-19 22:13:39.103 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:15:38.303 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:22:38.717 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 89, train_steps: 3560 +2026-06-19 22:22:39.425 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 22:22:39.613 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:22:39.735 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:22:39.866 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:23:46.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_7/index_1033_seed_16463_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 22:23:51.599 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:24:10.246 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:24:10.252 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:24:10.252 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:25:19.684 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_7/index_1033_seed_16464_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 22:25:24.684 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:25:43.333 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:25:43.683 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:25:44.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:27:38.162 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 22:27:38.886 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:29:38.493 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:36:39.419 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 90, train_steps: 3600 +2026-06-19 22:37:15.608 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=90 (missing=0, unexpected=0) +2026-06-19 22:37:17.308 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 22:37:17.515 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:37:17.632 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:37:17.726 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:38:24.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_7/index_4178_seed_19898_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 22:38:28.350 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:38:46.428 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:38:46.437 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:38:46.437 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:39:55.005 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_7/index_4178_seed_19899_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 22:40:00.419 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:40:18.926 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:40:19.241 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:40:20.239 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:42:14.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.70s (1.89min) +2026-06-19 22:42:14.863 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:44:13.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:51:14.858 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 91, train_steps: 3640 +2026-06-19 22:51:15.700 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 22:51:15.880 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:51:15.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:51:16.075 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:52:21.795 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000091/group_7/index_6489_seed_33232_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 22:52:27.037 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:52:45.397 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:52:45.404 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:52:45.404 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:53:51.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000091/group_7/index_6489_seed_33233_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 22:53:55.537 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:54:13.848 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:54:15.937 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:54:16.872 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:56:12.672 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.67s (1.93min) +2026-06-19 22:56:13.388 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:58:13.110 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:05:14.268 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 92, train_steps: 3680 +2026-06-19 23:05:15.118 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 23:05:15.283 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:05:15.358 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:05:15.499 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:06:21.097 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000092/group_7/index_8838_seed_22173_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 23:06:26.492 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:06:44.894 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:06:44.909 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:06:44.909 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:07:51.178 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000092/group_7/index_8838_seed_22174_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 23:07:55.269 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:08:13.631 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:08:16.058 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:08:16.993 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:10:11.751 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.62s (1.91min) +2026-06-19 23:10:12.471 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:12:11.281 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:19:13.488 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 93, train_steps: 3720 +2026-06-19 23:19:14.152 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 23:19:14.297 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:19:14.343 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:19:14.430 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:20:20.240 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_7/index_3185_seed_29232_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 23:20:23.388 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:20:41.979 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:20:41.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:20:41.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:21:50.141 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_7/index_3185_seed_29233_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 23:21:55.563 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:22:14.345 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:22:14.354 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:22:15.473 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:24:10.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.33s (1.92min) +2026-06-19 23:24:11.707 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:26:11.284 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:33:13.943 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 94, train_steps: 3760 +2026-06-19 23:33:14.640 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 23:33:14.781 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:33:14.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:33:14.909 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:34:20.610 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_7/index_1739_seed_24513_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 23:34:24.720 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:34:43.392 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:34:43.403 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:34:43.403 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:35:51.719 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_7/index_1739_seed_24514_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 23:35:56.941 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:36:15.612 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:36:16.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:36:17.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:38:12.062 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.89s (1.91min) +2026-06-19 23:38:12.784 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:40:11.320 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:47:11.876 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 95, train_steps: 3800 +2026-06-19 23:47:12.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-19 23:47:12.740 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:47:12.807 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:47:12.924 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:48:18.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_7/index_6920_seed_20677_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-19 23:48:23.169 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:48:42.174 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:48:42.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:48:42.186 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:49:48.248 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_7/index_6920_seed_20678_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-19 23:49:53.299 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:50:11.808 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:50:12.004 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:50:13.205 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:52:08.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.82s (1.91min) +2026-06-19 23:52:08.791 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:54:07.676 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:01:09.777 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 96, train_steps: 3840 +2026-06-20 00:01:10.476 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 00:01:10.705 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:01:10.924 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:01:11.034 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:02:16.904 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_7/index_8960_seed_28162_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 00:02:21.343 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:02:40.771 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:02:40.780 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:02:40.781 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:03:47.344 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_7/index_8960_seed_28163_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 00:03:52.643 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:04:11.773 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:04:11.783 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:04:12.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:06:07.928 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 00:06:08.644 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:08:07.688 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:15:08.510 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 97, train_steps: 3880 +2026-06-20 00:15:09.218 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 00:15:09.366 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:15:09.399 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:15:09.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:16:15.395 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_7/index_8885_seed_16241_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 00:16:20.281 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:16:40.787 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:16:40.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:16:40.794 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:17:47.291 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_7/index_8885_seed_16242_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 00:17:52.194 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:18:10.823 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:18:11.238 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:18:12.313 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:20:07.671 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.20s (1.92min) +2026-06-20 00:20:08.401 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:22:07.125 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:29:07.014 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 98, train_steps: 3920 +2026-06-20 00:29:07.692 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 00:29:07.847 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:29:07.905 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:29:08.037 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:30:13.852 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000098/group_7/index_8053_seed_15818_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 00:30:18.804 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:30:37.715 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:30:37.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:30:37.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:31:46.557 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000098/group_7/index_8053_seed_15819_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 00:31:50.693 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:32:08.904 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:32:10.419 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:32:11.391 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:34:06.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 00:34:07.118 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:36:03.849 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:43:05.350 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 99, train_steps: 3960 +2026-06-20 00:43:06.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 00:43:06.196 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:43:06.276 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:43:06.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:44:12.095 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_7/index_8117_seed_12406_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 00:44:15.254 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:44:33.614 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:44:33.623 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:44:33.623 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:45:46.271 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_7/index_8117_seed_12407_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 00:45:49.605 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:46:07.786 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:46:09.899 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:46:10.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:48:05.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.75s (1.91min) +2026-06-20 00:48:06.440 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:50:03.396 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:57:04.806 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 100, train_steps: 4000 +2026-06-20 00:58:47.538 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=100 (missing=0, unexpected=0) +2026-06-20 00:58:49.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 00:58:49.758 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:58:49.859 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:58:49.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:59:57.170 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_7/index_4642_seed_219_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 01:00:00.417 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:00:18.684 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:00:18.693 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:00:18.693 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:01:26.725 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_7/index_4642_seed_220_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 01:01:29.941 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:01:48.263 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:01:50.303 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:01:51.209 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:03:44.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-20 01:03:45.434 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:05:42.218 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:12:39.869 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 101, train_steps: 4040 +2026-06-20 01:12:40.568 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 01:12:40.764 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:12:40.855 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:12:40.932 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:13:47.792 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_7/index_5650_seed_4644_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 01:13:53.133 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:14:12.054 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:14:12.060 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:14:12.060 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:15:19.960 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_7/index_5650_seed_4645_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 01:15:24.854 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:15:43.181 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:15:43.499 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:15:44.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:17:38.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.62s (1.89min) +2026-06-20 01:17:38.931 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:19:38.324 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:26:37.810 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 102, train_steps: 4080 +2026-06-20 01:26:38.504 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 01:26:38.684 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:26:38.891 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:26:38.906 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:27:44.272 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_7/index_5132_seed_6377_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 01:27:49.093 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:28:07.735 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:28:07.745 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:28:07.747 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:29:16.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_7/index_5132_seed_6378_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 01:29:21.840 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:29:40.893 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:29:41.183 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:29:42.479 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:31:36.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-20 01:31:37.097 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:33:37.223 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:40:38.544 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 103, train_steps: 4120 +2026-06-20 01:40:39.244 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 01:40:39.335 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:40:39.379 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:40:39.426 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:41:46.054 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000103/group_7/index_1878_seed_29541_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 01:41:50.815 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:42:09.648 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:42:09.657 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:42:09.657 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:43:18.969 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000103/group_7/index_1878_seed_29542_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 01:43:23.931 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:43:42.760 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:43:43.334 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:43:44.361 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:45:37.767 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-20 01:45:38.487 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:47:38.265 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:54:37.795 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 104, train_steps: 4160 +2026-06-20 01:54:38.702 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 01:54:38.797 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:54:38.835 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:54:38.866 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:55:45.207 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_7/index_3917_seed_4020_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 01:55:49.414 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:56:08.329 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:56:08.338 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:56:08.338 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:57:18.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_7/index_3917_seed_4021_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 01:57:21.892 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:57:40.871 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:57:42.889 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:57:43.808 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:59:37.314 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-20 01:59:38.031 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:01:38.018 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:08:38.458 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 105, train_steps: 4200 +2026-06-20 02:08:39.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 02:08:39.258 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:08:39.297 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:08:39.343 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:09:46.401 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000105/group_7/index_7876_seed_4754_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 02:09:49.622 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:10:07.860 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:10:07.870 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:10:07.872 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:11:19.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000105/group_7/index_7876_seed_4755_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 02:11:22.299 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:11:40.842 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:11:43.945 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:11:44.871 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:13:38.217 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 02:13:38.925 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:15:38.267 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:22:37.857 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 106, train_steps: 4240 +2026-06-20 02:22:38.564 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 02:22:38.739 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:22:38.807 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:22:38.928 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:23:44.478 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000106/group_7/index_1463_seed_10759_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 02:23:49.797 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:24:08.777 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:24:08.787 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:24:08.787 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:25:16.693 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000106/group_7/index_1463_seed_10760_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 02:25:21.631 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:25:40.837 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:25:40.847 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:25:41.869 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:27:35.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.90s (1.90min) +2026-06-20 02:27:36.583 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:29:36.812 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:36:37.099 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 107, train_steps: 4280 +2026-06-20 02:36:37.802 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 02:36:37.978 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:36:38.077 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:36:38.189 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:37:43.683 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_7/index_3508_seed_15130_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 02:37:48.603 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:38:06.929 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:38:06.936 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:38:06.936 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:39:15.090 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_7/index_3508_seed_15131_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 02:39:19.374 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:39:37.498 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:39:39.229 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:39:40.504 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:41:35.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 02:41:36.072 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:43:35.144 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:50:35.010 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 108, train_steps: 4320 +2026-06-20 02:50:35.728 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 02:50:35.932 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:50:36.025 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:50:36.108 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:51:41.869 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_7/index_2244_seed_9317_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 02:51:47.076 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:52:06.905 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:52:06.915 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:52:06.915 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:53:13.107 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_7/index_2244_seed_9318_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 02:53:17.453 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:53:36.960 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:53:37.574 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:53:38.596 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:55:33.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.45s (1.91min) +2026-06-20 02:55:33.972 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:57:30.646 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:04:30.704 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 109, train_steps: 4360 +2026-06-20 03:04:31.405 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 03:04:31.580 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:04:31.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:04:31.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:05:37.134 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000109/group_7/index_8694_seed_5930_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 03:05:40.664 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:05:58.824 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:05:58.834 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:05:58.834 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:07:09.507 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000109/group_7/index_8694_seed_5931_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 03:07:12.756 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:07:31.063 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:07:33.407 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:07:34.328 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:09:29.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-20 03:09:30.015 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:11:26.575 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:18:25.334 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 110, train_steps: 4400 +2026-06-20 03:19:01.573 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=110 (missing=0, unexpected=0) +2026-06-20 03:19:03.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 03:19:03.433 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:19:03.500 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:19:03.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:20:11.502 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_7/index_1246_seed_5331_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 03:20:16.776 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:20:35.824 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:20:35.833 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:20:35.833 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:21:43.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_7/index_1246_seed_5332_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 03:21:47.421 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:22:05.677 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:22:07.524 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:22:08.466 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:24:03.966 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.41s (1.92min) +2026-06-20 03:24:04.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:26:02.472 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:33:05.402 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 111, train_steps: 4440 +2026-06-20 03:33:06.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 03:33:06.294 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:33:06.410 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:33:06.499 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:34:12.154 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_7/index_8494_seed_29816_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 03:34:15.356 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:34:33.467 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:34:33.476 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:34:33.476 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:35:46.494 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_7/index_8494_seed_29817_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 03:35:49.722 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:36:07.948 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:36:10.098 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:36:11.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:38:06.021 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 03:38:06.732 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:40:04.482 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:47:04.321 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 112, train_steps: 4480 +2026-06-20 03:47:05.042 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 03:47:05.139 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:47:05.181 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:47:05.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:48:10.996 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_7/index_6237_seed_3223_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 03:48:14.273 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:48:32.268 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:48:32.277 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:48:32.277 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:49:40.872 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_7/index_6237_seed_3224_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 03:49:46.194 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:50:04.530 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:50:04.983 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:50:06.289 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:52:00.943 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-20 03:52:01.657 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:54:00.963 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:01:04.009 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 113, train_steps: 4520 +2026-06-20 04:01:04.875 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 04:01:05.066 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:01:05.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:01:05.259 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:02:10.471 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000113/group_7/index_2104_seed_11880_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 04:02:14.172 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:02:32.186 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:02:32.195 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:02:32.195 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:03:42.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000113/group_7/index_2104_seed_11881_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 04:03:45.482 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:04:03.226 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:04:06.611 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:04:07.530 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:06:02.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.15s (1.92min) +2026-06-20 04:06:03.532 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:08:01.586 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:15:03.782 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 114, train_steps: 4560 +2026-06-20 04:15:04.473 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 04:15:04.639 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:15:04.710 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:15:04.847 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:16:10.681 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000114/group_7/index_1742_seed_29552_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 04:16:16.049 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:16:34.648 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:16:34.656 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:16:34.657 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:17:40.252 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000114/group_7/index_1742_seed_29553_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 04:17:44.231 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:18:02.559 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:18:04.152 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:18:05.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:20:00.180 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.72s (1.91min) +2026-06-20 04:20:00.895 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:22:00.099 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:29:01.474 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 115, train_steps: 4600 +2026-06-20 04:29:02.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 04:29:02.360 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:29:02.426 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:29:02.551 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:30:08.007 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_7/index_4321_seed_30350_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 04:30:13.285 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:30:31.518 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:30:31.527 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:30:31.527 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:31:38.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_7/index_4321_seed_30351_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 04:31:43.717 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:32:02.175 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:32:02.494 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:32:03.522 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:33:58.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.33s (1.91min) +2026-06-20 04:33:58.775 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:35:55.626 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:42:56.477 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 116, train_steps: 4640 +2026-06-20 04:42:57.160 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 04:42:57.342 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:42:57.430 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:42:57.548 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:44:03.212 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_7/index_530_seed_9292_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 04:44:07.138 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:44:25.520 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:44:25.529 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:44:25.529 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:45:33.790 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_7/index_530_seed_9293_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 04:45:38.716 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:45:58.373 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:45:58.380 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:45:59.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:47:53.698 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.25s (1.90min) +2026-06-20 04:47:54.407 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:49:50.888 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:56:50.231 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 117, train_steps: 4680 +2026-06-20 04:56:50.953 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 04:56:51.122 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:56:51.185 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:56:51.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:57:56.948 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_7/index_1458_seed_1169_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 04:58:01.698 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:58:20.108 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:58:20.117 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:58:20.117 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:59:27.894 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_7/index_1458_seed_1170_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 04:59:31.711 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:59:50.215 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:59:51.798 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:59:52.755 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:01:46.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.49s (1.89min) +2026-06-20 05:01:47.078 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:03:44.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:10:42.594 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 118, train_steps: 4720 +2026-06-20 05:10:43.296 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 05:10:43.443 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:10:43.496 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:10:43.611 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:11:49.503 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000118/group_7/index_7576_seed_18067_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 05:11:54.746 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:12:13.228 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:12:13.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:12:13.238 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:13:22.702 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000118/group_7/index_7576_seed_18068_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 05:13:26.485 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:13:44.979 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:13:46.957 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:13:47.917 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:15:41.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.26s (1.89min) +2026-06-20 05:15:42.001 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:17:38.993 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:24:37.913 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 119, train_steps: 4760 +2026-06-20 05:24:38.622 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 05:24:38.760 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:24:38.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:24:38.889 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:25:44.618 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_7/index_3062_seed_19646_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 05:25:47.829 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:26:05.937 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:26:05.945 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:26:05.946 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:27:14.871 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_7/index_3062_seed_19647_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 05:27:20.162 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:27:39.539 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:27:40.119 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:27:41.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:29:34.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-20 05:29:35.560 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:31:33.104 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:38:32.163 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 120, train_steps: 4800 +2026-06-20 05:39:08.700 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=120 (missing=0, unexpected=0) +2026-06-20 05:39:10.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 05:39:11.022 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:39:11.068 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:39:11.157 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:40:18.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000120/group_7/index_7070_seed_1841_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 05:40:22.904 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:40:41.482 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:40:41.491 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:40:41.493 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:41:48.551 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000120/group_7/index_7070_seed_1842_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 05:41:53.880 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:42:12.556 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:42:12.566 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:42:13.486 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:44:08.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.96s (1.92min) +2026-06-20 05:44:09.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:46:10.084 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:53:12.134 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 121, train_steps: 4840 +2026-06-20 05:53:12.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 05:53:12.995 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:53:13.105 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:53:13.184 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:54:19.387 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_7/index_6007_seed_32261_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 05:54:22.631 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:54:40.916 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:54:40.926 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:54:40.926 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:55:48.806 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_7/index_6007_seed_32262_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 05:55:54.224 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:56:13.148 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:56:13.156 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:56:14.190 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:58:09.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.97s (1.92min) +2026-06-20 05:58:10.076 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:00:10.574 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:07:11.959 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 122, train_steps: 4880 +2026-06-20 06:07:12.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 06:07:12.808 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:07:12.872 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:07:12.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:08:18.607 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_7/index_6530_seed_20321_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 06:08:23.278 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:08:41.897 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:08:41.905 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:08:41.906 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:09:48.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_7/index_6530_seed_20322_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 06:09:52.781 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:10:11.029 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:10:13.195 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:10:14.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:12:08.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.25s (1.90min) +2026-06-20 06:12:09.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:14:10.578 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:21:11.873 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 123, train_steps: 4920 +2026-06-20 06:21:12.561 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 06:21:12.731 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:21:12.807 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:21:12.904 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:22:18.659 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_7/index_8555_seed_33391_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 06:22:23.826 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:22:42.064 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:22:42.071 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:22:42.071 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:23:48.180 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_7/index_8555_seed_33392_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 06:23:53.143 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:24:11.834 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:24:12.071 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:24:13.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:26:08.788 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.47s (1.92min) +2026-06-20 06:26:09.511 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:28:09.832 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:35:11.851 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 124, train_steps: 4960 +2026-06-20 06:35:12.542 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 06:35:12.748 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:35:12.844 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:35:12.923 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:36:18.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_7/index_7531_seed_30710_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 06:36:22.233 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:36:40.807 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:36:40.816 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:36:40.816 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:37:48.142 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_7/index_7531_seed_30711_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 06:37:53.549 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:38:12.119 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:38:12.129 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:38:13.241 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:40:08.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.44s (1.92min) +2026-06-20 06:40:09.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:42:09.772 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:49:11.647 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 125, train_steps: 5000 +2026-06-20 06:50:15.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 06:50:15.377 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:50:15.464 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:50:15.570 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:51:21.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_7/index_1262_seed_33352_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 06:51:26.022 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:51:44.310 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:51:44.319 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:51:44.319 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:52:50.941 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_7/index_1262_seed_33353_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 06:52:55.878 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:53:14.290 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:53:15.926 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:53:16.949 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:55:11.654 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.68s (1.91min) +2026-06-20 06:55:12.372 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:57:11.763 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:04:11.924 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 126, train_steps: 5040 +2026-06-20 07:04:12.631 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 07:04:12.842 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:04:12.944 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:04:13.038 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:05:18.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_7/index_3243_seed_10882_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 07:05:23.561 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:05:41.739 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:05:41.749 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:05:41.749 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:06:49.346 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_7/index_3243_seed_10883_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 07:06:54.271 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:07:12.971 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:07:13.988 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:07:15.229 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:09:10.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.08s (1.92min) +2026-06-20 07:09:11.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:11:11.165 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:18:11.933 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 127, train_steps: 5080 +2026-06-20 07:18:12.631 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 07:18:12.831 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:18:12.948 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:18:13.025 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:19:18.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_7/index_2423_seed_21675_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 07:19:23.387 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:19:41.807 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:19:41.818 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:19:41.818 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:20:48.852 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_7/index_2423_seed_21676_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 07:20:53.816 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:21:12.468 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:21:13.050 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:21:14.074 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:23:09.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.50s (1.93min) +2026-06-20 07:23:10.327 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:25:10.158 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:32:11.179 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 128, train_steps: 5120 +2026-06-20 07:32:11.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 07:32:12.021 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:32:12.060 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:32:12.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:33:17.673 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_7/index_9709_seed_25089_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 07:33:22.556 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:33:40.911 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:33:40.919 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:33:40.920 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:34:48.565 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_7/index_9709_seed_25090_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 07:34:52.770 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:35:11.218 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:35:13.016 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:35:13.993 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:37:08.641 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.56s (1.91min) +2026-06-20 07:37:09.373 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:39:09.929 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:46:11.992 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 129, train_steps: 5160 +2026-06-20 07:46:12.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 07:46:12.849 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:46:12.891 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:46:12.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:47:19.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_7/index_3016_seed_11056_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 07:47:22.426 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:47:40.905 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:47:40.911 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:47:40.912 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:48:49.078 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_7/index_3016_seed_11057_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 07:48:54.504 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:49:13.055 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:49:13.070 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:49:14.101 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:51:09.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-20 07:51:09.758 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:53:10.303 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:00:11.787 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 130, train_steps: 5200 +2026-06-20 08:00:49.492 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=130 (missing=0, unexpected=0) +2026-06-20 08:00:51.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 08:00:51.447 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:00:51.551 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:00:51.645 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:01:59.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_7/index_6050_seed_885_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 08:02:04.757 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:02:23.638 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:02:23.646 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:02:23.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:03:28.852 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_7/index_6050_seed_886_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 08:03:33.829 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:03:52.440 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:03:52.509 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:03:53.510 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:05:47.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.72s (1.90min) +2026-06-20 08:05:48.018 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:07:44.843 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:14:45.319 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 131, train_steps: 5240 +2026-06-20 08:14:46.027 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 08:14:46.211 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:14:46.309 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:14:46.404 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:15:52.189 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_7/index_7621_seed_16738_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 08:15:56.313 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:16:14.679 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:16:14.689 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:16:14.689 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:17:23.306 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_7/index_7621_seed_16739_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 08:17:28.430 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:17:46.880 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:17:46.990 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:17:48.015 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:19:41.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.38s (1.89min) +2026-06-20 08:19:42.139 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:21:39.466 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:28:37.834 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 132, train_steps: 5280 +2026-06-20 08:28:38.522 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 08:28:38.731 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:28:38.842 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:28:38.917 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:29:44.925 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000132/group_7/index_6699_seed_26358_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 08:29:49.162 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:30:07.413 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:30:07.422 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:30:07.422 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:31:15.552 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000132/group_7/index_6699_seed_26359_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 08:31:20.883 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:31:40.841 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:31:40.849 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:31:41.875 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:33:35.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.44s (1.89min) +2026-06-20 08:33:36.119 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:35:35.261 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:42:34.338 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 133, train_steps: 5320 +2026-06-20 08:42:35.019 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 08:42:35.219 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:42:35.331 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:42:35.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:43:41.346 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_7/index_9380_seed_19655_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 08:43:45.521 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:44:03.485 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:44:03.495 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:44:03.495 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:45:11.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_7/index_9380_seed_19656_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 08:45:14.897 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:45:33.081 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:45:35.888 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:45:36.809 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:47:32.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.77s (1.93min) +2026-06-20 08:47:33.450 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:49:30.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:56:29.436 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 134, train_steps: 5360 +2026-06-20 08:56:30.118 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 08:56:30.338 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:56:30.457 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:56:30.536 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:57:36.330 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000134/group_7/index_6809_seed_15952_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 08:57:39.782 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:57:58.725 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:57:58.731 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:57:58.732 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:59:07.475 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000134/group_7/index_6809_seed_15953_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 08:59:10.824 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:59:29.196 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:59:32.041 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:59:32.954 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:01:26.319 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 09:01:27.048 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:03:23.803 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:10:22.996 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 135, train_steps: 5400 +2026-06-20 09:10:23.677 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 09:10:23.862 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:10:23.965 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:10:24.061 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:11:30.020 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000135/group_7/index_1261_seed_33366_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 09:11:35.306 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:11:53.897 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:11:53.907 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:11:53.907 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:13:00.324 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000135/group_7/index_1261_seed_33367_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 09:13:04.339 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:13:22.755 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:13:23.932 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:13:24.891 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:15:18.588 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.61s (1.89min) +2026-06-20 09:15:19.310 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:17:17.098 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:24:16.807 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 136, train_steps: 5440 +2026-06-20 09:24:17.533 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 09:24:17.725 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:24:17.824 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:24:17.904 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:25:23.758 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_7/index_8329_seed_24160_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 09:25:26.995 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:25:45.017 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:25:45.027 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:25:45.028 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:26:54.119 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_7/index_8329_seed_24161_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 09:26:57.332 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:27:15.765 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:27:18.656 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:27:19.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:29:12.959 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.15s (1.89min) +2026-06-20 09:29:13.686 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:31:13.176 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:38:13.220 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 137, train_steps: 5480 +2026-06-20 09:38:13.936 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 09:38:14.065 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:38:14.105 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:38:14.171 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:39:20.251 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000137/group_7/index_861_seed_19488_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 09:39:25.589 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:39:43.959 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:39:43.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:39:43.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:40:50.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000137/group_7/index_861_seed_19489_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 09:40:55.190 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:41:13.958 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:41:14.371 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:41:15.561 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:43:09.522 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-20 09:43:10.251 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:45:10.503 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:52:12.150 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 138, train_steps: 5520 +2026-06-20 09:52:12.866 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 09:52:13.022 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:52:13.074 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:52:13.154 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:53:18.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_7/index_604_seed_3410_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 09:53:23.805 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:53:42.134 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:53:42.143 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:53:42.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:54:49.305 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_7/index_604_seed_3411_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 09:54:54.209 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:55:13.210 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:55:13.383 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:55:14.610 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:57:10.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.46s (1.92min) +2026-06-20 09:57:10.922 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:59:11.152 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:06:12.681 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 139, train_steps: 5560 +2026-06-20 10:06:13.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 10:06:13.565 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:06:13.657 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:06:13.735 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:07:19.533 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_7/index_4896_seed_33001_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 10:07:24.383 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:07:42.628 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:07:42.634 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:07:42.635 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:08:49.927 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_7/index_4896_seed_33002_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 10:08:54.770 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:09:12.963 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:09:13.264 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:09:14.270 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:11:08.675 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.20s (1.90min) +2026-06-20 10:11:09.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:13:09.780 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:20:12.689 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 140, train_steps: 5600 +2026-06-20 10:20:49.872 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=140 (missing=0, unexpected=0) +2026-06-20 10:20:51.813 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 10:20:52.002 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:20:52.096 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:20:52.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:21:59.575 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000140/group_7/index_3735_seed_24712_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 10:22:04.469 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:22:22.953 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:22:22.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:22:22.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:23:28.804 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000140/group_7/index_3735_seed_24713_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 10:23:33.691 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:23:51.967 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:23:53.040 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:23:54.037 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:25:47.600 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-20 10:25:48.313 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:27:45.363 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:34:44.003 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 141, train_steps: 5640 +2026-06-20 10:34:44.690 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 10:34:44.897 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:34:45.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:34:45.102 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:35:50.544 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_7/index_6734_seed_5082_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 10:35:55.497 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:36:13.726 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:36:13.735 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:36:13.735 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:37:22.831 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_7/index_6734_seed_5083_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 10:37:27.832 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:37:46.014 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:37:46.846 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:37:47.866 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:39:41.415 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.51s (1.89min) +2026-06-20 10:39:42.148 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:41:39.463 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:48:38.469 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 142, train_steps: 5680 +2026-06-20 10:48:39.159 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 10:48:39.366 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:48:39.476 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:48:39.645 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:49:46.273 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000142/group_7/index_3542_seed_24097_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 10:49:50.644 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:50:08.860 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:50:08.871 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:50:08.871 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:51:19.140 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000142/group_7/index_3542_seed_24098_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 10:51:24.397 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:51:43.143 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:51:43.153 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:51:44.158 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:53:37.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-20 10:53:37.851 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:55:37.992 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:02:37.923 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 143, train_steps: 5720 +2026-06-20 11:02:38.635 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 11:02:38.839 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:02:38.951 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:02:39.015 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:03:44.810 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000143/group_7/index_6394_seed_10722_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 11:03:49.092 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:04:07.142 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:04:07.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:04:07.153 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:05:18.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000143/group_7/index_6394_seed_10723_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 11:05:23.411 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:05:42.016 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:05:42.169 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:05:43.334 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:07:36.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 11:07:37.313 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:09:36.842 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:16:37.103 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 144, train_steps: 5760 +2026-06-20 11:16:37.800 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 11:16:37.971 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:16:38.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:16:38.150 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:17:43.612 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_7/index_6114_seed_11112_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 11:17:47.594 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:18:05.678 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:18:05.688 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:18:05.688 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:19:13.340 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_7/index_6114_seed_11113_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 11:19:18.744 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:19:37.563 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:19:38.707 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:19:39.901 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:21:33.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.47s (1.89min) +2026-06-20 11:21:34.087 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:23:30.883 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:30:30.404 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 145, train_steps: 5800 +2026-06-20 11:30:31.098 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 11:30:31.286 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:30:31.388 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:30:31.468 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:31:36.811 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000145/group_7/index_3975_seed_5503_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 11:31:42.233 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:32:00.759 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:32:00.769 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:32:00.769 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:33:08.870 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000145/group_7/index_3975_seed_5504_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 11:33:13.830 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:33:32.598 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:33:33.207 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:33:34.222 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:35:28.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.89s (1.90min) +2026-06-20 11:35:28.900 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:37:25.684 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:44:23.921 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 146, train_steps: 5840 +2026-06-20 11:44:24.629 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 11:44:24.813 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:44:24.914 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:44:24.993 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:45:30.637 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_7/index_2947_seed_2987_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 11:45:35.519 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:45:53.733 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:45:53.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:45:53.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:46:59.071 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_7/index_2947_seed_2988_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 11:47:03.942 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:47:22.233 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:47:23.247 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:47:24.259 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:49:19.133 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-20 11:49:19.856 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:51:17.039 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:58:15.925 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 147, train_steps: 5880 +2026-06-20 11:58:16.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 11:58:16.735 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:58:16.777 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:58:16.846 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:59:22.995 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_7/index_5007_seed_10635_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 11:59:27.234 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:59:45.947 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:59:45.956 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:59:45.959 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:00:52.680 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_7/index_5007_seed_10636_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 12:00:57.626 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:01:16.221 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:01:16.232 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:01:17.271 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:03:12.586 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.29s (1.92min) +2026-06-20 12:03:13.293 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:05:12.903 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:12:12.700 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 148, train_steps: 5920 +2026-06-20 12:12:13.406 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 12:12:13.547 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:12:13.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:12:13.697 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:13:19.533 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000148/group_7/index_1344_seed_2330_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 12:13:24.344 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:13:42.657 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:13:42.667 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:13:42.667 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:14:49.523 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000148/group_7/index_1344_seed_2331_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 12:14:54.331 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:15:12.899 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:15:13.144 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:15:14.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:17:09.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.69s (1.91min) +2026-06-20 12:17:09.789 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:19:10.587 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:26:11.391 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 149, train_steps: 5960 +2026-06-20 12:26:12.136 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 12:26:12.301 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:26:12.372 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:26:12.472 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:27:18.165 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000149/group_7/index_9979_seed_536_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 12:27:23.142 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:27:41.805 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:27:41.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:27:41.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:28:48.694 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000149/group_7/index_9979_seed_537_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 12:28:53.536 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:29:12.164 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:29:13.029 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:29:14.039 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:31:08.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.72s (1.91min) +2026-06-20 12:31:09.533 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:33:09.830 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:40:10.993 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 150, train_steps: 6000 +2026-06-20 12:41:53.677 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=150 (missing=0, unexpected=0) +2026-06-20 12:41:56.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 12:41:56.253 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:41:56.371 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:41:56.488 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:43:03.718 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_7/index_9935_seed_20265_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 12:43:08.641 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:43:26.836 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:43:26.844 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:43:26.845 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:44:32.562 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_7/index_9935_seed_20266_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 12:44:37.562 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:44:58.106 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:44:59.744 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:45:00.773 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:46:55.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.70s (1.91min) +2026-06-20 12:46:56.236 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:48:52.987 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:55:51.607 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 151, train_steps: 6040 +2026-06-20 12:55:52.306 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 12:55:52.432 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:55:52.470 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:55:52.539 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:56:58.440 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_7/index_6917_seed_31838_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 12:57:03.418 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:57:21.939 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:57:21.949 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:57:21.949 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:58:27.153 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_7/index_6917_seed_31839_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 12:58:32.123 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:58:50.711 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:58:50.721 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:58:51.753 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:00:46.190 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.41s (1.91min) +2026-06-20 13:00:46.908 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:02:43.651 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:09:41.904 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 152, train_steps: 6080 +2026-06-20 13:09:42.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 13:09:42.755 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:09:42.835 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:09:42.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:10:48.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_7/index_6286_seed_11495_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 13:10:53.652 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:11:11.942 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:11:11.949 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:11:11.949 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:12:20.248 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_7/index_6286_seed_11496_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 13:12:25.264 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:12:43.900 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:12:43.910 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:12:44.926 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:14:38.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-20 13:14:39.492 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:16:38.549 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:23:37.225 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 153, train_steps: 6120 +2026-06-20 13:23:38.095 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 13:23:38.307 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:23:38.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:23:38.533 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:24:45.107 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_7/index_3682_seed_32179_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 13:24:50.074 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:25:08.251 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:25:08.260 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:25:08.261 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:26:15.065 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_7/index_3682_seed_32180_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 13:26:20.067 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:26:39.535 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:26:40.127 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:26:41.297 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:28:34.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.39s (1.89min) +2026-06-20 13:28:35.473 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:30:32.908 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:37:31.701 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 154, train_steps: 6160 +2026-06-20 13:37:32.394 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 13:37:32.583 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:37:32.679 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:37:32.758 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:38:38.501 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_7/index_6840_seed_16368_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 13:38:42.732 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:39:00.998 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:39:01.007 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:39:01.008 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:40:09.341 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_7/index_6840_seed_16369_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 13:40:14.521 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:40:33.065 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:40:33.788 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:40:34.826 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:42:29.329 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.47s (1.91min) +2026-06-20 13:42:30.062 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:44:27.159 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:51:24.977 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 155, train_steps: 6200 +2026-06-20 13:51:25.702 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 13:51:25.916 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:51:26.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:51:26.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:52:31.617 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_7/index_5812_seed_2514_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 13:52:36.476 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:52:54.561 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:52:54.570 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:52:54.570 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:54:01.127 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_7/index_5812_seed_2515_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 13:54:06.106 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:54:24.358 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:54:24.447 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:54:25.460 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:56:19.129 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.63s (1.89min) +2026-06-20 13:56:19.859 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:58:16.749 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:05:15.347 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 156, train_steps: 6240 +2026-06-20 14:05:16.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 14:05:16.180 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:05:16.221 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:05:16.315 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:06:22.296 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_7/index_7835_seed_17674_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 14:06:27.218 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:06:45.567 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:06:45.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:06:45.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:07:51.655 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_7/index_7835_seed_17675_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 14:07:55.953 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:08:14.267 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:08:15.931 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:08:16.904 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:10:11.722 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.74s (1.91min) +2026-06-20 14:10:12.457 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:12:11.103 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:19:12.246 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 157, train_steps: 6280 +2026-06-20 14:19:12.925 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 14:19:13.102 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:19:13.205 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:19:13.299 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:20:18.844 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000157/group_7/index_4500_seed_10810_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 14:20:23.937 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:20:42.306 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:20:42.315 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:20:42.315 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:21:47.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000157/group_7/index_4500_seed_10811_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 14:21:52.987 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:22:11.380 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:22:12.398 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:22:13.536 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:24:08.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.22s (1.92min) +2026-06-20 14:24:09.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:26:09.904 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:33:10.831 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 158, train_steps: 6320 +2026-06-20 14:33:11.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 14:33:11.704 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:33:11.804 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:33:11.883 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:34:17.334 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_7/index_9167_seed_20355_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 14:34:22.233 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:34:40.831 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:34:40.838 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:34:40.838 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:35:47.107 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_7/index_9167_seed_20356_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 14:35:52.026 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:36:10.523 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:36:10.935 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:36:12.038 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:38:07.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.88s (1.91min) +2026-06-20 14:38:07.784 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:40:05.257 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:47:04.569 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 159, train_steps: 6360 +2026-06-20 14:47:05.291 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 14:47:05.387 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:47:05.429 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:47:05.476 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:48:11.076 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_7/index_977_seed_188_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 14:48:15.060 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:48:33.327 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:48:33.337 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:48:33.337 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:49:40.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_7/index_977_seed_189_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 14:49:43.240 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:50:01.601 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:50:04.049 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:50:05.205 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:51:59.193 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.86s (1.90min) +2026-06-20 14:51:59.929 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:53:57.322 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:00:59.185 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 160, train_steps: 6400 +2026-06-20 15:01:37.084 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=160 (missing=0, unexpected=0) +2026-06-20 15:01:39.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 15:01:39.412 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:01:39.455 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:01:39.547 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:02:47.374 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_7/index_8595_seed_20939_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 15:02:50.900 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:03:09.333 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:03:09.340 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:03:09.340 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:04:18.827 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_7/index_8595_seed_20940_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 15:04:24.108 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:04:42.641 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:04:42.879 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:04:43.911 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:06:37.409 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.48s (1.89min) +2026-06-20 15:06:38.136 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:08:38.302 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:15:37.526 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 161, train_steps: 6440 +2026-06-20 15:15:38.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 15:15:38.389 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:15:38.495 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:15:38.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:16:44.216 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_7/index_3558_seed_20763_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 15:16:48.346 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:17:06.498 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:17:06.508 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:17:06.508 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:18:16.521 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_7/index_3558_seed_20764_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 15:18:19.690 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:18:37.950 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:18:40.375 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:18:41.526 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:20:35.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-20 15:20:36.332 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:22:36.155 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:29:36.905 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 162, train_steps: 6480 +2026-06-20 15:29:37.569 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 15:29:37.732 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:29:37.797 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:29:37.912 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:30:43.731 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_7/index_5912_seed_4949_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 15:30:46.884 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:31:05.109 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:31:05.116 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:31:05.119 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:32:11.812 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_7/index_5912_seed_4950_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 15:32:15.089 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:32:33.455 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:32:37.444 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:32:38.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:34:33.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.48s (1.91min) +2026-06-20 15:34:33.827 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:36:31.315 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:43:31.488 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 163, train_steps: 6520 +2026-06-20 15:43:32.172 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 15:43:32.345 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:43:32.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:43:32.537 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:44:38.470 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000163/group_7/index_9140_seed_25101_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 15:44:41.741 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:44:59.810 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:44:59.819 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:44:59.820 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:46:09.448 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000163/group_7/index_9140_seed_25102_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 15:46:12.674 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:46:31.347 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:46:33.676 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:46:34.593 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:48:29.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.47s (1.91min) +2026-06-20 15:48:29.934 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:50:26.537 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:57:25.517 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 164, train_steps: 6560 +2026-06-20 15:57:26.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 15:57:26.438 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:57:26.542 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:57:26.637 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:58:32.037 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_7/index_4340_seed_28845_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 15:58:37.246 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:58:55.708 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:58:55.719 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:58:55.719 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:00:02.524 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_7/index_4340_seed_28846_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 16:00:06.782 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:00:25.144 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:00:26.402 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:00:27.333 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:02:21.064 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.61s (1.89min) +2026-06-20 16:02:21.793 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:04:18.345 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:11:18.377 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 165, train_steps: 6600 +2026-06-20 16:11:19.086 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 16:11:19.240 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:11:19.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:11:19.403 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:12:25.224 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_7/index_7523_seed_26470_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 16:12:28.421 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:12:46.515 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:12:46.524 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:12:46.524 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:13:54.682 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_7/index_7523_seed_26471_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 16:13:57.851 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:14:16.090 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:14:19.106 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:14:20.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:16:14.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.75s (1.90min) +2026-06-20 16:16:14.870 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:18:14.124 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:25:13.936 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 166, train_steps: 6640 +2026-06-20 16:25:14.621 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 16:25:14.804 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:25:14.907 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:25:15.013 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:26:20.805 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_7/index_3807_seed_14231_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 16:26:24.072 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:26:42.105 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:26:42.115 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:26:42.115 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:27:50.071 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_7/index_3807_seed_14232_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 16:27:53.297 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:28:11.244 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:28:13.831 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:28:14.862 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:30:08.844 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.78s (1.90min) +2026-06-20 16:30:09.574 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:32:10.305 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:39:12.159 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 167, train_steps: 6680 +2026-06-20 16:39:12.847 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 16:39:12.969 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:39:13.004 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:39:13.069 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:40:19.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_7/index_3069_seed_10996_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 16:40:24.452 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:40:42.658 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:40:42.667 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:40:42.667 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:41:48.335 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_7/index_3069_seed_10997_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 16:41:52.414 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:42:10.719 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:42:12.095 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:42:13.249 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:44:08.745 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.36s (1.92min) +2026-06-20 16:44:09.477 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:46:10.018 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:53:11.387 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 168, train_steps: 6720 +2026-06-20 16:53:12.090 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 16:53:12.256 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:53:12.337 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:53:12.465 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:54:18.363 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_7/index_35_seed_19317_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 16:54:21.602 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:54:40.816 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:54:40.823 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:54:40.824 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:55:47.476 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_7/index_35_seed_19318_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 16:55:52.772 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:56:11.532 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:56:11.767 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:56:12.810 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:58:07.736 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-20 16:58:08.464 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:00:07.578 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:07:09.181 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 169, train_steps: 6760 +2026-06-20 17:07:09.877 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 17:07:10.009 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 17:07:10.071 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:07:10.184 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:08:16.121 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_7/index_5483_seed_8180_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 17:08:20.413 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 17:08:39.285 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:08:39.295 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:08:39.295 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:09:47.923 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_7/index_5483_seed_8181_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 17:09:53.155 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 17:10:11.608 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:10:11.904 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 17:10:12.940 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:12:08.255 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.20s (1.92min) +2026-06-20 17:12:08.969 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:14:09.319 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:21:11.053 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 170, train_steps: 6800 +2026-06-20 17:21:47.246 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=170 (missing=0, unexpected=0) +2026-06-20 17:21:49.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 17:21:49.388 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 17:21:49.459 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:21:49.592 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:22:56.400 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_7/index_8789_seed_4148_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 17:23:00.487 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 17:23:18.421 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:23:18.431 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:23:18.431 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:24:25.566 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_7/index_8789_seed_4149_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 17:24:30.879 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 17:24:49.128 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:24:49.925 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 17:24:50.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:26:44.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.05s (1.88min) +2026-06-20 17:26:44.743 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:28:41.425 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:35:39.792 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 30, sp_rank 0: update step: 171, train_steps: 6840 +2026-06-20 17:35:40.432 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 30, dp_rank 15: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=7, rank_in_group=1 +2026-06-20 17:35:40.658 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 17:35:40.926 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:35:41.011 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:36:47.947 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000171/group_7/index_6228_seed_10045_dp_15_rank_30_sample_2_slot_0.mp4 +2026-06-20 17:36:52.286 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 17:37:10.911 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:37:10.921 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 30 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:37:10.921 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 30 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:38:19.445 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 30: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000171/group_7/index_6228_seed_10046_dp_15_rank_30_sample_3_slot_1.mp4 +2026-06-20 17:38:22.979 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 17:38:42.497 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:38:42.537 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 30: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 17:38:43.681 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 30: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:40:36.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 30: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.89s (1.88min) +2026-06-20 17:40:37.378 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 30, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-29 01:30:50.945 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 01:30:51.624 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 01:30:51.624 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 01:53:15.071 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 01:53:15.734 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 01:53:15.735 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 02:22:43.035 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 02:22:43.709 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 02:22:43.712 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 02:48:38.422 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 02:48:39.091 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 02:48:39.092 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:09:53.533 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 03:09:54.216 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:09:54.216 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:33:45.548 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 03:33:46.215 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:33:46.215 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:00:50.933 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 04:00:51.606 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:00:51.607 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:27:42.693 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 04:27:43.379 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:27:43.380 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:54:00.342 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 04:54:00.998 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:54:01.002 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 05:18:10.282 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 05:18:10.948 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 05:18:10.949 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 05:41:14.592 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 05:41:15.268 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 05:41:15.269 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:04:29.327 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 06:04:30.004 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:04:30.005 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:26:59.349 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 06:27:00.008 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:27:00.011 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:51:34.862 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 06:51:35.519 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:51:35.519 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 07:15:17.036 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 07:15:17.719 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 07:15:17.720 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 07:36:27.381 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 07:36:28.038 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 07:36:28.038 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:08:49.213 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 08:08:49.882 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:08:49.882 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:33:26.885 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 08:33:27.558 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:33:27.559 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:01:16.357 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 09:01:17.034 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:01:17.035 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:24:41.707 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 09:24:42.385 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:24:42.387 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:49:41.127 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 09:49:41.793 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:49:41.793 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 10:12:11.200 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 10:12:11.850 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 10:12:11.853 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 10:38:25.204 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 10:38:25.859 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 10:38:25.859 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:07:04.315 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 11:07:04.975 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:07:04.976 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:29:17.203 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 11:29:17.861 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:29:17.864 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:57:16.387 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 11:57:17.052 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:57:17.055 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 12:21:48.990 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 12:21:49.642 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 12:21:49.643 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 12:48:42.768 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 12:48:43.444 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 12:48:43.445 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 13:14:42.912 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 13:14:43.598 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 13:14:43.598 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 13:39:09.199 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 13:39:09.877 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 13:39:09.878 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:05:47.440 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 14:05:48.120 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:05:48.120 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:30:31.199 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 14:30:31.866 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:30:31.867 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:51:45.123 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 14:51:45.794 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:51:45.795 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 15:23:02.385 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 15:23:03.062 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 15:23:03.063 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 15:48:53.767 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 15:48:54.434 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 15:48:54.435 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 16:17:45.196 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 16:17:45.855 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 16:17:45.858 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 16:43:02.206 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 16:43:02.884 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 16:43:02.885 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:11:39.711 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 17:11:40.379 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:11:40.379 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:34:38.508 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 17:34:39.168 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:34:39.171 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:56:06.182 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 17:56:06.847 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:56:06.849 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 18:18:46.101 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 18:18:46.766 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 18:18:46.766 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 18:40:43.582 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 18:40:44.240 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 18:40:44.241 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_37.log b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_37.log new file mode 100644 index 0000000000000000000000000000000000000000..9952019d63f525d78dc49d6ab6db9a3127eb5046 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_37.log @@ -0,0 +1,1994 @@ +2026-06-19 01:23:02.971 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-19 01:23:02.972 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-19 01:30:17.294 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 01:30:17.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:30:17.418 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:31:52.359 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:31:52.362 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:33:45.514 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 01:34:01.016 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 01:35:56.996 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.94s (1.93min) +2026-06-19 01:35:57.692 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 01:37:54.189 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 01:37:57.220 | INFO | hyvideo.utils.grpo_utils:train_one_step:3127 - Rank 37, sp_rank 0: init adaptive KL controller (init=1.000e-07, min=1.000e-07, max=1.000e-05) +2026-06-19 01:45:11.464 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 1, train_steps: 40 +2026-06-19 01:45:12.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 01:45:12.709 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:45:12.761 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:46:17.133 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:46:17.135 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:48:10.944 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 01:48:11.838 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 01:50:07.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.70s (1.93min) +2026-06-19 01:50:08.466 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 01:52:06.954 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 01:59:06.758 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 2, train_steps: 80 +2026-06-19 01:59:07.380 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 01:59:07.472 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:59:07.475 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:00:11.779 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:00:11.782 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:02:10.528 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 02:02:11.571 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:04:06.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.88s (1.91min) +2026-06-19 02:04:07.374 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:06:03.754 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:13:04.851 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 3, train_steps: 120 +2026-06-19 02:13:05.464 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 02:13:05.554 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:13:05.556 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:14:09.750 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:14:09.752 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:16:04.655 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 02:16:05.818 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:17:59.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.77s (1.90min) +2026-06-19 02:18:00.448 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:19:59.351 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:26:59.877 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 4, train_steps: 160 +2026-06-19 02:27:00.547 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 02:27:00.719 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:27:00.829 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:28:05.295 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:28:05.326 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:30:00.316 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 02:30:01.252 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:31:55.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.62s (1.91min) +2026-06-19 02:31:56.651 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:33:53.225 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:40:53.224 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 5, train_steps: 200 +2026-06-19 02:40:53.901 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 02:40:54.097 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:40:54.240 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:41:58.485 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:41:58.488 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:43:51.811 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 02:43:52.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:45:46.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.50s (1.89min) +2026-06-19 02:45:46.993 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:47:43.352 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:54:40.655 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 6, train_steps: 240 +2026-06-19 02:54:41.360 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 02:54:41.486 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:54:41.530 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:55:46.959 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:55:46.963 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:57:42.949 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 02:57:43.762 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:59:37.353 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.38s (1.89min) +2026-06-19 02:59:38.043 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:01:37.730 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:08:37.469 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 7, train_steps: 280 +2026-06-19 03:08:38.148 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 03:08:38.352 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:08:38.472 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:09:42.877 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:09:42.877 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:11:38.337 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 03:11:39.492 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:13:34.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.68s (1.91min) +2026-06-19 03:13:34.976 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:15:32.042 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:22:31.143 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 8, train_steps: 320 +2026-06-19 03:22:31.814 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 03:22:31.956 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:22:32.006 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:23:37.097 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:23:37.097 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:25:33.100 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 03:25:33.909 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:27:28.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-19 03:27:29.431 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:29:26.904 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:36:25.635 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 9, train_steps: 360 +2026-06-19 03:36:26.346 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 03:36:26.435 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:36:26.484 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:37:31.016 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:37:31.020 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:39:26.010 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 03:39:26.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:41:20.399 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.36s (1.89min) +2026-06-19 03:41:21.105 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:43:17.560 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:50:16.922 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 10, train_steps: 400 +2026-06-19 03:50:52.950 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=10 (missing=0, unexpected=0) +2026-06-19 03:50:56.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 03:50:56.315 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:50:56.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:52:02.172 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:52:02.174 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:53:57.827 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 03:53:58.785 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:55:52.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.65s (1.89min) +2026-06-19 03:55:53.209 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:57:49.764 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:04:48.376 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 11, train_steps: 440 +2026-06-19 04:04:49.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 04:04:49.201 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:04:49.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:05:53.859 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:05:53.859 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:07:48.852 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 04:07:49.785 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:09:43.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.25s (1.89min) +2026-06-19 04:09:43.797 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:11:40.269 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:18:37.286 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 12, train_steps: 480 +2026-06-19 04:18:37.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 04:18:38.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:18:38.255 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:19:42.826 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:19:42.828 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:21:37.180 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 04:21:38.122 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:23:33.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.19s (1.92min) +2026-06-19 04:23:34.070 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:25:31.243 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:32:29.818 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 13, train_steps: 520 +2026-06-19 04:32:30.512 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 04:32:30.688 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:32:30.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:33:35.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:33:35.354 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:35:30.312 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 04:35:31.158 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:37:24.102 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.77s (1.88min) +2026-06-19 04:37:24.791 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:39:21.076 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:46:18.580 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 14, train_steps: 560 +2026-06-19 04:46:19.247 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 04:46:19.414 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:46:19.517 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:47:24.202 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:47:24.206 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:49:17.445 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 04:49:18.242 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:51:12.963 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.49s (1.91min) +2026-06-19 04:51:13.650 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:53:12.995 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:00:11.824 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 15, train_steps: 600 +2026-06-19 05:00:12.526 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 05:00:12.727 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:00:12.830 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:01:17.682 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:01:17.684 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:03:11.349 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 05:03:12.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:05:07.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.64s (1.91min) +2026-06-19 05:05:08.060 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:07:05.977 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:14:04.893 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 16, train_steps: 640 +2026-06-19 05:14:05.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 05:14:05.766 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:14:05.874 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:15:10.262 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:15:10.266 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:17:06.988 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 05:17:07.922 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:19:03.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.02s (1.92min) +2026-06-19 05:19:03.741 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:21:01.962 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:28:03.324 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 17, train_steps: 680 +2026-06-19 05:28:04.006 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 05:28:04.097 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:28:04.147 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:29:08.704 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:29:08.706 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:31:03.926 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 05:31:04.984 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:32:58.688 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.48s (1.89min) +2026-06-19 05:32:59.387 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:34:56.265 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:41:58.094 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 18, train_steps: 720 +2026-06-19 05:41:58.777 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 05:41:58.952 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:41:59.066 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:43:03.278 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:43:03.282 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:45:00.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 05:45:01.016 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:46:55.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.88s (1.90min) +2026-06-19 05:46:55.797 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:48:52.338 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:55:50.490 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 19, train_steps: 760 +2026-06-19 05:55:51.177 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 05:55:51.287 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:55:51.319 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:56:55.781 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:56:55.785 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:58:49.068 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 05:58:49.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:00:43.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.10s (1.89min) +2026-06-19 06:00:43.876 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:02:40.341 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:09:39.749 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 20, train_steps: 800 +2026-06-19 06:10:16.190 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=20 (missing=0, unexpected=0) +2026-06-19 06:10:18.747 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 06:10:18.844 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:10:18.901 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:11:24.672 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:11:24.675 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:13:19.050 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 06:13:19.965 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:15:13.224 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.90s (1.88min) +2026-06-19 06:15:13.910 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:17:12.687 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:24:12.546 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 21, train_steps: 840 +2026-06-19 06:24:13.270 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 06:24:13.405 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:24:13.452 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:25:17.633 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:25:17.636 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:27:11.260 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 06:27:12.138 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:29:06.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.51s (1.91min) +2026-06-19 06:29:07.667 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:31:04.577 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:38:04.273 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 22, train_steps: 880 +2026-06-19 06:38:04.947 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 06:38:05.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:38:05.162 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:39:09.766 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:39:09.770 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:41:03.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 06:41:04.130 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:42:58.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.99s (1.90min) +2026-06-19 06:42:59.165 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:44:56.504 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:51:58.163 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 23, train_steps: 920 +2026-06-19 06:51:58.842 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 06:51:58.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:51:58.984 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:53:03.273 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:53:03.277 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:54:59.789 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 06:55:00.599 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:56:54.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.11s (1.90min) +2026-06-19 06:56:55.622 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:58:52.320 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:05:51.312 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 24, train_steps: 960 +2026-06-19 07:05:52.004 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 07:05:52.164 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:05:52.248 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:06:56.859 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:06:56.862 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:08:52.217 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 07:08:53.013 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:10:46.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-19 07:10:46.871 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:12:44.154 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:19:41.968 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 25, train_steps: 1000 +2026-06-19 07:20:52.104 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 07:20:52.212 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:20:52.265 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:21:56.864 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:21:56.873 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:23:52.666 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 07:23:53.488 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:25:46.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.81s (1.88min) +2026-06-19 07:25:47.205 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:27:44.381 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:34:43.670 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 26, train_steps: 1040 +2026-06-19 07:34:44.357 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 07:34:44.540 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:34:44.642 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:35:48.766 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:35:48.769 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:37:45.443 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 07:37:46.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:39:39.638 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.14s (1.89min) +2026-06-19 07:39:40.348 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:41:39.073 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:48:39.672 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 27, train_steps: 1080 +2026-06-19 07:48:40.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 07:48:40.481 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:48:40.517 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:49:45.937 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:49:45.940 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:51:43.285 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 07:51:44.232 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:53:38.298 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.99s (1.90min) +2026-06-19 07:53:39.007 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:55:38.423 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:02:38.173 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 28, train_steps: 1120 +2026-06-19 08:02:38.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 08:02:38.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:02:39.016 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:03:43.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:03:43.690 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:05:41.901 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 08:05:43.060 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:07:37.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.24s (1.90min) +2026-06-19 08:07:38.091 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:09:37.934 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:16:37.847 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 29, train_steps: 1160 +2026-06-19 08:16:38.691 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 08:16:38.783 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:16:38.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:17:43.537 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:17:43.543 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:19:42.881 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 08:19:43.830 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:21:37.264 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 08:21:37.961 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:23:37.408 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:30:39.540 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 30, train_steps: 1200 +2026-06-19 08:31:15.460 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=30 (missing=0, unexpected=0) +2026-06-19 08:31:18.258 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 08:31:18.414 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:31:18.489 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:32:24.899 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:32:24.902 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:34:19.980 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 08:34:21.080 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:36:14.510 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-19 08:36:15.226 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:38:14.173 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:45:14.275 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 31, train_steps: 1240 +2026-06-19 08:45:14.992 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 08:45:15.138 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:45:15.194 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:46:19.713 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:46:19.714 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:48:15.288 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 08:48:16.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:50:10.322 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.84s (1.90min) +2026-06-19 08:50:11.023 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:52:11.150 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:59:13.086 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 32, train_steps: 1280 +2026-06-19 08:59:13.778 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 08:59:13.980 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:59:14.108 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:00:18.536 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:00:18.539 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:02:13.436 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 09:02:14.599 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:04:08.713 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.91s (1.90min) +2026-06-19 09:04:09.420 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:06:10.409 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:13:13.351 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 33, train_steps: 1320 +2026-06-19 09:13:13.953 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 09:13:14.139 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:13:14.250 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:14:18.925 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:14:18.927 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:16:14.822 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 09:16:15.826 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:18:11.066 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.06s (1.92min) +2026-06-19 09:18:11.785 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:20:11.627 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:27:13.734 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 34, train_steps: 1360 +2026-06-19 09:27:14.444 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 09:27:14.605 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:27:14.656 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:28:19.020 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:28:19.022 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:30:13.695 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 09:30:14.946 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:32:08.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.81s (1.90min) +2026-06-19 09:32:09.592 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:34:09.780 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:41:12.395 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 35, train_steps: 1400 +2026-06-19 09:41:13.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 09:41:13.174 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:41:13.205 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:42:17.835 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:42:17.836 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:44:13.080 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 09:44:14.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:46:10.686 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 116.56s (1.94min) +2026-06-19 09:46:11.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:48:10.896 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:55:11.958 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 36, train_steps: 1440 +2026-06-19 09:55:12.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 09:55:12.828 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:55:12.932 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:56:17.664 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:56:17.664 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:58:12.646 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 09:58:13.802 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:00:08.966 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-19 10:00:09.671 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:02:10.008 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:09:11.698 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 37, train_steps: 1480 +2026-06-19 10:09:12.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 10:09:12.566 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:09:12.656 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:10:17.137 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:10:17.140 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:12:13.042 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 10:12:13.993 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:14:09.029 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.94s (1.92min) +2026-06-19 10:14:09.737 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:16:10.368 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:23:12.345 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 38, train_steps: 1520 +2026-06-19 10:23:13.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 10:23:13.258 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:23:13.398 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:24:17.675 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:24:17.678 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:26:12.666 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 10:26:13.823 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:28:09.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.16s (1.92min) +2026-06-19 10:28:09.954 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:30:10.746 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:37:11.945 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 39, train_steps: 1560 +2026-06-19 10:37:12.646 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 10:37:12.780 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:37:12.814 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:38:17.367 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:38:17.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:41:42.526 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 10:41:43.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:43:36.471 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.70s (1.88min) +2026-06-19 10:43:37.180 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:45:37.176 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:52:37.854 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 40, train_steps: 1600 +2026-06-19 10:53:14.132 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=40 (missing=0, unexpected=0) +2026-06-19 10:53:15.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 10:53:15.870 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:53:15.908 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:54:21.984 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:54:21.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:56:16.410 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 10:56:17.378 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:58:11.717 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.27s (1.90min) +2026-06-19 10:58:12.416 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:00:11.327 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:07:12.579 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 41, train_steps: 1640 +2026-06-19 11:07:13.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 11:07:13.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:07:13.558 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:08:17.923 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:08:17.923 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:10:12.668 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 11:10:13.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:12:08.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.28s (1.90min) +2026-06-19 11:12:09.087 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:14:08.757 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:21:10.888 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 42, train_steps: 1680 +2026-06-19 11:21:11.804 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 11:21:11.952 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:21:12.005 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:22:16.288 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:22:16.291 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:24:12.073 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 11:24:13.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:26:07.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.32s (1.91min) +2026-06-19 11:26:08.278 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:28:06.796 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:35:06.765 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 43, train_steps: 1720 +2026-06-19 11:35:07.459 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 11:35:07.656 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:35:07.768 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:36:11.837 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:36:11.841 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:38:10.298 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 11:38:11.123 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:40:05.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.38s (1.91min) +2026-06-19 11:40:06.549 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:42:03.524 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:49:04.423 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 44, train_steps: 1760 +2026-06-19 11:49:05.113 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 11:49:05.203 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:49:05.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:50:09.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:50:09.719 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:52:06.491 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 11:52:07.298 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:54:01.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.69s (1.89min) +2026-06-19 11:54:01.909 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:56:01.234 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:03:03.861 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 45, train_steps: 1800 +2026-06-19 12:03:04.558 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 12:03:04.770 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:03:04.912 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:04:09.281 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:04:09.286 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:06:05.347 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 12:06:06.165 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:08:01.417 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.05s (1.92min) +2026-06-19 12:08:02.101 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:10:01.382 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:17:03.650 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 46, train_steps: 1840 +2026-06-19 12:17:04.279 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 12:17:04.388 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:17:04.405 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:18:09.215 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:18:09.218 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:20:05.333 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 12:20:06.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:22:01.733 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.17s (1.92min) +2026-06-19 12:22:02.442 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:24:01.449 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:31:02.926 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 47, train_steps: 1880 +2026-06-19 12:31:03.607 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 12:31:03.715 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:31:03.754 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:32:07.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:32:07.970 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:34:03.182 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 12:34:04.337 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:35:58.279 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-19 12:35:58.982 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:37:55.892 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:44:57.667 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 48, train_steps: 1920 +2026-06-19 12:44:58.355 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 12:44:58.539 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:44:58.663 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:46:03.064 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:46:03.067 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:47:58.930 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 12:47:59.761 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:49:54.620 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-19 12:49:55.342 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:51:52.288 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:58:52.076 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 49, train_steps: 1960 +2026-06-19 12:58:52.784 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 12:58:52.979 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:58:53.081 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:59:57.338 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:59:57.341 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:01:54.369 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 13:01:55.184 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:03:48.446 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.04s (1.88min) +2026-06-19 13:03:49.148 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:05:45.467 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:12:45.263 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 50, train_steps: 2000 +2026-06-19 13:14:21.126 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=50 (missing=0, unexpected=0) +2026-06-19 13:14:23.203 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 13:14:23.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:14:23.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:15:29.631 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:15:29.634 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:17:25.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 13:17:25.807 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:19:20.081 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.84s (1.90min) +2026-06-19 13:19:20.807 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:21:17.524 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:28:18.459 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 51, train_steps: 2040 +2026-06-19 13:28:19.184 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 13:28:19.337 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:28:19.418 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:29:23.454 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:29:23.458 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:31:18.028 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 13:31:18.899 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:33:13.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.83s (1.90min) +2026-06-19 13:33:13.742 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:35:13.228 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:42:13.603 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 52, train_steps: 2080 +2026-06-19 13:42:14.276 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 13:42:14.447 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:42:14.524 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:43:19.204 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:43:19.207 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:45:15.287 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 13:45:16.218 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:47:10.481 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.03s (1.90min) +2026-06-19 13:47:11.196 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:49:11.264 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:56:12.597 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 53, train_steps: 2120 +2026-06-19 13:56:13.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 13:56:13.465 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:56:13.550 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:57:18.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:57:18.027 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:59:13.130 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 13:59:14.090 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:01:08.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.41s (1.91min) +2026-06-19 14:01:09.529 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:03:10.474 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:10:12.161 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 54, train_steps: 2160 +2026-06-19 14:10:12.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 14:10:13.009 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:10:13.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:11:17.401 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:11:17.402 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:13:12.866 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 14:13:13.834 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:15:08.493 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.42s (1.91min) +2026-06-19 14:15:09.217 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:17:09.306 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:24:11.905 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 55, train_steps: 2200 +2026-06-19 14:24:12.602 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 14:24:12.762 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:24:12.832 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:25:17.358 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:25:17.362 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:27:12.753 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 14:27:13.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:29:09.888 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.82s (1.93min) +2026-06-19 14:29:10.604 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:31:10.764 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:38:12.115 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 56, train_steps: 2240 +2026-06-19 14:38:12.811 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 14:38:12.980 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:38:13.062 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:39:17.551 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:39:17.554 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:41:13.092 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 14:41:14.053 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:43:08.780 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.66s (1.91min) +2026-06-19 14:43:09.475 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:45:09.717 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:52:11.839 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 57, train_steps: 2280 +2026-06-19 14:52:12.526 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 14:52:12.658 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:52:12.699 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:53:17.181 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:53:17.184 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:55:13.057 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 14:55:14.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:57:08.883 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.79s (1.91min) +2026-06-19 14:57:09.597 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:59:09.880 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:06:12.280 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 58, train_steps: 2320 +2026-06-19 15:06:12.985 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 15:06:13.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:06:13.282 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:07:17.973 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:07:17.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:09:13.062 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 15:09:14.042 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:11:08.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.71s (1.91min) +2026-06-19 15:11:09.498 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:13:09.809 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:20:12.050 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 59, train_steps: 2360 +2026-06-19 15:20:12.738 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 15:20:12.832 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:20:12.871 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:21:17.527 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:21:17.530 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:23:12.372 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 15:23:13.512 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:25:09.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.28s (1.92min) +2026-06-19 15:25:09.742 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:27:10.367 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:34:12.067 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 60, train_steps: 2400 +2026-06-19 15:34:48.309 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=60 (missing=0, unexpected=0) +2026-06-19 15:34:50.615 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 15:34:50.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:34:50.786 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:35:57.102 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:35:57.105 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:37:53.529 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 15:37:54.692 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:39:47.707 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-19 15:39:48.402 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:41:45.233 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:48:44.573 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 61, train_steps: 2440 +2026-06-19 15:48:45.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 15:48:45.465 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:48:45.542 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:49:50.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:49:50.075 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:51:47.609 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 15:51:48.421 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:53:41.681 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.04s (1.88min) +2026-06-19 15:53:42.399 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:55:39.325 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:02:38.994 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 62, train_steps: 2480 +2026-06-19 16:02:39.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 16:02:39.806 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:02:39.854 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:03:45.187 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:03:45.190 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:05:44.132 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 16:05:44.971 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:07:38.761 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.59s (1.89min) +2026-06-19 16:07:39.470 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:09:38.852 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:16:39.310 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 63, train_steps: 2520 +2026-06-19 16:16:40.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 16:16:40.172 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:16:40.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:17:45.904 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:17:45.904 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:19:45.197 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 16:19:46.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:21:39.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.36s (1.89min) +2026-06-19 16:21:40.303 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:23:38.881 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:30:40.923 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 64, train_steps: 2560 +2026-06-19 16:30:41.625 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 16:30:41.761 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:30:41.805 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:31:47.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:31:47.318 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:33:45.471 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 16:33:46.288 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:35:40.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.63s (1.89min) +2026-06-19 16:35:40.844 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:37:39.397 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:44:41.286 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 65, train_steps: 2600 +2026-06-19 16:44:41.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 16:44:42.096 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:44:42.142 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:45:47.201 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:45:47.201 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:47:47.481 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 16:47:48.297 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:49:42.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.72s (1.90min) +2026-06-19 16:49:42.958 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:51:40.036 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:58:41.267 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 66, train_steps: 2640 +2026-06-19 16:58:41.990 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 16:58:42.171 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:58:42.278 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:59:47.222 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:59:47.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:01:43.599 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 17:01:44.420 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:03:37.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.34s (1.89min) +2026-06-19 17:03:38.681 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:05:38.381 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:12:40.112 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 67, train_steps: 2680 +2026-06-19 17:12:40.909 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 17:12:41.311 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:12:41.476 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:13:47.017 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:13:47.020 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:15:44.529 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 17:15:45.343 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:17:38.956 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 17:17:39.679 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:19:39.111 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:26:39.539 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 68, train_steps: 2720 +2026-06-19 17:26:40.216 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 17:26:40.418 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:26:40.527 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:27:46.430 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:27:46.430 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:29:43.886 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 17:29:44.723 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:31:38.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.76s (1.90min) +2026-06-19 17:31:39.420 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:33:38.460 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:40:38.352 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 69, train_steps: 2760 +2026-06-19 17:40:39.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 17:40:39.127 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:40:39.158 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:41:43.754 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:41:43.757 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:43:41.069 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 17:43:41.977 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:45:35.854 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.55s (1.89min) +2026-06-19 17:45:36.569 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:47:37.118 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:54:38.567 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 70, train_steps: 2800 +2026-06-19 17:55:14.690 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=70 (missing=0, unexpected=0) +2026-06-19 17:55:16.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 17:55:17.115 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:55:17.169 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:56:23.763 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:56:23.766 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:58:18.228 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 17:58:19.277 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:00:13.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.58s (1.89min) +2026-06-19 18:00:13.838 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:02:13.852 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:09:13.884 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 71, train_steps: 2840 +2026-06-19 18:09:14.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 18:09:14.801 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:09:14.939 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:10:19.402 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:10:19.402 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:12:16.021 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 18:12:16.852 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:14:12.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.67s (1.93min) +2026-06-19 18:14:13.460 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:16:12.691 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:23:13.823 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 72, train_steps: 2880 +2026-06-19 18:23:14.506 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 18:23:14.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:23:14.765 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:24:19.335 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:24:19.336 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:26:14.711 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 18:26:15.694 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:28:10.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.45s (1.91min) +2026-06-19 18:28:11.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:30:10.988 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:37:12.782 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 73, train_steps: 2920 +2026-06-19 18:37:13.456 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 18:37:13.658 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:37:13.798 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:38:18.409 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:38:18.410 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:40:13.249 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 18:40:14.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:42:09.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.16s (1.92min) +2026-06-19 18:42:10.284 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:44:10.444 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:51:12.745 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 74, train_steps: 2960 +2026-06-19 18:51:13.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 18:51:13.642 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:51:13.744 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:52:18.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:52:18.233 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:54:13.982 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 18:54:14.806 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:56:10.312 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.21s (1.92min) +2026-06-19 18:56:11.031 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:58:11.441 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:05:14.078 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 75, train_steps: 3000 +2026-06-19 19:06:23.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 19:06:23.414 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:06:23.526 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:07:28.306 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:07:28.310 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:09:24.209 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 19:09:25.185 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:11:18.397 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.14s (1.89min) +2026-06-19 19:11:19.100 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:13:16.252 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:20:16.312 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 76, train_steps: 3040 +2026-06-19 19:20:16.989 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 19:20:17.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:20:17.165 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:21:21.734 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:21:21.734 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:23:18.455 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 19:23:19.614 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:25:13.441 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.66s (1.89min) +2026-06-19 19:25:14.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:27:12.734 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:34:13.268 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 77, train_steps: 3080 +2026-06-19 19:34:13.950 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 19:34:14.126 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:34:14.209 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:35:18.507 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:35:18.510 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:37:13.073 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 19:37:13.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:39:08.734 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.60s (1.91min) +2026-06-19 19:39:09.441 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:41:09.769 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:48:12.635 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 78, train_steps: 3120 +2026-06-19 19:48:13.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 19:48:13.537 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:48:13.642 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:49:18.258 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:49:18.258 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:51:13.323 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 19:51:14.157 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:53:09.267 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-19 19:53:09.975 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:55:09.984 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:02:13.174 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 79, train_steps: 3160 +2026-06-19 20:02:13.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 20:02:14.050 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:02:14.127 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:03:18.528 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:03:18.529 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:05:13.092 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 20:05:13.925 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:07:09.660 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.53s (1.93min) +2026-06-19 20:07:10.385 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:09:11.032 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:16:13.934 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 80, train_steps: 3200 +2026-06-19 20:16:51.327 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=80 (missing=0, unexpected=0) +2026-06-19 20:16:53.092 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 20:16:53.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:16:53.264 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:18:00.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:18:00.285 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:19:54.752 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 20:19:55.731 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:21:49.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-19 20:21:50.363 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:23:47.271 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:30:47.432 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 81, train_steps: 3240 +2026-06-19 20:30:48.143 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 20:30:48.238 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:30:48.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:31:53.131 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:31:53.131 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:33:49.273 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 20:33:50.266 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:35:44.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.70s (1.89min) +2026-06-19 20:35:44.752 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:37:41.678 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:44:39.997 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 82, train_steps: 3280 +2026-06-19 20:44:40.696 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 20:44:40.829 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:44:40.864 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:45:46.571 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:45:46.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:47:44.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 20:47:44.992 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:49:38.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.41s (1.89min) +2026-06-19 20:49:39.275 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:51:38.423 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:58:38.984 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 83, train_steps: 3320 +2026-06-19 20:58:39.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 20:58:39.994 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:58:40.076 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:59:45.008 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:59:45.009 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:01:44.303 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 21:01:45.132 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:03:38.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.28s (1.89min) +2026-06-19 21:03:39.339 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:05:38.818 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:12:38.918 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 84, train_steps: 3360 +2026-06-19 21:12:39.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 21:12:39.790 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:12:39.838 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:13:45.254 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:13:45.258 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:15:43.791 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 21:15:44.611 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:17:38.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.31s (1.89min) +2026-06-19 21:17:38.863 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:19:38.523 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:26:39.495 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 85, train_steps: 3400 +2026-06-19 21:26:40.167 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 21:26:40.288 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:26:40.318 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:27:45.802 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:27:45.806 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:29:43.697 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 21:29:44.665 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:31:38.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-19 21:31:38.988 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:33:38.756 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:40:38.642 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 86, train_steps: 3440 +2026-06-19 21:40:39.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 21:40:39.474 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:40:39.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:41:44.887 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:41:44.887 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:43:42.867 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 21:43:43.841 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:45:37.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 21:45:37.987 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:47:37.709 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:54:39.041 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 87, train_steps: 3480 +2026-06-19 21:54:39.742 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 21:54:39.856 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:54:39.912 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:55:45.227 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:55:45.227 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:57:43.169 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 21:57:44.133 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:59:37.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.39s (1.89min) +2026-06-19 21:59:38.317 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:01:38.555 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:08:39.976 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 88, train_steps: 3520 +2026-06-19 22:08:40.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 22:08:40.868 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:08:40.997 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:09:46.109 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:09:46.113 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:11:43.755 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 22:11:44.582 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:13:38.389 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.57s (1.89min) +2026-06-19 22:13:39.103 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:15:38.304 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:22:38.691 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 89, train_steps: 3560 +2026-06-19 22:22:39.424 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 22:22:39.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:22:39.685 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:23:44.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:23:44.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:25:43.683 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 22:25:44.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:27:38.162 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 22:27:38.886 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:29:38.494 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:36:39.410 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 90, train_steps: 3600 +2026-06-19 22:37:15.302 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=90 (missing=0, unexpected=0) +2026-06-19 22:37:17.307 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 22:37:17.392 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:37:17.508 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:38:23.202 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:38:23.205 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:40:19.240 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 22:40:20.094 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:42:14.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.70s (1.89min) +2026-06-19 22:42:14.863 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:44:13.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:51:14.846 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 91, train_steps: 3640 +2026-06-19 22:51:15.700 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 22:51:15.894 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:51:15.979 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:52:20.512 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:52:20.515 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:54:15.937 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 22:54:16.767 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:56:12.671 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.67s (1.93min) +2026-06-19 22:56:13.388 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:58:13.112 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:05:14.257 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 92, train_steps: 3680 +2026-06-19 23:05:15.118 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 23:05:15.285 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:05:15.382 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:06:19.690 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:06:19.693 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:08:16.058 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 23:08:16.916 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:10:11.751 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.62s (1.91min) +2026-06-19 23:10:12.471 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:12:11.281 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:19:13.477 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 93, train_steps: 3720 +2026-06-19 23:19:14.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 23:19:14.288 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:19:14.318 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:20:19.070 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:20:19.070 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:22:14.354 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 23:22:15.376 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:24:10.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.33s (1.92min) +2026-06-19 23:24:11.707 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:26:11.284 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:33:13.939 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 94, train_steps: 3760 +2026-06-19 23:33:14.640 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 23:33:14.822 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:33:14.950 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:34:19.154 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:34:19.157 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:36:16.135 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 23:36:17.100 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:38:12.062 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.89s (1.91min) +2026-06-19 23:38:12.784 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:40:11.321 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:47:11.936 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 95, train_steps: 3800 +2026-06-19 23:47:12.588 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-19 23:47:12.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:47:12.750 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:48:17.248 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:48:17.251 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:50:12.004 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-19 23:50:13.177 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:52:08.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.82s (1.91min) +2026-06-19 23:52:08.791 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:54:07.677 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:01:09.771 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 96, train_steps: 3840 +2026-06-20 00:01:10.475 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 00:01:10.641 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:01:10.834 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:02:15.408 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:02:15.412 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:04:11.783 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 00:04:12.922 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:06:07.927 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 00:06:08.644 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:08:07.689 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:15:08.500 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 97, train_steps: 3880 +2026-06-20 00:15:09.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 00:15:09.368 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:15:09.422 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:16:13.969 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:16:13.972 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:18:11.237 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 00:18:12.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:20:07.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.20s (1.92min) +2026-06-20 00:20:08.400 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:22:07.126 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:29:07.005 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 98, train_steps: 3920 +2026-06-20 00:29:07.692 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 00:29:07.877 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:29:07.978 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:30:12.385 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:30:12.392 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:32:10.419 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 00:32:11.387 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:34:06.389 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 00:34:07.118 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:36:03.851 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:43:05.342 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 99, train_steps: 3960 +2026-06-20 00:43:06.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 00:43:06.154 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:43:06.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:44:10.730 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:44:10.735 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:46:09.899 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 00:46:10.895 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:48:05.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.75s (1.91min) +2026-06-20 00:48:06.440 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:50:03.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:57:04.800 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 100, train_steps: 4000 +2026-06-20 00:58:47.835 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=100 (missing=0, unexpected=0) +2026-06-20 00:58:49.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 00:58:49.704 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:58:49.755 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:59:55.680 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:59:55.683 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:01:50.303 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 01:01:51.288 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:03:44.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-20 01:03:45.434 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:05:42.219 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:12:39.860 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 101, train_steps: 4040 +2026-06-20 01:12:40.567 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 01:12:40.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:12:40.812 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:13:47.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:13:47.114 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:15:43.499 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 01:15:44.485 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:17:38.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.62s (1.89min) +2026-06-20 01:17:38.931 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:19:38.325 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:26:37.801 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 102, train_steps: 4080 +2026-06-20 01:26:38.504 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 01:26:38.659 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:26:38.727 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:27:43.360 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:27:43.364 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:29:41.183 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 01:29:42.450 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:31:36.383 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-20 01:31:37.097 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:33:37.225 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:40:38.536 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 103, train_steps: 4120 +2026-06-20 01:40:39.244 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 01:40:39.415 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:40:39.499 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:41:44.795 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:41:44.798 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:43:43.334 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 01:43:44.329 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:45:37.767 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-20 01:45:38.487 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:47:38.267 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:54:37.783 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 104, train_steps: 4160 +2026-06-20 01:54:38.702 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 01:54:38.808 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:54:38.842 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:55:43.698 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:55:43.701 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:57:42.889 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 01:57:43.870 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:59:37.314 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-20 01:59:38.031 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:01:38.020 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:08:38.447 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 105, train_steps: 4200 +2026-06-20 02:08:39.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 02:08:39.346 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:08:39.460 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:09:44.431 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:09:44.433 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:11:43.945 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 02:11:44.916 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:13:38.217 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 02:13:38.924 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:15:38.269 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:22:37.840 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 106, train_steps: 4240 +2026-06-20 02:22:38.566 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 02:22:38.763 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:22:38.885 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:23:42.985 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:23:42.989 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:25:40.846 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 02:25:41.690 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:27:35.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.90s (1.90min) +2026-06-20 02:27:36.583 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:29:36.811 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:36:37.086 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 107, train_steps: 4280 +2026-06-20 02:36:37.802 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 02:36:37.998 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:36:38.149 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:37:42.496 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:37:42.498 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:39:39.229 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 02:39:40.406 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:41:35.347 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 02:41:36.072 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:43:35.145 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:50:34.994 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 108, train_steps: 4320 +2026-06-20 02:50:35.727 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 02:50:35.903 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:50:36.003 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:51:40.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:51:40.228 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:53:37.573 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 02:53:38.409 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:55:33.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.45s (1.91min) +2026-06-20 02:55:33.972 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:57:30.647 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:04:30.696 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 109, train_steps: 4360 +2026-06-20 03:04:31.406 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 03:04:31.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:04:31.719 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:05:36.068 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:05:36.070 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:07:33.406 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 03:07:34.222 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:09:29.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-20 03:09:30.014 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:11:26.575 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:18:25.324 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 110, train_steps: 4400 +2026-06-20 03:19:01.158 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=110 (missing=0, unexpected=0) +2026-06-20 03:19:03.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 03:19:03.409 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:19:03.426 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:20:10.218 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:20:10.222 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:22:07.523 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 03:22:08.347 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:24:03.966 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.41s (1.92min) +2026-06-20 03:24:04.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:26:02.472 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:33:05.391 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 111, train_steps: 4440 +2026-06-20 03:33:06.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 03:33:06.236 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:33:06.311 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:34:10.865 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:34:10.868 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:36:10.098 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 03:36:10.949 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:38:06.021 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 03:38:06.732 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:40:04.482 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:47:04.303 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 112, train_steps: 4480 +2026-06-20 03:47:05.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 03:47:05.229 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:47:05.328 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:48:09.397 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:48:09.401 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:50:04.983 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 03:50:06.085 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:52:00.943 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-20 03:52:01.657 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:54:00.965 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:01:04.071 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 113, train_steps: 4520 +2026-06-20 04:01:04.876 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 04:01:05.093 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:01:05.186 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:02:09.534 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:02:09.538 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:04:06.611 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 04:04:07.590 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:06:02.814 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.15s (1.92min) +2026-06-20 04:06:03.532 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:08:01.586 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:15:03.852 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 114, train_steps: 4560 +2026-06-20 04:15:04.473 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 04:15:04.588 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:15:04.634 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:16:09.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:16:09.195 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:18:04.152 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 04:18:05.205 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:20:00.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.72s (1.91min) +2026-06-20 04:20:00.895 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:22:00.099 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:29:01.463 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 115, train_steps: 4600 +2026-06-20 04:29:02.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 04:29:02.418 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:29:02.542 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:30:06.678 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:30:06.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:32:02.494 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 04:32:03.350 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:33:58.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.33s (1.91min) +2026-06-20 04:33:58.775 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:35:55.625 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:42:56.459 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 116, train_steps: 4640 +2026-06-20 04:42:57.160 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 04:42:57.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:42:57.487 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:44:01.929 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:44:01.932 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:45:58.380 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 04:45:59.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:47:53.698 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.25s (1.90min) +2026-06-20 04:47:54.407 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:49:50.889 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:56:50.228 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 117, train_steps: 4680 +2026-06-20 04:56:50.953 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 04:56:51.083 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:56:51.106 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:57:55.640 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:57:55.643 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:59:51.798 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 04:59:52.782 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:01:46.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.49s (1.89min) +2026-06-20 05:01:47.079 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:03:44.145 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:10:42.588 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 118, train_steps: 4720 +2026-06-20 05:10:43.296 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 05:10:43.489 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:10:43.586 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:11:48.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:11:48.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:13:46.957 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 05:13:47.941 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:15:41.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.26s (1.89min) +2026-06-20 05:15:42.001 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:17:38.996 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:24:37.898 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 119, train_steps: 4760 +2026-06-20 05:24:38.622 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 05:24:38.714 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:24:38.730 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:25:43.456 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:25:43.459 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:27:40.119 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 05:27:41.244 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:29:34.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-20 05:29:35.559 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:31:33.105 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:38:32.145 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 120, train_steps: 4800 +2026-06-20 05:39:09.057 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=120 (missing=0, unexpected=0) +2026-06-20 05:39:10.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 05:39:11.011 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:39:11.042 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:40:17.402 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:40:17.405 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:42:12.566 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 05:42:13.546 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:44:08.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.96s (1.92min) +2026-06-20 05:44:09.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:46:10.087 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:53:12.061 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 121, train_steps: 4840 +2026-06-20 05:53:12.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 05:53:12.917 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:53:12.951 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:54:17.775 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:54:17.778 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:56:13.155 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 05:56:14.039 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:58:09.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.97s (1.92min) +2026-06-20 05:58:10.076 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:00:10.574 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:07:11.951 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 122, train_steps: 4880 +2026-06-20 06:07:12.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 06:07:12.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:07:12.895 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:08:17.197 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:08:17.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:10:13.195 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 06:10:14.017 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:12:08.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.25s (1.90min) +2026-06-20 06:12:09.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:14:10.579 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:21:11.858 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 123, train_steps: 4920 +2026-06-20 06:21:12.561 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 06:21:12.700 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:21:12.742 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:22:16.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:22:16.960 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:24:12.071 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 06:24:13.093 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:26:08.788 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.47s (1.92min) +2026-06-20 06:26:09.511 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:28:09.833 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:35:11.844 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 124, train_steps: 4960 +2026-06-20 06:35:12.538 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 06:35:12.718 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:35:12.810 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:36:16.893 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:36:16.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:38:12.129 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 06:38:13.050 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:40:08.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.44s (1.92min) +2026-06-20 06:40:09.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:42:09.772 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:49:11.639 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 125, train_steps: 5000 +2026-06-20 06:50:15.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 06:50:15.361 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:50:15.452 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:51:19.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:51:19.979 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:53:15.926 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 06:53:16.897 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:55:11.653 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.68s (1.91min) +2026-06-20 06:55:12.372 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:57:11.763 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:04:11.917 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 126, train_steps: 5040 +2026-06-20 07:04:12.629 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 07:04:12.722 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:04:12.755 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:05:17.448 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:05:17.451 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:07:13.988 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 07:07:15.193 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:09:10.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.08s (1.92min) +2026-06-20 07:09:11.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:11:11.166 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:18:11.925 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 127, train_steps: 5080 +2026-06-20 07:18:12.630 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 07:18:12.825 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:18:12.941 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:19:17.270 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:19:17.271 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:21:13.050 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 07:21:13.938 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:23:09.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.50s (1.93min) +2026-06-20 07:23:10.327 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:25:10.157 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:32:11.165 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 128, train_steps: 5120 +2026-06-20 07:32:11.883 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 07:32:12.042 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:32:12.120 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:33:16.324 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:33:16.327 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:35:13.016 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 07:35:13.859 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:37:08.641 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.56s (1.91min) +2026-06-20 07:37:09.373 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:39:09.930 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:46:11.987 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 129, train_steps: 5160 +2026-06-20 07:46:12.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 07:46:12.843 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:46:12.885 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:47:17.860 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:47:17.863 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:49:13.070 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 07:49:13.899 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:51:09.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-20 07:51:09.758 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:53:10.303 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:00:11.780 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 130, train_steps: 5200 +2026-06-20 08:00:49.143 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=130 (missing=0, unexpected=0) +2026-06-20 08:00:51.244 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 08:00:51.458 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:00:51.545 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:01:58.582 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:01:58.585 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:03:52.509 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 08:03:53.325 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:05:47.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.72s (1.90min) +2026-06-20 08:05:48.018 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:07:44.844 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:14:45.313 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 131, train_steps: 5240 +2026-06-20 08:14:46.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 08:14:46.152 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:14:46.191 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:15:50.657 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:15:50.660 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:17:46.989 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 08:17:47.992 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:19:41.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.38s (1.89min) +2026-06-20 08:19:42.138 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:21:39.466 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:28:37.825 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 132, train_steps: 5280 +2026-06-20 08:28:38.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 08:28:38.644 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:28:38.678 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:29:43.680 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:29:43.682 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:31:40.849 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 08:31:41.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:33:35.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.44s (1.89min) +2026-06-20 08:33:36.119 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:35:35.263 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:42:34.340 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 133, train_steps: 5320 +2026-06-20 08:42:35.017 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 08:42:35.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:42:35.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:43:40.026 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:43:40.028 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:45:35.887 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 08:45:36.865 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:47:32.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.77s (1.93min) +2026-06-20 08:47:33.450 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:49:30.692 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:56:29.429 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 134, train_steps: 5360 +2026-06-20 08:56:30.115 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 08:56:30.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:56:30.330 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:57:34.792 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:57:34.795 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:59:32.040 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 08:59:32.866 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:01:26.319 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 09:01:27.048 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:03:23.804 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:10:22.979 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 135, train_steps: 5400 +2026-06-20 09:10:23.673 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 09:10:23.784 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:10:23.826 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:11:28.340 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:11:28.343 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:13:23.932 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 09:13:24.766 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:15:18.588 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.61s (1.89min) +2026-06-20 09:15:19.310 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:17:17.099 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:24:16.790 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 136, train_steps: 5440 +2026-06-20 09:24:17.532 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 09:24:17.664 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:24:17.707 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:25:22.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:25:22.373 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:27:18.656 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 09:27:19.671 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:29:12.958 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.15s (1.89min) +2026-06-20 09:29:13.686 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:31:13.177 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:38:13.221 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 137, train_steps: 5480 +2026-06-20 09:38:13.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 09:38:14.037 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:38:14.077 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:39:18.677 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:39:18.685 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:41:14.371 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 09:41:15.412 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:43:09.522 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-20 09:43:10.251 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:45:10.503 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:52:12.142 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 138, train_steps: 5520 +2026-06-20 09:52:12.873 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 09:52:13.062 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:52:13.159 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:53:17.640 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:53:17.643 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:55:13.383 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 09:55:14.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:57:10.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.46s (1.92min) +2026-06-20 09:57:10.921 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:59:11.151 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:06:12.677 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 139, train_steps: 5560 +2026-06-20 10:06:13.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 10:06:13.573 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:06:13.664 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:07:18.123 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:07:18.127 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:09:13.264 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 10:09:14.250 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:11:08.675 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.20s (1.90min) +2026-06-20 10:11:09.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:13:09.781 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:20:12.684 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 140, train_steps: 5600 +2026-06-20 10:20:50.172 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=140 (missing=0, unexpected=0) +2026-06-20 10:20:51.812 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 10:20:51.944 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:20:51.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:21:58.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:21:58.131 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:23:53.040 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 10:23:53.798 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:25:47.600 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-20 10:25:48.313 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:27:45.364 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:34:44.000 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 141, train_steps: 5640 +2026-06-20 10:34:44.688 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 10:34:44.871 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:34:44.944 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:35:49.363 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:35:49.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:37:46.846 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 10:37:47.842 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:39:41.415 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.51s (1.89min) +2026-06-20 10:39:42.148 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:41:39.464 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:48:38.461 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 142, train_steps: 5680 +2026-06-20 10:48:39.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 10:48:39.262 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:48:39.300 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:49:45.015 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:49:45.015 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:51:43.153 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 10:51:44.057 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:53:37.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-20 10:53:37.851 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:55:37.993 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:02:37.909 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 143, train_steps: 5720 +2026-06-20 11:02:38.630 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 11:02:38.758 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:02:38.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:03:43.823 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:03:43.826 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:05:42.169 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 11:05:43.329 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:07:36.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 11:07:37.313 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:09:36.842 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:16:37.092 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 144, train_steps: 5760 +2026-06-20 11:16:37.801 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 11:16:38.003 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:16:38.123 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:17:42.475 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:17:42.478 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:19:38.707 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 11:19:39.825 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:21:33.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.47s (1.89min) +2026-06-20 11:21:34.087 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:23:30.886 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:30:30.395 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 145, train_steps: 5800 +2026-06-20 11:30:31.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 11:30:31.289 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:30:31.392 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:31:35.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:31:35.617 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:33:33.207 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 11:33:34.042 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:35:28.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.89s (1.90min) +2026-06-20 11:35:28.900 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:37:25.683 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:44:23.908 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 146, train_steps: 5840 +2026-06-20 11:44:24.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 11:44:24.747 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:44:24.790 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:45:29.411 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:45:29.414 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:47:23.247 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 11:47:24.080 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:49:19.133 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-20 11:49:19.856 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:51:17.039 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:58:15.914 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 147, train_steps: 5880 +2026-06-20 11:58:16.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 11:58:16.744 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:58:16.788 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:59:21.750 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:59:21.790 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:01:16.232 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 12:01:17.080 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:03:12.586 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.29s (1.92min) +2026-06-20 12:03:13.294 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:05:12.904 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:12:12.704 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 148, train_steps: 5920 +2026-06-20 12:12:13.405 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 12:12:13.564 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:12:13.634 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:13:18.263 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:13:18.266 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:15:13.144 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 12:15:14.139 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:17:09.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.69s (1.91min) +2026-06-20 12:17:09.789 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:19:10.588 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:26:11.399 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 149, train_steps: 5960 +2026-06-20 12:26:12.136 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 12:26:12.333 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:26:12.445 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:27:16.947 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:27:16.950 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:29:13.029 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 12:29:14.028 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:31:08.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.72s (1.91min) +2026-06-20 12:31:09.532 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:33:09.830 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:40:10.984 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 150, train_steps: 6000 +2026-06-20 12:41:53.757 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=150 (missing=0, unexpected=0) +2026-06-20 12:41:56.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 12:41:56.235 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:41:56.251 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:43:02.375 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:43:02.377 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:44:59.743 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 12:45:00.749 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:46:55.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.70s (1.91min) +2026-06-20 12:46:56.236 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:48:52.989 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:55:51.599 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 151, train_steps: 6040 +2026-06-20 12:55:52.305 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 12:55:52.401 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:55:52.448 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:56:56.974 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:56:56.977 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:58:50.721 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 12:58:51.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:00:46.189 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.41s (1.91min) +2026-06-20 13:00:46.908 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:02:43.653 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:09:41.898 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 152, train_steps: 6080 +2026-06-20 13:09:42.599 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 13:09:42.807 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:09:42.941 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:10:47.817 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:10:47.817 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:12:43.910 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 13:12:44.892 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:14:38.763 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-20 13:14:39.492 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:16:38.551 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:23:37.280 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 153, train_steps: 6120 +2026-06-20 13:23:38.091 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 13:23:38.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:23:38.241 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:24:43.441 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:24:43.444 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:26:40.127 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 13:26:41.289 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:28:34.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.39s (1.89min) +2026-06-20 13:28:35.473 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:30:32.908 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:37:31.693 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 154, train_steps: 6160 +2026-06-20 13:37:32.391 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 13:37:32.510 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:37:32.543 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:38:37.042 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:38:37.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:40:33.788 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 13:40:34.779 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:42:29.329 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.47s (1.91min) +2026-06-20 13:42:30.062 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:44:27.158 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:51:24.973 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 155, train_steps: 6200 +2026-06-20 13:51:25.701 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 13:51:25.856 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:51:25.910 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:52:30.060 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:52:30.063 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:54:24.447 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 13:54:25.434 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:56:19.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.63s (1.89min) +2026-06-20 13:56:19.859 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:58:16.750 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:05:15.347 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 156, train_steps: 6240 +2026-06-20 14:05:16.035 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 14:05:16.238 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:05:16.368 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:06:21.210 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:06:21.213 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:08:15.931 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 14:08:16.922 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:10:11.722 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.74s (1.91min) +2026-06-20 14:10:12.457 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:12:11.104 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:19:12.235 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 157, train_steps: 6280 +2026-06-20 14:19:12.924 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 14:19:13.102 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:19:13.184 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:20:17.666 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:20:17.669 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:22:12.398 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 14:22:13.489 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:24:08.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.22s (1.92min) +2026-06-20 14:24:09.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:26:09.905 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:33:10.830 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 158, train_steps: 6320 +2026-06-20 14:33:11.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 14:33:11.609 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:33:11.655 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:34:16.166 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:34:16.170 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:36:10.935 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 14:36:12.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:38:07.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.88s (1.91min) +2026-06-20 14:38:07.784 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:40:05.258 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:47:04.563 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 159, train_steps: 6360 +2026-06-20 14:47:05.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 14:47:05.421 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:47:05.458 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:48:09.926 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:48:09.929 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:50:04.048 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 14:50:05.212 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:51:59.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.86s (1.90min) +2026-06-20 14:51:59.929 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:53:57.323 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:00:59.172 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 160, train_steps: 6400 +2026-06-20 15:01:36.923 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=160 (missing=0, unexpected=0) +2026-06-20 15:01:39.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 15:01:39.442 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:01:39.474 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:02:46.295 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:02:46.295 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:04:42.879 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 15:04:43.707 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:06:37.408 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.48s (1.89min) +2026-06-20 15:06:38.136 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:08:38.300 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:15:37.513 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 161, train_steps: 6440 +2026-06-20 15:15:38.207 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 15:15:38.369 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:15:38.451 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:16:42.986 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:16:42.989 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:18:40.375 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 15:18:41.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:20:35.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-20 15:20:36.332 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:22:36.147 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:29:36.892 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 162, train_steps: 6480 +2026-06-20 15:29:37.571 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 15:29:37.783 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:29:37.889 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:30:42.263 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:30:42.266 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:32:37.444 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 15:32:38.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:34:33.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.48s (1.91min) +2026-06-20 15:34:33.827 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:36:31.317 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:43:31.478 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 163, train_steps: 6520 +2026-06-20 15:43:32.172 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 15:43:32.372 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:43:32.502 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:44:36.964 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:44:36.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:46:33.676 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 15:46:34.563 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:48:29.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.47s (1.91min) +2026-06-20 15:48:29.934 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:50:26.538 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:57:25.497 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 164, train_steps: 6560 +2026-06-20 15:57:26.229 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 15:57:26.418 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:57:26.525 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:58:30.521 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:58:30.525 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:00:26.401 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 16:00:27.392 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:02:21.064 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.61s (1.89min) +2026-06-20 16:02:21.793 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:04:18.344 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:11:18.368 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 165, train_steps: 6600 +2026-06-20 16:11:19.087 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 16:11:19.261 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:11:19.335 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:12:24.170 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:12:24.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:14:19.105 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 16:14:20.235 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:16:14.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.75s (1.90min) +2026-06-20 16:16:14.870 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:18:14.126 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:25:13.928 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 166, train_steps: 6640 +2026-06-20 16:25:14.621 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 16:25:14.797 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:25:14.898 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:26:19.597 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:26:19.600 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:28:13.831 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 16:28:14.937 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:30:08.844 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.78s (1.90min) +2026-06-20 16:30:09.574 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:32:10.307 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:39:12.154 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 167, train_steps: 6680 +2026-06-20 16:39:12.848 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 16:39:12.961 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:39:12.998 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:40:17.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:40:17.602 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:42:12.094 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 16:42:13.323 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:44:08.745 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.36s (1.92min) +2026-06-20 16:44:09.477 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:46:10.020 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:53:11.384 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 168, train_steps: 6720 +2026-06-20 16:53:12.088 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 16:53:12.220 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:53:12.262 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:54:16.895 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:54:16.898 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:56:11.767 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 16:56:12.771 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:58:07.736 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-20 16:58:08.464 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:00:07.578 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:07:09.243 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 169, train_steps: 6760 +2026-06-20 17:07:09.876 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 17:07:09.983 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:07:10.041 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:08:14.946 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:08:14.948 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:10:11.903 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 17:10:12.889 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:12:08.254 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.20s (1.92min) +2026-06-20 17:12:08.969 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:14:09.320 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:21:11.048 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 170, train_steps: 6800 +2026-06-20 17:21:47.488 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=170 (missing=0, unexpected=0) +2026-06-20 17:21:49.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 17:21:49.414 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:21:49.531 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:22:54.826 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:22:54.829 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:24:49.925 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 17:24:50.701 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:26:44.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.05s (1.88min) +2026-06-20 17:26:44.743 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:28:41.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:35:39.738 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 37, sp_rank 0: update step: 171, train_steps: 6840 +2026-06-20 17:35:40.431 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 37, dp_rank 18: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=0 +2026-06-20 17:35:40.591 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:35:40.798 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:36:46.161 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 37 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:36:46.164 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 37 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:38:42.536 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 37: Using local metric names without global sync: ['avg'] +2026-06-20 17:38:43.553 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 37: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:40:36.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 37: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.89s (1.88min) +2026-06-20 17:40:37.378 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 37, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-29 01:27:08.632 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 01:27:08.632 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 01:51:25.496 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 01:51:25.497 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 02:20:16.165 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 02:20:16.166 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 02:46:50.659 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 02:46:50.661 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:06:52.596 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:06:52.597 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:31:13.912 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:31:13.912 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:57:53.990 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:57:53.990 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:25:32.756 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:25:32.756 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:53:31.859 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:53:31.860 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 05:14:51.274 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 05:14:51.275 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 05:39:01.357 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 05:39:01.359 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:02:15.282 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:02:15.284 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:25:05.512 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:25:05.513 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:48:17.455 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:48:17.455 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 07:11:56.909 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 07:11:56.909 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 07:33:36.083 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 07:33:36.084 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:07:34.200 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:07:34.201 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:29:39.779 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:29:39.780 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:58:44.499 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:58:44.499 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:22:10.874 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:22:10.878 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:47:13.719 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:47:13.720 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 10:10:21.196 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 10:10:21.197 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 10:35:25.255 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 10:35:25.256 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:02:55.895 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:02:55.896 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:26:43.833 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:26:43.834 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:54:48.832 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:54:48.832 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 12:17:59.535 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 12:17:59.536 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 12:46:12.538 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 12:46:12.541 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 13:12:11.198 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 13:12:11.198 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 13:36:41.867 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 13:36:41.868 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:03:52.664 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:03:52.665 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:26:16.840 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:26:16.840 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:49:03.202 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:49:03.204 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 15:21:24.224 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 15:21:24.224 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 15:46:20.467 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 15:46:20.468 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 16:15:56.631 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 16:15:56.632 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 16:40:27.995 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 16:40:27.995 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:09:23.253 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:09:23.255 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:32:40.901 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:32:40.903 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:53:11.837 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:53:11.838 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 18:16:51.885 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 18:16:51.887 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 18:38:51.345 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 18:38:51.346 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_39.log b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_39.log new file mode 100644 index 0000000000000000000000000000000000000000..b58eb1fdc258a5abf036edeeeb6660e9521c7f88 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_39.log @@ -0,0 +1,1994 @@ +2026-06-19 01:23:02.950 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-19 01:23:02.953 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-19 01:30:17.294 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 01:30:17.391 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:30:17.393 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:31:52.591 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:31:52.593 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:33:45.514 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 01:34:00.994 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 01:35:56.996 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.94s (1.93min) +2026-06-19 01:35:57.692 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 01:37:54.190 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 01:37:57.219 | INFO | hyvideo.utils.grpo_utils:train_one_step:3127 - Rank 39, sp_rank 0: init adaptive KL controller (init=1.000e-07, min=1.000e-07, max=1.000e-05) +2026-06-19 01:45:11.457 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 1, train_steps: 40 +2026-06-19 01:45:12.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 01:45:12.701 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:45:12.739 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:46:17.168 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:46:17.170 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:48:10.944 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 01:48:11.830 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 01:50:07.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.70s (1.93min) +2026-06-19 01:50:08.465 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 01:52:06.954 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 01:59:06.750 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 2, train_steps: 80 +2026-06-19 01:59:07.381 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 01:59:07.476 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:59:07.492 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:00:11.699 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:00:11.702 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:02:10.528 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 02:02:11.500 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:04:06.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.88s (1.91min) +2026-06-19 02:04:07.374 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:06:03.753 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:13:04.845 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 3, train_steps: 120 +2026-06-19 02:13:05.469 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 02:13:05.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:13:05.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:14:09.745 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:14:09.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:16:04.655 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 02:16:05.658 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:17:59.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.77s (1.90min) +2026-06-19 02:18:00.448 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:19:59.349 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:26:59.873 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 4, train_steps: 160 +2026-06-19 02:27:00.548 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 02:27:00.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:27:00.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:28:05.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:28:05.310 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:30:00.316 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 02:30:01.094 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:31:55.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.62s (1.91min) +2026-06-19 02:31:56.651 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:33:53.223 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:40:53.231 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 5, train_steps: 200 +2026-06-19 02:40:53.901 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 02:40:54.061 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:40:54.149 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:41:58.557 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:41:58.560 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:43:51.811 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 02:43:52.592 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:45:46.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.50s (1.89min) +2026-06-19 02:45:46.993 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:47:43.353 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:54:40.648 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 6, train_steps: 240 +2026-06-19 02:54:41.360 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 02:54:41.461 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:54:41.495 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:55:47.002 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:55:47.006 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:57:42.949 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 02:57:43.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:59:37.353 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.38s (1.89min) +2026-06-19 02:59:38.043 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:01:37.729 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:08:37.467 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 7, train_steps: 280 +2026-06-19 03:08:38.148 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 03:08:38.337 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:08:38.458 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:09:42.555 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:09:42.558 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:11:38.337 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 03:11:39.501 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:13:34.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.68s (1.91min) +2026-06-19 03:13:34.976 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:15:32.041 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:22:31.154 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 8, train_steps: 320 +2026-06-19 03:22:31.814 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 03:22:31.942 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:22:31.973 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:23:37.049 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:23:37.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:25:33.100 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 03:25:33.890 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:27:28.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-19 03:27:29.432 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:29:26.903 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:36:25.638 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 9, train_steps: 360 +2026-06-19 03:36:26.347 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 03:36:26.537 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:36:26.636 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:37:31.053 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:37:31.055 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:39:26.010 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 03:39:26.800 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:41:20.399 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.36s (1.89min) +2026-06-19 03:41:21.105 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:43:17.558 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:50:16.922 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 10, train_steps: 400 +2026-06-19 03:50:53.153 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=10 (missing=0, unexpected=0) +2026-06-19 03:50:56.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 03:50:56.302 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:50:56.322 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:52:02.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:52:02.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:53:57.828 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 03:53:58.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:55:52.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.65s (1.89min) +2026-06-19 03:55:53.209 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:57:49.763 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:04:48.374 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 11, train_steps: 440 +2026-06-19 04:04:49.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 04:04:49.220 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:04:49.271 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:05:53.927 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:05:53.931 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:07:48.852 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 04:07:49.800 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:09:43.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.25s (1.89min) +2026-06-19 04:09:43.797 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:11:40.269 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:18:37.290 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 12, train_steps: 480 +2026-06-19 04:18:37.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 04:18:38.107 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:18:38.142 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:19:42.787 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:19:42.790 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:21:37.180 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 04:21:38.131 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:23:33.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.19s (1.92min) +2026-06-19 04:23:34.070 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:25:31.243 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:32:29.814 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 13, train_steps: 520 +2026-06-19 04:32:30.513 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 04:32:30.709 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:32:30.836 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:33:35.084 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:33:35.084 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:35:30.312 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 04:35:31.251 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:37:24.102 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.77s (1.88min) +2026-06-19 04:37:24.791 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:39:21.077 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:46:18.589 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 14, train_steps: 560 +2026-06-19 04:46:19.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 04:46:19.336 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:46:19.378 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:47:24.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:47:24.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:49:17.445 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 04:49:18.393 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:51:12.963 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.49s (1.91min) +2026-06-19 04:51:13.650 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:53:12.995 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:00:11.833 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 15, train_steps: 600 +2026-06-19 05:00:12.525 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 05:00:12.688 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:00:12.761 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:01:17.386 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:01:17.389 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:03:11.349 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 05:03:12.607 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:05:07.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.64s (1.91min) +2026-06-19 05:05:08.061 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:07:05.976 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:14:04.896 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 16, train_steps: 640 +2026-06-19 05:14:05.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 05:14:05.784 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:14:05.909 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:15:10.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:15:10.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:17:06.988 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 05:17:07.931 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:19:03.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.02s (1.92min) +2026-06-19 05:19:03.742 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:21:01.961 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:28:03.324 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 17, train_steps: 680 +2026-06-19 05:28:04.007 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 05:28:04.203 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:28:04.301 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:29:08.768 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:29:08.771 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:31:03.926 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 05:31:04.887 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:32:58.688 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.48s (1.89min) +2026-06-19 05:32:59.387 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:34:56.266 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:41:58.103 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 18, train_steps: 720 +2026-06-19 05:41:58.777 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 05:41:58.941 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:41:59.026 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:43:03.335 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:43:03.339 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:45:00.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 05:45:01.158 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:46:55.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.88s (1.90min) +2026-06-19 05:46:55.797 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:48:52.338 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:55:50.491 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 19, train_steps: 760 +2026-06-19 05:55:51.178 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 05:55:51.334 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:55:51.426 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:56:55.816 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:56:55.819 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:58:49.068 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 05:58:50.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:00:43.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.10s (1.89min) +2026-06-19 06:00:43.876 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:02:40.342 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:09:39.758 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 20, train_steps: 800 +2026-06-19 06:10:16.388 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=20 (missing=0, unexpected=0) +2026-06-19 06:10:18.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 06:10:19.177 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:10:19.196 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:11:24.754 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:11:24.754 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:13:19.050 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 06:13:20.124 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:15:13.224 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.90s (1.88min) +2026-06-19 06:15:13.910 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:17:12.688 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:24:12.556 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 21, train_steps: 840 +2026-06-19 06:24:13.271 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 06:24:13.445 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:24:13.524 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:25:17.923 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:25:17.926 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:27:11.260 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 06:27:12.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:29:06.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.51s (1.91min) +2026-06-19 06:29:07.668 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:31:04.578 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:38:04.278 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 22, train_steps: 880 +2026-06-19 06:38:04.947 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 06:38:05.164 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:38:05.306 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:39:09.845 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:39:09.848 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:41:03.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 06:41:04.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:42:58.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.99s (1.90min) +2026-06-19 06:42:59.165 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:44:56.505 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:51:58.169 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 23, train_steps: 920 +2026-06-19 06:51:58.843 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 06:51:59.017 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:51:59.122 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:53:03.328 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:53:03.331 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:54:59.789 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 06:55:00.751 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:56:54.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.11s (1.90min) +2026-06-19 06:56:55.622 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:58:52.321 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:05:51.314 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 24, train_steps: 960 +2026-06-19 07:05:52.005 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 07:05:52.191 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:05:52.306 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:06:56.641 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:06:56.646 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:08:52.217 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 07:08:53.160 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:10:46.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-19 07:10:46.871 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:12:44.155 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:19:41.971 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 25, train_steps: 1000 +2026-06-19 07:20:52.105 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 07:20:52.298 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:20:52.410 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:21:56.964 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:21:56.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:23:52.666 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 07:23:53.653 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:25:46.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.81s (1.88min) +2026-06-19 07:25:47.205 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:27:44.382 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:34:43.683 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 26, train_steps: 1040 +2026-06-19 07:34:44.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 07:34:44.468 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:34:44.502 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:35:48.833 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:35:48.838 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:37:45.443 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 07:37:46.407 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:39:39.638 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.14s (1.89min) +2026-06-19 07:39:40.348 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:41:39.073 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:48:39.667 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 27, train_steps: 1080 +2026-06-19 07:48:40.347 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 07:48:40.465 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:48:40.484 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:49:45.629 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:49:45.632 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:51:43.285 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 07:51:44.249 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:53:38.298 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.99s (1.90min) +2026-06-19 07:53:39.007 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:55:38.423 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:02:38.173 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 28, train_steps: 1120 +2026-06-19 08:02:38.880 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 08:02:39.074 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:02:39.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:03:43.698 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:03:43.699 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:05:41.901 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 08:05:43.042 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:07:37.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.24s (1.90min) +2026-06-19 08:07:38.092 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:09:37.933 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:16:37.840 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 29, train_steps: 1160 +2026-06-19 08:16:38.691 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 08:16:38.822 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:16:38.870 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:17:43.536 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:17:43.539 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:19:42.881 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 08:19:43.657 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:21:37.264 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 08:21:37.961 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:23:37.405 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:30:39.548 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 30, train_steps: 1200 +2026-06-19 08:31:16.063 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=30 (missing=0, unexpected=0) +2026-06-19 08:31:18.258 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 08:31:18.387 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:31:18.430 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:32:24.918 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:32:24.921 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:34:19.980 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 08:34:20.977 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:36:14.510 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-19 08:36:15.226 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:38:14.172 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:45:14.272 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 31, train_steps: 1240 +2026-06-19 08:45:14.992 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 08:45:15.123 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:45:15.159 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:46:19.706 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:46:19.706 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:48:15.288 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 08:48:16.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:50:10.322 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.84s (1.90min) +2026-06-19 08:50:11.023 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:52:11.151 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:59:13.098 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 32, train_steps: 1280 +2026-06-19 08:59:13.775 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 08:59:13.869 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:59:13.914 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:00:18.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:00:18.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:02:13.436 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 09:02:14.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:04:08.713 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.91s (1.90min) +2026-06-19 09:04:09.420 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:06:10.408 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:13:13.280 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 33, train_steps: 1320 +2026-06-19 09:13:13.952 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 09:13:14.098 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:13:14.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:14:18.705 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:14:18.709 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:16:14.822 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 09:16:15.835 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:18:11.066 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.06s (1.92min) +2026-06-19 09:18:11.785 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:20:11.627 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:27:13.792 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 34, train_steps: 1360 +2026-06-19 09:27:14.444 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 09:27:14.606 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:27:14.676 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:28:19.113 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:28:19.116 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:30:13.695 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 09:30:14.836 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:32:08.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.81s (1.90min) +2026-06-19 09:32:09.592 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:34:09.779 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:41:12.383 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 35, train_steps: 1400 +2026-06-19 09:41:13.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 09:41:13.259 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:41:13.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:42:17.817 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:42:17.820 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:44:13.080 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 09:44:13.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:46:10.686 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 116.56s (1.94min) +2026-06-19 09:46:11.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:48:10.895 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:55:11.964 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 36, train_steps: 1440 +2026-06-19 09:55:12.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 09:55:12.803 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:55:12.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:56:17.528 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:56:17.531 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:58:12.646 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 09:58:13.819 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:00:08.966 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-19 10:00:09.672 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:02:10.008 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:09:11.701 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 37, train_steps: 1480 +2026-06-19 10:09:12.397 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 10:09:12.618 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:09:12.775 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:10:17.155 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:10:17.158 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:12:13.042 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 10:12:14.010 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:14:09.029 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.94s (1.92min) +2026-06-19 10:14:09.737 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:16:10.368 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:23:12.346 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 38, train_steps: 1520 +2026-06-19 10:23:13.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 10:23:13.216 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:23:13.294 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:24:17.693 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:24:17.697 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:26:12.667 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 10:26:13.840 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:28:09.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.16s (1.92min) +2026-06-19 10:28:09.954 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:30:10.745 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:37:11.947 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 39, train_steps: 1560 +2026-06-19 10:37:12.646 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 10:37:12.794 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:37:12.846 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:38:17.479 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:38:17.482 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:41:42.526 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 10:41:43.679 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:43:36.471 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.70s (1.88min) +2026-06-19 10:43:37.180 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:45:37.176 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:52:37.855 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 40, train_steps: 1600 +2026-06-19 10:53:14.002 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=40 (missing=0, unexpected=0) +2026-06-19 10:53:15.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 10:53:15.958 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:53:16.049 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:54:22.084 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:54:22.087 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:56:16.410 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 10:56:17.386 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:58:11.717 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.27s (1.90min) +2026-06-19 10:58:12.416 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:00:11.325 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:07:12.578 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 41, train_steps: 1640 +2026-06-19 11:07:13.291 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 11:07:13.411 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:07:13.441 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:08:17.860 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:08:17.860 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:10:12.668 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 11:10:13.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:12:08.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.28s (1.90min) +2026-06-19 11:12:09.087 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:14:08.757 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:21:10.883 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 42, train_steps: 1680 +2026-06-19 11:21:11.805 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 11:21:12.016 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:21:12.165 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:22:16.667 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:22:16.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:24:12.073 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 11:24:13.190 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:26:07.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.32s (1.91min) +2026-06-19 11:26:08.278 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:28:06.796 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:35:06.771 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 43, train_steps: 1720 +2026-06-19 11:35:07.457 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 11:35:07.567 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:35:07.602 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:36:11.978 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:36:11.981 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:38:10.298 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 11:38:11.275 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:40:05.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.38s (1.91min) +2026-06-19 11:40:06.549 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:42:03.524 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:49:04.421 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 44, train_steps: 1760 +2026-06-19 11:49:05.113 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 11:49:05.247 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:49:05.294 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:50:09.798 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:50:09.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:52:06.491 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 11:52:07.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:54:01.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.69s (1.89min) +2026-06-19 11:54:01.909 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:56:01.233 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:03:03.858 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 45, train_steps: 1800 +2026-06-19 12:03:04.557 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 12:03:04.722 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:03:04.814 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:04:09.402 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:04:09.404 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:06:05.347 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 12:06:06.133 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:08:01.417 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.05s (1.92min) +2026-06-19 12:08:02.101 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:10:01.380 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:17:03.584 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 46, train_steps: 1840 +2026-06-19 12:17:04.279 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 12:17:04.431 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:17:04.486 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:18:09.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:18:09.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:20:05.333 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 12:20:06.485 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:22:01.733 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.17s (1.92min) +2026-06-19 12:22:02.442 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:24:01.451 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:31:02.923 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 47, train_steps: 1880 +2026-06-19 12:31:03.607 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 12:31:03.768 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:31:03.823 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:32:08.119 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:32:08.122 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:34:03.182 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 12:34:04.312 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:35:58.279 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-19 12:35:58.982 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:37:55.890 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:44:57.677 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 48, train_steps: 1920 +2026-06-19 12:44:58.355 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 12:44:58.563 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:44:58.706 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:46:03.213 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:46:03.214 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:47:58.930 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 12:47:59.916 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:49:54.620 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-19 12:49:55.342 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:51:52.289 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:58:52.078 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 49, train_steps: 1960 +2026-06-19 12:58:52.782 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 12:58:52.911 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:58:52.937 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:59:57.279 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:59:57.282 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:01:54.369 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 13:01:55.222 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:03:48.446 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.04s (1.88min) +2026-06-19 13:03:49.148 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:05:45.466 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:12:45.259 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 50, train_steps: 2000 +2026-06-19 13:14:21.662 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=50 (missing=0, unexpected=0) +2026-06-19 13:14:23.203 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 13:14:23.286 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:14:23.381 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:15:29.910 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:15:29.913 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:17:25.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 13:17:26.166 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:19:20.081 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.84s (1.90min) +2026-06-19 13:19:20.807 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:21:17.524 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:28:18.459 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 51, train_steps: 2040 +2026-06-19 13:28:19.185 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 13:28:19.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:28:19.489 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:29:23.733 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:29:23.738 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:31:18.028 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 13:31:19.040 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:33:13.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.83s (1.90min) +2026-06-19 13:33:13.742 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:35:13.229 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:42:13.613 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 52, train_steps: 2080 +2026-06-19 13:42:14.276 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 13:42:14.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:42:14.407 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:43:19.488 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:43:19.491 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:45:15.287 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 13:45:16.377 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:47:10.481 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.03s (1.90min) +2026-06-19 13:47:11.196 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:49:11.265 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:56:12.602 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 53, train_steps: 2120 +2026-06-19 13:56:13.303 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 13:56:13.457 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:56:13.513 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:57:18.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:57:18.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:59:13.130 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 13:59:13.983 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:01:08.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.41s (1.91min) +2026-06-19 14:01:09.529 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:03:10.472 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:10:12.153 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 54, train_steps: 2160 +2026-06-19 14:10:12.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 14:10:12.993 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:10:13.010 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:11:17.402 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:11:17.403 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:13:12.866 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 14:13:13.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:15:08.493 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.42s (1.91min) +2026-06-19 14:15:09.217 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:17:09.304 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:24:11.894 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 55, train_steps: 2200 +2026-06-19 14:24:12.602 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 14:24:12.692 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:24:12.733 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:25:17.346 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:25:17.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:27:12.753 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 14:27:13.902 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:29:09.888 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.82s (1.93min) +2026-06-19 14:29:10.604 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:31:10.763 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:38:12.119 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 56, train_steps: 2240 +2026-06-19 14:38:12.811 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 14:38:12.939 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:38:12.970 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:39:17.343 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:39:17.346 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:41:13.092 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 14:41:13.943 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:43:08.780 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.66s (1.91min) +2026-06-19 14:43:09.475 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:45:09.716 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:52:11.830 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 57, train_steps: 2280 +2026-06-19 14:52:12.533 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 14:52:12.729 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:52:12.863 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:53:17.186 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:53:17.186 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:55:13.057 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 14:55:14.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:57:08.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.79s (1.91min) +2026-06-19 14:57:09.597 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:59:09.880 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:06:12.286 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 58, train_steps: 2320 +2026-06-19 15:06:12.984 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 15:06:13.121 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:06:13.169 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:07:17.981 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:07:17.981 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:09:13.062 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 15:09:14.025 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:11:08.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.71s (1.91min) +2026-06-19 15:11:09.498 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:13:09.809 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:20:12.055 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 59, train_steps: 2360 +2026-06-19 15:20:12.738 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 15:20:12.855 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:20:12.887 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:21:17.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:21:17.580 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:23:12.372 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 15:23:13.495 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:25:09.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.28s (1.92min) +2026-06-19 15:25:09.743 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:27:10.367 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:34:12.073 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 60, train_steps: 2400 +2026-06-19 15:34:48.213 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=60 (missing=0, unexpected=0) +2026-06-19 15:34:50.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 15:34:50.759 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:34:50.818 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:35:57.137 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:35:57.140 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:37:53.529 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 15:37:54.701 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:39:47.707 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-19 15:39:48.402 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:41:45.234 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:48:44.581 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 61, train_steps: 2440 +2026-06-19 15:48:45.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 15:48:45.414 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:48:45.438 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:49:50.113 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:49:50.117 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:51:47.609 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 15:51:48.472 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:53:41.681 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.04s (1.88min) +2026-06-19 15:53:42.399 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:55:39.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:02:38.991 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 62, train_steps: 2480 +2026-06-19 16:02:39.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 16:02:39.790 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:02:39.837 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:03:45.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:03:45.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:05:44.132 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 16:05:44.928 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:07:38.761 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.59s (1.89min) +2026-06-19 16:07:39.470 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:09:38.852 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:16:39.307 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 63, train_steps: 2520 +2026-06-19 16:16:40.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 16:16:40.220 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:16:40.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:17:45.662 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:17:45.665 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:19:45.197 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 16:19:45.997 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:21:39.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.36s (1.89min) +2026-06-19 16:21:40.303 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:23:38.880 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:30:40.923 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 64, train_steps: 2560 +2026-06-19 16:30:41.625 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 16:30:41.739 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:30:41.774 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:31:47.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:31:47.318 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:33:45.471 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 16:33:46.275 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:35:40.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.63s (1.89min) +2026-06-19 16:35:40.844 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:37:39.397 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:44:41.287 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 65, train_steps: 2600 +2026-06-19 16:44:41.969 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 16:44:42.165 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:44:42.311 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:45:47.181 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:45:47.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:47:47.481 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 16:47:48.276 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:49:42.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.72s (1.90min) +2026-06-19 16:49:42.958 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:51:40.035 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:58:41.266 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 66, train_steps: 2640 +2026-06-19 16:58:41.989 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 16:58:42.084 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:58:42.116 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:59:47.222 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:59:47.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:01:43.599 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 17:01:44.411 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:03:37.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.34s (1.89min) +2026-06-19 17:03:38.681 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:05:38.380 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:12:40.113 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 67, train_steps: 2680 +2026-06-19 17:12:40.909 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 17:12:41.278 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:12:41.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:13:46.558 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:13:46.561 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:15:44.529 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 17:15:45.335 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:17:38.956 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 17:17:39.679 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:19:39.111 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:26:39.537 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 68, train_steps: 2720 +2026-06-19 17:26:40.214 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 17:26:40.350 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:26:40.379 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:27:46.142 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:27:46.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:29:43.886 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 17:29:44.701 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:31:38.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.76s (1.90min) +2026-06-19 17:31:39.420 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:33:38.460 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:40:38.357 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 69, train_steps: 2760 +2026-06-19 17:40:39.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 17:40:39.232 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:40:39.359 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:41:43.755 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:41:43.757 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:43:41.069 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 17:43:41.948 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:45:35.854 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.55s (1.89min) +2026-06-19 17:45:36.569 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:47:37.117 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:54:38.564 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 70, train_steps: 2800 +2026-06-19 17:55:14.860 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=70 (missing=0, unexpected=0) +2026-06-19 17:55:16.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 17:55:17.085 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:55:17.134 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:56:23.754 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:56:23.757 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:58:18.228 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 17:58:19.259 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:00:13.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.58s (1.89min) +2026-06-19 18:00:13.838 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:02:13.852 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:09:13.885 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 71, train_steps: 2840 +2026-06-19 18:09:14.615 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 18:09:14.761 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:09:14.797 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:10:19.395 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:10:19.398 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:12:16.021 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 18:12:16.838 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:14:12.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.67s (1.93min) +2026-06-19 18:14:13.460 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:16:12.691 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:23:13.820 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 72, train_steps: 2880 +2026-06-19 18:23:14.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 18:23:14.619 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:23:14.669 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:24:19.322 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:24:19.325 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:26:14.711 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 18:26:15.660 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:28:10.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.45s (1.91min) +2026-06-19 18:28:11.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:30:10.987 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:37:12.773 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 73, train_steps: 2920 +2026-06-19 18:37:13.455 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 18:37:13.618 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:37:13.697 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:38:18.398 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:38:18.400 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:40:13.249 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 18:40:14.235 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:42:09.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.16s (1.92min) +2026-06-19 18:42:10.284 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:44:10.444 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:51:12.750 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 74, train_steps: 2960 +2026-06-19 18:51:13.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 18:51:13.659 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:51:13.783 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:52:18.392 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:52:18.393 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:54:13.982 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 18:54:14.948 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:56:10.312 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.21s (1.92min) +2026-06-19 18:56:11.031 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:58:11.442 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:05:14.006 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 75, train_steps: 3000 +2026-06-19 19:06:23.229 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 19:06:23.334 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:06:23.388 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:07:28.229 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:07:28.232 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:09:24.209 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 19:09:25.070 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:11:18.397 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.14s (1.89min) +2026-06-19 19:11:19.100 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:13:16.253 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:20:16.311 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 76, train_steps: 3040 +2026-06-19 19:20:16.989 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 19:20:17.174 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:20:17.271 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:21:21.688 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:21:21.691 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:23:18.455 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 19:23:19.622 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:25:13.441 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.66s (1.89min) +2026-06-19 19:25:14.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:27:12.734 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:34:13.267 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 77, train_steps: 3080 +2026-06-19 19:34:13.950 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 19:34:14.084 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:34:14.118 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:35:18.532 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:35:18.535 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:37:13.073 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 19:37:13.936 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:39:08.734 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.60s (1.91min) +2026-06-19 19:39:09.441 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:41:09.768 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:48:12.637 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 78, train_steps: 3120 +2026-06-19 19:48:13.368 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 19:48:13.568 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:48:13.714 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:49:18.248 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:49:18.251 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:51:13.323 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 19:51:14.148 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:53:09.267 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-19 19:53:09.975 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:55:09.984 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:02:13.171 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 79, train_steps: 3160 +2026-06-19 20:02:13.885 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 20:02:14.062 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:02:14.165 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:03:18.446 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:03:18.450 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:05:13.092 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 20:05:13.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:07:09.660 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.53s (1.93min) +2026-06-19 20:07:10.385 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:09:11.032 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:16:13.930 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 80, train_steps: 3200 +2026-06-19 20:16:50.909 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=80 (missing=0, unexpected=0) +2026-06-19 20:16:53.092 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 20:16:53.189 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:16:53.247 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:18:00.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:18:00.195 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:19:54.752 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 20:19:55.562 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:21:49.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-19 20:21:50.363 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:23:47.269 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:30:47.419 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 81, train_steps: 3240 +2026-06-19 20:30:48.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 20:30:48.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:30:48.499 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:31:53.028 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:31:53.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:33:49.273 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 20:33:50.103 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:35:44.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.70s (1.89min) +2026-06-19 20:35:44.752 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:37:41.676 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:44:40.054 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 82, train_steps: 3280 +2026-06-19 20:44:40.696 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 20:44:40.857 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:44:40.921 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:45:46.186 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:45:46.187 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:47:44.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 20:47:44.931 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:49:38.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.41s (1.89min) +2026-06-19 20:49:39.275 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:51:38.423 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:58:38.985 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 83, train_steps: 3320 +2026-06-19 20:58:39.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 20:58:40.009 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:58:40.114 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:59:45.021 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:59:45.022 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:01:44.303 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 21:01:45.115 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:03:38.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.28s (1.89min) +2026-06-19 21:03:39.339 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:05:38.817 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:12:38.914 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 84, train_steps: 3360 +2026-06-19 21:12:39.651 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 21:12:39.852 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:12:39.977 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:13:45.334 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:13:45.334 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:15:43.791 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 21:15:44.599 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:17:38.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.31s (1.89min) +2026-06-19 21:17:38.863 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:19:38.522 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:26:39.489 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 85, train_steps: 3400 +2026-06-19 21:26:40.168 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 21:26:40.375 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:26:40.496 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:27:45.806 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:27:45.806 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:29:43.697 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 21:29:44.686 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:31:38.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-19 21:31:38.988 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:33:38.756 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:40:38.649 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 86, train_steps: 3440 +2026-06-19 21:40:39.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 21:40:39.514 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:40:39.554 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:41:44.880 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:41:44.883 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:43:42.867 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 21:43:43.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:45:37.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 21:45:37.987 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:47:37.708 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:54:39.042 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 87, train_steps: 3480 +2026-06-19 21:54:39.743 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 21:54:39.911 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:54:39.982 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:55:45.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:55:45.231 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:57:43.169 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 21:57:43.977 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:59:37.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.39s (1.89min) +2026-06-19 21:59:38.317 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:01:38.555 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:08:39.987 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 88, train_steps: 3520 +2026-06-19 22:08:40.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 22:08:40.845 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:08:40.939 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:09:46.126 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:09:46.127 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:11:43.755 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 22:11:44.626 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:13:38.389 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.57s (1.89min) +2026-06-19 22:13:39.103 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:15:38.302 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:22:38.701 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 89, train_steps: 3560 +2026-06-19 22:22:39.424 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 22:22:39.550 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:22:39.585 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:23:44.907 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:23:44.908 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:25:43.683 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 22:25:44.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:27:38.162 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 22:27:38.887 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:29:38.493 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:36:39.408 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 90, train_steps: 3600 +2026-06-19 22:37:15.466 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=90 (missing=0, unexpected=0) +2026-06-19 22:37:17.307 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 22:37:17.391 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:37:17.473 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:38:23.136 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:38:23.140 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:40:19.240 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 22:40:20.076 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:42:14.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.70s (1.89min) +2026-06-19 22:42:14.863 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:44:13.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:51:14.841 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 91, train_steps: 3640 +2026-06-19 22:51:15.699 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 22:51:15.809 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:51:15.840 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:52:20.518 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:52:20.518 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:54:15.937 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 22:54:16.756 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:56:12.671 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.67s (1.93min) +2026-06-19 22:56:13.388 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:58:13.110 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:05:14.251 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 92, train_steps: 3680 +2026-06-19 23:05:15.118 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 23:05:15.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:05:15.445 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:06:19.628 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:06:19.632 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:08:16.058 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 23:08:16.870 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:10:11.751 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.62s (1.91min) +2026-06-19 23:10:12.471 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:12:11.280 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:19:13.470 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 93, train_steps: 3720 +2026-06-19 23:19:14.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 23:19:14.320 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:19:14.404 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:20:19.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:20:19.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:22:14.354 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 23:22:15.345 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:24:10.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.33s (1.92min) +2026-06-19 23:24:11.707 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:26:11.282 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:33:13.940 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 94, train_steps: 3760 +2026-06-19 23:33:14.639 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 23:33:14.814 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:33:14.915 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:34:19.342 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:34:19.345 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:36:16.135 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 23:36:16.931 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:38:12.062 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.89s (1.91min) +2026-06-19 23:38:12.784 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:40:11.319 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:47:11.862 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 95, train_steps: 3800 +2026-06-19 23:47:12.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-19 23:47:12.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:47:12.906 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:48:17.069 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:48:17.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:50:12.004 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-19 23:50:13.018 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:52:08.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.82s (1.91min) +2026-06-19 23:52:08.791 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:54:07.675 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:01:09.767 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 96, train_steps: 3840 +2026-06-20 00:01:10.475 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 00:01:10.654 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:01:10.874 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:02:15.395 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:02:15.398 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:04:11.783 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 00:04:12.766 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:06:07.927 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 00:06:08.644 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:08:07.688 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:15:08.564 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 97, train_steps: 3880 +2026-06-20 00:15:09.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 00:15:09.414 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:15:09.527 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:16:14.047 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:16:14.047 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:18:11.237 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 00:18:12.121 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:20:07.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.20s (1.92min) +2026-06-20 00:20:08.400 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:22:07.125 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:29:06.996 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 98, train_steps: 3920 +2026-06-20 00:29:07.692 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 00:29:07.892 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:29:07.997 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:30:12.367 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:30:12.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:32:10.419 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 00:32:11.224 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:34:06.389 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 00:34:07.117 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:36:03.849 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:43:05.332 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 99, train_steps: 3960 +2026-06-20 00:43:06.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 00:43:06.185 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:43:06.265 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:44:10.637 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:44:10.640 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:46:09.899 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 00:46:10.735 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:48:05.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.75s (1.91min) +2026-06-20 00:48:06.440 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:50:03.397 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:57:04.790 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 100, train_steps: 4000 +2026-06-20 00:58:47.815 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=100 (missing=0, unexpected=0) +2026-06-20 00:58:49.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 00:58:49.738 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:58:49.813 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:59:55.684 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:59:55.685 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:01:50.303 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 01:01:51.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:03:44.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-20 01:03:45.434 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:05:42.218 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:12:39.867 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 101, train_steps: 4040 +2026-06-20 01:12:40.567 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 01:12:40.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:12:40.704 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:13:46.602 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:13:46.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:15:43.499 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 01:15:44.493 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:17:38.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.62s (1.89min) +2026-06-20 01:17:38.931 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:19:38.325 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:26:37.806 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 102, train_steps: 4080 +2026-06-20 01:26:38.504 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 01:26:38.672 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:26:38.747 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:27:43.362 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:27:43.364 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:29:41.183 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 01:29:42.357 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:31:36.383 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-20 01:31:37.097 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:33:37.224 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:40:38.543 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 103, train_steps: 4120 +2026-06-20 01:40:39.244 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 01:40:39.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:40:39.402 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:41:44.797 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:41:44.798 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:43:43.334 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 01:43:44.338 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:45:37.767 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-20 01:45:38.487 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:47:38.265 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:54:37.791 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 104, train_steps: 4160 +2026-06-20 01:54:38.702 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 01:54:38.830 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:54:38.857 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:55:43.707 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:55:43.710 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:57:42.889 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 01:57:43.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:59:37.314 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-20 01:59:38.031 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:01:38.021 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:08:38.448 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 105, train_steps: 4200 +2026-06-20 02:08:39.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 02:08:39.332 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:08:39.425 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:09:44.429 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:09:44.433 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:11:43.945 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 02:11:44.822 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:13:38.217 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 02:13:38.924 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:15:38.267 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:22:37.838 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 106, train_steps: 4240 +2026-06-20 02:22:38.564 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 02:22:38.697 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:22:38.737 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:23:43.137 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:23:43.140 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:25:40.846 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 02:25:41.677 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:27:35.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.90s (1.90min) +2026-06-20 02:27:36.583 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:29:36.811 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:36:37.084 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 107, train_steps: 4280 +2026-06-20 02:36:37.802 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 02:36:37.985 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:36:38.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:37:42.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:37:42.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:39:39.229 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 02:39:40.399 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:41:35.347 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 02:41:36.072 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:43:35.145 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:50:34.993 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 108, train_steps: 4320 +2026-06-20 02:50:35.727 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 02:50:35.891 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:50:35.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:51:40.442 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:51:40.445 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:53:37.573 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 02:53:38.387 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:55:33.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.45s (1.91min) +2026-06-20 02:55:33.972 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:57:30.647 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:04:30.686 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 109, train_steps: 4360 +2026-06-20 03:04:31.408 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 03:04:31.611 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:04:31.750 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:05:35.994 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:05:35.997 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:07:33.406 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 03:07:34.210 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:09:29.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-20 03:09:30.014 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:11:26.574 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:18:25.320 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 110, train_steps: 4400 +2026-06-20 03:19:01.205 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=110 (missing=0, unexpected=0) +2026-06-20 03:19:03.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 03:19:03.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:19:03.442 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:20:10.103 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:20:10.106 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:22:07.524 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 03:22:08.499 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:24:03.966 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.41s (1.92min) +2026-06-20 03:24:04.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:26:02.473 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:33:05.397 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 111, train_steps: 4440 +2026-06-20 03:33:06.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 03:33:06.207 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:33:06.239 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:34:10.889 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:34:10.892 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:36:10.098 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 03:36:11.102 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:38:06.021 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 03:38:06.732 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:40:04.482 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:47:04.312 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 112, train_steps: 4480 +2026-06-20 03:47:05.043 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 03:47:05.176 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:47:05.221 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:48:09.558 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:48:09.562 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:50:04.983 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 03:50:06.160 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:52:00.943 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-20 03:52:01.658 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:54:00.964 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:01:04.003 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 113, train_steps: 4520 +2026-06-20 04:01:04.875 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 04:01:05.057 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:01:05.152 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:02:09.350 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:02:09.354 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:04:06.611 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 04:04:07.486 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:06:02.814 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.15s (1.92min) +2026-06-20 04:06:03.532 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:08:01.584 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:15:03.777 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 114, train_steps: 4560 +2026-06-20 04:15:04.474 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 04:15:04.652 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:15:04.753 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:16:09.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:16:09.176 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:18:04.152 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 04:18:05.136 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:20:00.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.72s (1.91min) +2026-06-20 04:20:00.895 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:22:00.099 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:29:01.457 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 115, train_steps: 4600 +2026-06-20 04:29:02.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 04:29:02.336 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:29:02.378 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:30:06.706 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:30:06.706 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:32:02.494 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 04:32:03.489 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:33:58.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.33s (1.91min) +2026-06-20 04:33:58.775 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:35:55.624 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:42:56.473 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 116, train_steps: 4640 +2026-06-20 04:42:57.159 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 04:42:57.273 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:42:57.302 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:44:01.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:44:01.995 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:45:58.380 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 04:45:59.352 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:47:53.698 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.25s (1.90min) +2026-06-20 04:47:54.408 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:49:50.887 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:56:50.216 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 117, train_steps: 4680 +2026-06-20 04:56:50.953 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 04:56:51.098 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:56:51.138 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:57:55.545 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:57:55.548 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:59:51.798 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 04:59:52.615 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:01:46.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.49s (1.89min) +2026-06-20 05:01:47.079 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:03:44.144 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:10:42.577 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 118, train_steps: 4720 +2026-06-20 05:10:43.296 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 05:10:43.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:10:43.473 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:11:47.789 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:11:47.792 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:13:46.957 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 05:13:47.774 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:15:41.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.26s (1.89min) +2026-06-20 05:15:42.001 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:17:38.994 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:24:37.893 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 119, train_steps: 4760 +2026-06-20 05:24:38.623 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 05:24:38.786 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:24:38.847 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:25:43.456 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:25:43.459 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:27:40.119 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 05:27:41.079 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:29:34.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-20 05:29:35.559 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:31:33.103 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:38:32.150 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 120, train_steps: 4800 +2026-06-20 05:39:09.176 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=120 (missing=0, unexpected=0) +2026-06-20 05:39:10.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 05:39:11.025 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:39:11.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:40:17.162 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:40:17.165 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:42:12.566 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 05:42:13.443 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:44:08.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.96s (1.92min) +2026-06-20 05:44:09.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:46:10.086 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:53:12.050 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 121, train_steps: 4840 +2026-06-20 05:53:12.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 05:53:12.933 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:53:12.986 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:54:17.685 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:54:17.689 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:56:13.155 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 05:56:13.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:58:09.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.97s (1.92min) +2026-06-20 05:58:10.076 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:00:10.575 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:07:11.953 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 122, train_steps: 4880 +2026-06-20 06:07:12.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 06:07:12.760 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:07:12.798 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:08:17.227 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:08:17.229 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:10:13.195 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 06:10:14.075 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:12:08.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.25s (1.90min) +2026-06-20 06:12:09.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:14:10.579 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:21:11.856 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 123, train_steps: 4920 +2026-06-20 06:21:12.561 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 06:21:12.731 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:21:12.810 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:22:17.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:22:17.061 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:24:12.071 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 06:24:13.102 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:26:08.788 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.47s (1.92min) +2026-06-20 06:26:09.511 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:28:09.832 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:35:11.839 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 124, train_steps: 4960 +2026-06-20 06:35:12.538 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 06:35:12.691 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:35:12.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:36:16.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:36:16.824 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:38:12.129 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 06:38:13.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:40:08.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.44s (1.92min) +2026-06-20 06:40:09.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:42:09.771 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:49:11.638 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 125, train_steps: 5000 +2026-06-20 06:50:15.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 06:50:15.318 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:50:15.361 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:51:19.977 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:51:19.979 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:53:15.926 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 06:53:16.732 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:55:11.653 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.68s (1.91min) +2026-06-20 06:55:12.372 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:57:11.762 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:04:11.908 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 126, train_steps: 5040 +2026-06-20 07:04:12.630 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 07:04:12.758 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:04:12.802 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:05:17.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:05:17.438 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:07:13.988 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 07:07:15.025 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:09:10.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.08s (1.92min) +2026-06-20 07:09:11.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:11:11.164 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:18:11.922 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 127, train_steps: 5080 +2026-06-20 07:18:12.629 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 07:18:12.812 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:18:12.907 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:19:17.264 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:19:17.267 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:21:13.050 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 07:21:13.866 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:23:09.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.50s (1.93min) +2026-06-20 07:23:10.327 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:25:10.157 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:32:11.162 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 128, train_steps: 5120 +2026-06-20 07:32:11.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 07:32:11.994 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:32:12.034 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:33:16.324 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:33:16.327 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:35:13.016 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 07:35:13.829 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:37:08.641 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.56s (1.91min) +2026-06-20 07:37:09.373 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:39:09.929 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:46:11.976 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 129, train_steps: 5160 +2026-06-20 07:46:12.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 07:46:12.873 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:46:12.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:47:17.852 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:47:17.855 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:49:13.070 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 07:49:13.876 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:51:09.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-20 07:51:09.758 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:53:10.302 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:00:11.772 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 130, train_steps: 5200 +2026-06-20 08:00:48.775 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=130 (missing=0, unexpected=0) +2026-06-20 08:00:51.244 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 08:00:51.334 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:00:51.377 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:01:58.676 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:01:58.678 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:03:52.509 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 08:03:53.310 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:05:47.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.72s (1.90min) +2026-06-20 08:05:48.018 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:07:44.843 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:14:45.322 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 131, train_steps: 5240 +2026-06-20 08:14:46.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 08:14:46.169 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:14:46.224 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:15:50.907 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:15:50.910 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:17:46.989 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 08:17:47.983 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:19:41.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.38s (1.89min) +2026-06-20 08:19:42.139 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:21:39.466 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:28:37.832 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 132, train_steps: 5280 +2026-06-20 08:28:38.525 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 08:28:38.728 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:28:38.830 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:29:43.683 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:29:43.684 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:31:40.849 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 08:31:41.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:33:35.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.44s (1.89min) +2026-06-20 08:33:36.119 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:35:35.262 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:42:34.344 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 133, train_steps: 5320 +2026-06-20 08:42:35.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 08:42:35.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:42:35.337 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:43:40.034 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:43:40.038 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:45:35.887 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 08:45:36.874 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:47:32.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.77s (1.93min) +2026-06-20 08:47:33.450 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:49:30.692 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:56:29.436 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 134, train_steps: 5360 +2026-06-20 08:56:30.114 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 08:56:30.238 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:56:30.278 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:57:34.872 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:57:34.875 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:59:32.040 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 08:59:32.907 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:01:26.319 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 09:01:27.048 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:03:23.804 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:10:22.978 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 135, train_steps: 5400 +2026-06-20 09:10:23.673 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 09:10:23.835 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:10:23.914 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:11:28.466 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:11:28.469 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:13:23.932 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 09:13:24.753 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:15:18.588 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.61s (1.89min) +2026-06-20 09:15:19.310 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:17:17.098 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:24:16.798 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 136, train_steps: 5440 +2026-06-20 09:24:17.532 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 09:24:17.718 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:24:17.813 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:25:22.388 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:25:22.389 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:27:18.656 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 09:27:19.568 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:29:12.958 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.15s (1.89min) +2026-06-20 09:29:13.686 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:31:13.177 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:38:13.279 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 137, train_steps: 5480 +2026-06-20 09:38:13.936 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 09:38:14.087 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:38:14.142 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:39:18.672 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:39:18.676 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:41:14.371 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 09:41:15.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:43:09.522 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-20 09:43:10.251 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:45:10.503 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:52:12.148 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 138, train_steps: 5520 +2026-06-20 09:52:12.866 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 09:52:13.000 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:52:13.029 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:53:17.667 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:53:17.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:55:13.383 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 09:55:14.571 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:57:10.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.46s (1.92min) +2026-06-20 09:57:10.921 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:59:11.151 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:06:12.684 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 139, train_steps: 5560 +2026-06-20 10:06:13.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 10:06:13.542 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:06:13.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:07:18.157 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:07:18.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:09:13.264 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 10:09:14.258 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:11:08.675 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.20s (1.90min) +2026-06-20 10:11:09.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:13:09.781 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:20:12.692 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 140, train_steps: 5600 +2026-06-20 10:20:50.000 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=140 (missing=0, unexpected=0) +2026-06-20 10:20:51.812 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 10:20:52.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:20:52.201 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:21:58.263 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:21:58.266 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:23:53.040 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 10:23:54.013 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:25:47.600 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-20 10:25:48.313 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:27:45.363 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:34:44.005 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 141, train_steps: 5640 +2026-06-20 10:34:44.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 10:34:44.824 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:34:44.855 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:35:49.445 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:35:49.445 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:37:46.846 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 10:37:47.859 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:39:41.415 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.51s (1.89min) +2026-06-20 10:39:42.148 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:41:39.463 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:48:38.463 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 142, train_steps: 5680 +2026-06-20 10:48:39.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 10:48:39.281 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:48:39.316 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:49:44.890 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:49:44.890 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:51:43.153 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 10:51:44.013 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:53:37.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-20 10:53:37.851 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:55:37.992 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:02:37.909 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 143, train_steps: 5720 +2026-06-20 11:02:38.632 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 11:02:38.805 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:02:38.897 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:03:43.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:03:43.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:05:42.169 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 11:05:43.321 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:07:36.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 11:07:37.313 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:09:36.843 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:16:37.098 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 144, train_steps: 5760 +2026-06-20 11:16:37.797 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 11:16:37.893 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:16:37.933 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:17:42.521 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:17:42.524 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:19:38.707 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 11:19:39.839 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:21:33.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.47s (1.89min) +2026-06-20 11:21:34.087 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:23:30.885 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:30:30.387 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 145, train_steps: 5800 +2026-06-20 11:30:31.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 11:30:31.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:30:31.409 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:31:35.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:31:35.607 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:33:33.207 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 11:33:34.198 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:35:28.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.89s (1.90min) +2026-06-20 11:35:28.900 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:37:25.683 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:44:23.919 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 146, train_steps: 5840 +2026-06-20 11:44:24.628 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 11:44:24.769 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:44:24.807 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:45:29.443 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:45:29.447 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:47:23.247 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 11:47:24.234 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:49:19.133 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-20 11:49:19.856 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:51:17.040 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:58:15.926 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 147, train_steps: 5880 +2026-06-20 11:58:16.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 11:58:16.732 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:58:16.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:59:21.783 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:59:21.798 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:01:16.232 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 12:01:17.126 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:03:12.586 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.29s (1.92min) +2026-06-20 12:03:13.294 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:05:12.904 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:12:12.696 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 148, train_steps: 5920 +2026-06-20 12:12:13.405 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 12:12:13.552 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:12:13.617 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:13:18.281 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:13:18.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:15:13.144 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 12:15:13.972 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:17:09.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.69s (1.91min) +2026-06-20 12:17:09.789 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:19:10.588 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:26:11.390 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 149, train_steps: 5960 +2026-06-20 12:26:12.136 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 12:26:12.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:26:12.377 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:27:16.947 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:27:16.950 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:29:13.029 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 12:29:14.010 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:31:08.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.72s (1.91min) +2026-06-20 12:31:09.532 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:33:09.829 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:40:10.991 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 150, train_steps: 6000 +2026-06-20 12:41:53.502 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=150 (missing=0, unexpected=0) +2026-06-20 12:41:56.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 12:41:56.300 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:41:56.437 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:43:02.563 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:43:02.566 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:44:59.743 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 12:45:00.757 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:46:55.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.70s (1.91min) +2026-06-20 12:46:56.236 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:48:52.989 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:55:51.601 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 151, train_steps: 6040 +2026-06-20 12:55:52.305 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 12:55:52.460 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:55:52.514 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:56:56.992 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:56:56.992 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:58:50.721 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 12:58:51.704 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:00:46.189 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.41s (1.91min) +2026-06-20 13:00:46.908 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:02:43.653 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:09:41.910 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 152, train_steps: 6080 +2026-06-20 13:09:42.596 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 13:09:42.720 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:09:42.753 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:10:47.759 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:10:47.762 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:12:43.910 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 13:12:44.909 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:14:38.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-20 13:14:39.492 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:16:38.550 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:23:37.219 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 153, train_steps: 6120 +2026-06-20 13:23:38.091 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 13:23:38.221 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:23:38.258 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:24:43.422 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:24:43.426 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:26:40.127 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 13:26:41.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:28:34.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.39s (1.89min) +2026-06-20 13:28:35.473 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:30:32.906 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:37:31.688 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 154, train_steps: 6160 +2026-06-20 13:37:32.392 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 13:37:32.558 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:37:32.634 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:38:37.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:38:37.036 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:40:33.788 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 13:40:34.612 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:42:29.329 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.47s (1.91min) +2026-06-20 13:42:30.062 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:44:27.157 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:51:24.975 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 155, train_steps: 6200 +2026-06-20 13:51:25.702 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 13:51:25.920 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:51:26.049 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:52:30.279 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:52:30.282 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:54:24.447 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 13:54:25.450 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:56:19.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.63s (1.89min) +2026-06-20 13:56:19.859 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:58:16.751 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:05:15.336 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 156, train_steps: 6240 +2026-06-20 14:05:16.034 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 14:05:16.224 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:05:16.330 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:06:21.093 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:06:21.102 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:08:15.931 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 14:08:16.804 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:10:11.722 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.74s (1.91min) +2026-06-20 14:10:12.457 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:12:11.103 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:19:12.229 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 157, train_steps: 6280 +2026-06-20 14:19:12.923 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 14:19:13.038 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:19:13.078 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:20:17.666 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:20:17.669 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:22:12.398 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 14:22:13.329 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:24:08.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.22s (1.92min) +2026-06-20 14:24:09.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:26:09.904 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:33:10.821 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 158, train_steps: 6320 +2026-06-20 14:33:11.520 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 14:33:11.728 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:33:11.843 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:34:16.168 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:34:16.170 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:36:10.935 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 14:36:11.846 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:38:07.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.88s (1.91min) +2026-06-20 14:38:07.784 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:40:05.257 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:47:04.559 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 159, train_steps: 6360 +2026-06-20 14:47:05.291 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 14:47:05.405 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:47:05.442 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:48:09.889 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:48:09.892 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:50:04.048 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 14:50:05.051 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:51:59.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.86s (1.90min) +2026-06-20 14:51:59.929 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:53:57.321 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:00:59.172 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 160, train_steps: 6400 +2026-06-20 15:01:36.726 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=160 (missing=0, unexpected=0) +2026-06-20 15:01:39.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 15:01:39.448 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:01:39.509 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:02:46.036 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:02:46.036 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:04:42.879 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 15:04:43.684 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:06:37.408 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.48s (1.89min) +2026-06-20 15:06:38.136 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:08:38.300 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:15:37.514 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 161, train_steps: 6440 +2026-06-20 15:15:38.206 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 15:15:38.322 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:15:38.361 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:16:42.898 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:16:42.902 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:18:40.375 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 15:18:41.382 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:20:35.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-20 15:20:36.332 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:22:36.147 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:29:36.888 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 162, train_steps: 6480 +2026-06-20 15:29:37.568 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 15:29:37.680 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:29:37.713 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:30:42.266 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:30:42.267 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:32:37.444 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 15:32:38.271 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:34:33.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.48s (1.91min) +2026-06-20 15:34:33.827 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:36:31.316 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:43:31.477 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 163, train_steps: 6520 +2026-06-20 15:43:32.171 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 15:43:32.328 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:43:32.386 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:44:37.069 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:44:37.070 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:46:33.676 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 15:46:34.501 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:48:29.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.47s (1.91min) +2026-06-20 15:48:29.934 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:50:26.537 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:57:25.498 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 164, train_steps: 6560 +2026-06-20 15:57:26.227 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 15:57:26.319 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:57:26.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:58:30.675 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:58:30.678 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:00:26.401 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 16:00:27.221 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:02:21.064 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.61s (1.89min) +2026-06-20 16:02:21.793 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:04:18.343 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:11:18.366 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 165, train_steps: 6600 +2026-06-20 16:11:19.087 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 16:11:19.231 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:11:19.264 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:12:24.091 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:12:24.094 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:14:19.105 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 16:14:20.252 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:16:14.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.75s (1.90min) +2026-06-20 16:16:14.870 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:18:14.126 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:25:13.928 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 166, train_steps: 6640 +2026-06-20 16:25:14.618 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 16:25:14.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:25:14.755 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:26:19.599 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:26:19.600 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:28:13.831 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 16:28:14.920 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:30:08.844 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.78s (1.90min) +2026-06-20 16:30:09.574 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:32:10.307 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:39:12.150 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 167, train_steps: 6680 +2026-06-20 16:39:12.855 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 16:39:13.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:39:13.160 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:40:17.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:40:17.711 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:42:12.094 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 16:42:13.306 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:44:08.745 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.36s (1.92min) +2026-06-20 16:44:09.477 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:46:10.020 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:53:11.385 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 168, train_steps: 6720 +2026-06-20 16:53:12.091 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 16:53:12.255 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:53:12.334 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:54:17.034 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:54:17.035 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:56:11.767 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 16:56:12.763 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:58:07.736 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-20 16:58:08.464 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:00:07.579 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:07:09.174 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 169, train_steps: 6760 +2026-06-20 17:07:09.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 17:07:10.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:07:10.150 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:08:14.949 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:08:14.950 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:10:11.903 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 17:10:12.906 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:12:08.254 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.20s (1.92min) +2026-06-20 17:12:08.969 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:14:09.320 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:21:11.055 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 170, train_steps: 6800 +2026-06-20 17:21:47.266 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=170 (missing=0, unexpected=0) +2026-06-20 17:21:49.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 17:21:49.433 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:21:49.550 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:22:55.027 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:22:55.030 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:24:49.925 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 17:24:50.792 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:26:44.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.05s (1.88min) +2026-06-20 17:26:44.743 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:28:41.424 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:35:39.727 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 39, sp_rank 0: update step: 171, train_steps: 6840 +2026-06-20 17:35:40.431 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 39, dp_rank 19: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=9, rank_in_group=1 +2026-06-20 17:35:40.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:35:40.895 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:36:46.210 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 39 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:36:46.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 39 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:38:42.536 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 39: Using local metric names without global sync: ['avg'] +2026-06-20 17:38:43.544 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 39: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:40:36.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 39: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.89s (1.88min) +2026-06-20 17:40:37.378 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 39, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-29 01:27:08.198 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 01:27:08.198 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 01:51:25.467 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 01:51:25.469 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 02:20:16.089 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 02:20:16.090 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 02:46:50.978 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 02:46:50.979 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:06:52.585 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:06:52.587 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:31:13.829 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:31:13.829 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:57:54.304 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:57:54.304 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:25:32.922 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:25:32.923 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:53:31.625 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:53:31.626 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 05:14:51.457 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 05:14:51.458 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 05:39:03.400 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 05:39:03.401 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:02:15.358 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:02:15.360 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:25:05.675 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:25:05.677 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:48:17.460 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:48:17.461 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 07:11:57.377 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 07:11:57.378 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 07:33:36.211 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 07:33:36.211 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:07:34.008 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:07:34.009 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:29:39.395 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:29:39.396 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:58:44.520 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:58:44.523 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:22:11.109 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:22:11.112 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:47:13.877 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:47:13.878 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 10:10:21.112 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 10:10:21.115 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 10:35:25.136 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 10:35:25.137 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:02:55.870 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:02:55.871 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:26:43.815 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:26:43.816 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:54:47.534 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:54:47.534 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 12:17:59.714 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 12:17:59.715 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 12:46:13.061 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 12:46:13.064 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 13:12:11.464 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 13:12:11.465 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 13:36:41.911 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 13:36:41.912 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:03:52.307 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:03:52.309 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:26:16.818 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:26:16.819 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:49:03.283 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:49:03.286 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 15:21:24.825 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 15:21:24.826 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 15:46:20.441 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 15:46:20.443 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 16:15:56.532 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 16:15:56.532 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 16:40:28.585 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 16:40:28.587 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:09:23.115 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:09:23.117 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:32:40.981 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:32:40.983 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:53:11.993 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:53:11.993 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 18:16:51.886 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 18:16:51.887 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 18:38:51.324 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 18:38:51.332 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_42.log b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_42.log new file mode 100644 index 0000000000000000000000000000000000000000..1098d83b1a41aecfebc220086ddaa13107be9418 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_42.log @@ -0,0 +1,3241 @@ +2026-06-19 01:22:18.343 | INFO | __main__:__init__:933 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-19 01:22:18.997 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-19 01:22:18.998 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-19 01:30:17.294 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 01:30:31.533 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 01:30:31.537 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:30:31.538 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:31:52.827 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_10/index_2048_seed_1800_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 01:31:56.655 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 01:32:14.242 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:32:14.252 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:32:14.252 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:33:22.362 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_10/index_2048_seed_1801_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 01:33:25.130 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 01:33:42.746 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:33:45.515 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 01:34:00.984 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 01:35:56.997 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.94s (1.93min) +2026-06-19 01:35:57.688 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 01:37:54.185 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 01:37:57.221 | INFO | hyvideo.utils.grpo_utils:train_one_step:3127 - Rank 42, sp_rank 0: init adaptive KL controller (init=1.000e-07, min=1.000e-07, max=1.000e-05) +2026-06-19 01:45:11.841 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 1, train_steps: 40 +2026-06-19 01:45:12.575 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 01:45:12.738 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 01:45:12.806 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:45:12.847 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:46:18.731 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000001/group_10/index_6903_seed_2427_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 01:46:23.587 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 01:46:40.990 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:46:40.998 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:46:41.000 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:47:47.343 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000001/group_10/index_6903_seed_2428_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 01:47:51.967 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 01:48:08.863 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:48:10.945 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 01:48:11.889 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 01:50:07.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.70s (1.93min) +2026-06-19 01:50:08.466 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 01:52:06.956 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 01:59:06.753 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 2, train_steps: 80 +2026-06-19 01:59:07.381 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 01:59:07.519 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 01:59:07.558 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:59:07.622 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:00:13.117 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000002/group_10/index_8507_seed_3373_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 02:00:17.809 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 02:00:36.136 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:00:36.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:00:36.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:01:46.258 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000002/group_10/index_8507_seed_3374_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 02:01:50.964 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 02:02:09.397 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:02:10.528 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:02:11.572 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:04:06.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.88s (1.91min) +2026-06-19 02:04:07.374 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:06:03.755 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:13:04.851 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 3, train_steps: 120 +2026-06-19 02:13:05.466 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 02:13:05.620 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:13:05.685 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:13:05.739 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:14:11.421 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_10/index_6864_seed_5045_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 02:14:15.698 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 02:14:33.333 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:14:33.341 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:14:33.344 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:15:40.513 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_10/index_6864_seed_5046_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 02:15:43.820 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 02:16:01.870 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:16:04.655 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:16:05.870 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:17:59.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.77s (1.90min) +2026-06-19 02:18:00.448 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:19:59.353 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:26:59.885 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 4, train_steps: 160 +2026-06-19 02:27:00.546 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 02:27:00.698 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:27:00.760 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:27:00.873 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:28:06.827 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000004/group_10/index_6010_seed_290_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 02:28:11.713 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 02:28:29.132 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:28:29.141 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:28:29.141 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:29:35.233 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000004/group_10/index_6010_seed_291_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 02:29:39.007 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 02:29:56.858 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:30:00.316 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:30:01.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:31:55.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.62s (1.91min) +2026-06-19 02:31:56.651 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:33:53.226 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:40:53.219 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 5, train_steps: 200 +2026-06-19 02:40:53.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 02:40:54.007 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:40:54.040 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:40:54.088 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:41:59.886 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_10/index_3767_seed_26097_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 02:42:04.812 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 02:42:22.250 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:42:22.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:42:22.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:43:28.707 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_10/index_3767_seed_26098_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 02:43:33.537 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 02:43:51.121 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:43:51.812 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:43:52.707 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:45:46.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.50s (1.89min) +2026-06-19 02:45:46.994 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:47:43.353 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:54:40.649 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 6, train_steps: 240 +2026-06-19 02:54:41.361 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 02:54:41.558 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:54:41.896 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:54:41.961 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:55:48.863 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_10/index_9163_seed_24870_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 02:55:52.647 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 02:56:10.301 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:56:10.310 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:56:10.311 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:57:19.373 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_10/index_9163_seed_24871_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 02:57:24.277 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 02:57:42.942 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:57:42.949 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:57:43.832 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:59:37.353 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.38s (1.89min) +2026-06-19 02:59:38.043 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:01:37.732 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:08:37.471 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 7, train_steps: 280 +2026-06-19 03:08:38.148 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 03:08:38.346 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:08:38.449 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:08:38.543 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:09:44.159 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_10/index_9143_seed_10367_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 03:09:48.244 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 03:10:06.116 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:10:06.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:10:06.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:11:14.808 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_10/index_9143_seed_10368_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 03:11:17.742 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 03:11:35.584 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:11:38.337 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:11:39.489 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:13:34.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.68s (1.91min) +2026-06-19 03:13:34.976 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:15:32.045 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:22:31.144 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 8, train_steps: 320 +2026-06-19 03:22:31.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 03:22:31.953 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:22:31.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:22:32.078 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:23:38.932 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000008/group_10/index_8958_seed_4822_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 03:23:44.304 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 03:24:01.800 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:24:01.806 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:24:01.806 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:25:09.087 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000008/group_10/index_8958_seed_4823_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 03:25:13.754 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 03:25:31.653 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:25:33.100 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:25:33.965 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:27:28.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-19 03:27:29.432 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:29:26.908 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:36:25.652 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 9, train_steps: 360 +2026-06-19 03:36:26.346 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 03:36:26.503 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:36:26.559 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:36:26.662 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:37:32.792 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_10/index_3532_seed_17055_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 03:37:37.525 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 03:37:55.044 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:37:55.053 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:37:55.053 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:39:02.007 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_10/index_3532_seed_17056_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 03:39:06.785 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 03:39:24.630 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:39:26.010 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:39:26.915 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:41:20.400 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.36s (1.89min) +2026-06-19 03:41:21.106 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:43:17.559 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:50:16.926 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 10, train_steps: 400 +2026-06-19 03:50:53.145 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=10 (missing=0, unexpected=0) +2026-06-19 03:50:56.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 03:50:56.308 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:50:56.332 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:50:56.399 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:52:03.728 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_10/index_5505_seed_9831_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 03:52:08.515 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 03:52:24.746 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:52:24.754 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:52:24.757 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:53:32.575 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_10/index_5505_seed_9832_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 03:53:37.212 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 03:53:53.913 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:53:57.828 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:53:58.706 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:55:52.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.65s (1.89min) +2026-06-19 03:55:53.209 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:57:49.765 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:04:48.369 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 11, train_steps: 440 +2026-06-19 04:04:49.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 04:04:49.254 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:04:49.342 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:04:49.458 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:05:55.118 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_10/index_5521_seed_27593_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 04:05:58.833 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 04:06:16.910 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:06:16.916 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:06:16.916 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:07:24.736 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_10/index_5521_seed_27594_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 04:07:29.834 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 04:07:47.793 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:07:48.853 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:07:49.735 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:09:43.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.25s (1.89min) +2026-06-19 04:09:43.797 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:11:40.269 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:18:37.294 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 12, train_steps: 480 +2026-06-19 04:18:37.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 04:18:38.164 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:18:38.267 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:18:38.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:19:44.599 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_10/index_4238_seed_23425_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 04:19:49.453 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 04:20:06.992 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:20:07.000 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:20:07.000 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:21:12.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_10/index_4238_seed_23426_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 04:21:17.676 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 04:21:35.463 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:21:37.180 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:21:38.090 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:23:33.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.19s (1.92min) +2026-06-19 04:23:34.070 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:25:31.244 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:32:29.813 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 13, train_steps: 520 +2026-06-19 04:32:30.512 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 04:32:30.669 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:32:30.729 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:32:30.862 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:33:36.868 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_10/index_4607_seed_31262_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 04:33:41.816 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 04:33:59.993 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:34:00.001 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:34:00.001 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:35:06.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_10/index_4607_seed_31263_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 04:35:10.492 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 04:35:28.761 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:35:30.313 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:35:31.321 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:37:24.102 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.77s (1.88min) +2026-06-19 04:37:24.792 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:39:21.078 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:46:18.587 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 14, train_steps: 560 +2026-06-19 04:46:19.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 04:46:19.367 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:46:19.406 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:46:19.472 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:47:25.983 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_10/index_2467_seed_10643_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 04:47:30.954 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 04:47:48.614 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:47:48.622 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:47:48.622 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:48:54.277 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_10/index_2467_seed_10644_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 04:48:59.143 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 04:49:16.736 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:49:17.445 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:49:18.360 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:51:12.963 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.49s (1.91min) +2026-06-19 04:51:13.650 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:53:12.996 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:00:11.825 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 15, train_steps: 600 +2026-06-19 05:00:12.525 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 05:00:12.663 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:00:12.698 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:00:12.789 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:01:18.986 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000015/group_10/index_513_seed_2568_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 05:01:23.800 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 05:01:42.246 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:01:42.252 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:01:42.252 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:02:47.609 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000015/group_10/index_513_seed_2569_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 05:02:52.446 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 05:03:10.206 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:03:11.349 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:03:12.520 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:05:07.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.64s (1.91min) +2026-06-19 05:05:08.060 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:07:05.975 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:14:04.888 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 16, train_steps: 640 +2026-06-19 05:14:05.573 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 05:14:05.667 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:14:05.719 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:14:05.766 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:15:11.681 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_10/index_1611_seed_9658_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 05:15:16.008 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 05:15:33.723 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:15:33.732 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:15:33.735 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:16:42.993 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_10/index_1611_seed_9659_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 05:16:45.901 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 05:17:03.635 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:17:06.988 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:17:07.987 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:19:03.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.02s (1.92min) +2026-06-19 05:19:03.742 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:21:01.963 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:28:03.324 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 17, train_steps: 680 +2026-06-19 05:28:04.007 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 05:28:04.213 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:28:04.308 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:28:04.391 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:29:10.600 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_10/index_7129_seed_6657_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 05:29:13.645 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 05:29:31.548 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:29:31.557 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:29:31.557 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:30:39.476 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_10/index_7129_seed_6658_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 05:30:42.971 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 05:31:01.004 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:31:03.926 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:31:05.036 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:32:58.689 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.48s (1.89min) +2026-06-19 05:32:59.387 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:34:56.267 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:41:58.104 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 18, train_steps: 720 +2026-06-19 05:41:58.777 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 05:41:58.948 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:41:59.034 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:41:59.149 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:43:05.080 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_10/index_6936_seed_24578_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 05:43:08.184 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 05:43:26.233 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:43:26.243 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:43:26.243 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:44:35.397 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_10/index_6936_seed_24579_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 05:44:38.459 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 05:44:56.609 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:45:00.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:45:01.200 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:46:55.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.88s (1.90min) +2026-06-19 05:46:55.797 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:48:52.333 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:55:50.490 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 19, train_steps: 760 +2026-06-19 05:55:51.177 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 05:55:51.308 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:55:51.343 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:55:51.435 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:56:57.461 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_10/index_4790_seed_30649_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 05:57:02.511 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 05:57:20.373 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:57:20.382 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:57:20.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:58:25.680 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_10/index_4790_seed_30650_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 05:58:30.595 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 05:58:48.562 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:58:49.068 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:58:49.946 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:00:43.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.10s (1.89min) +2026-06-19 06:00:43.876 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:02:40.342 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:09:39.753 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 20, train_steps: 800 +2026-06-19 06:10:16.199 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=20 (missing=0, unexpected=0) +2026-06-19 06:10:18.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 06:10:18.917 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:10:19.000 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:10:19.100 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:11:26.414 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_10/index_7886_seed_1912_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 06:11:30.535 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 06:11:48.729 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:11:48.738 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:11:48.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:12:54.810 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_10/index_7886_seed_1913_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 06:12:57.908 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 06:13:16.209 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:13:19.050 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:13:20.165 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:15:13.224 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.90s (1.88min) +2026-06-19 06:15:13.910 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:17:12.688 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:24:12.552 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 21, train_steps: 840 +2026-06-19 06:24:13.270 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 06:24:13.444 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:24:13.537 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:24:13.645 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:25:19.409 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_10/index_6990_seed_32014_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 06:25:22.417 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 06:25:40.284 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:25:40.290 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:25:40.291 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:26:47.694 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_10/index_6990_seed_32015_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 06:26:50.614 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 06:27:08.699 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:27:11.261 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:27:12.340 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:29:06.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.51s (1.91min) +2026-06-19 06:29:07.668 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:31:04.579 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:38:04.278 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 22, train_steps: 880 +2026-06-19 06:38:04.948 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 06:38:05.109 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:38:05.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:38:05.314 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:39:11.120 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_10/index_696_seed_3369_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 06:39:16.228 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 06:39:33.902 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:39:33.910 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:39:33.910 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:40:40.105 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_10/index_696_seed_3370_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 06:40:44.364 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 06:41:02.450 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:41:03.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:41:04.324 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:42:58.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.99s (1.90min) +2026-06-19 06:42:59.165 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:44:56.506 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:51:58.169 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 23, train_steps: 920 +2026-06-19 06:51:58.842 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 06:51:58.942 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:51:58.981 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:51:59.028 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:53:04.953 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000023/group_10/index_5017_seed_29113_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 06:53:10.092 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 06:53:27.900 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:53:27.909 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:53:27.909 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:54:35.027 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000023/group_10/index_5017_seed_29114_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 06:54:39.069 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 06:54:57.659 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:54:59.789 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:55:00.805 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:56:54.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.11s (1.90min) +2026-06-19 06:56:55.623 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:58:52.322 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:05:51.320 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 24, train_steps: 960 +2026-06-19 07:05:52.005 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 07:05:52.224 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:05:52.352 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:05:52.447 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:06:58.302 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_10/index_6542_seed_9458_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 07:07:02.086 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 07:07:19.972 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:07:19.981 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:07:19.982 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:08:27.618 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_10/index_6542_seed_9459_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 07:08:31.996 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 07:08:50.267 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:08:52.217 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:08:53.216 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:10:46.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-19 07:10:46.872 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:12:44.157 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:19:41.978 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 25, train_steps: 1000 +2026-06-19 07:20:52.105 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 07:20:52.257 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:20:52.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:20:52.451 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:21:58.445 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_10/index_9816_seed_5932_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 07:22:01.510 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 07:22:19.645 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:22:19.654 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:22:19.655 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:23:28.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_10/index_9816_seed_5933_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 07:23:34.041 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 07:23:52.347 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:23:52.666 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:23:53.607 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:25:46.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.81s (1.88min) +2026-06-19 07:25:47.206 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:27:44.383 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:34:43.680 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 26, train_steps: 1040 +2026-06-19 07:34:44.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 07:34:44.473 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:34:44.509 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:34:44.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:35:51.036 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_10/index_8786_seed_1976_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 07:35:55.798 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 07:36:13.231 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:36:13.240 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:36:13.242 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:37:21.152 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_10/index_8786_seed_1977_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 07:37:25.107 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 07:37:43.069 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:37:45.443 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:37:46.334 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:39:39.638 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.14s (1.89min) +2026-06-19 07:39:40.348 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:41:39.073 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:48:39.667 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 27, train_steps: 1080 +2026-06-19 07:48:40.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 07:48:40.479 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:48:40.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:48:40.610 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:49:46.950 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_10/index_2598_seed_1870_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 07:49:51.741 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 07:50:09.891 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:50:09.899 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:50:09.902 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:51:19.437 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_10/index_2598_seed_1871_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 07:51:23.348 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 07:51:42.853 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:51:43.285 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:51:44.273 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:53:38.298 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.99s (1.90min) +2026-06-19 07:53:39.007 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:55:38.423 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:02:38.176 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 28, train_steps: 1120 +2026-06-19 08:02:38.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 08:02:39.088 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:02:39.203 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:02:39.300 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:03:45.262 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_10/index_9784_seed_4859_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 08:03:50.240 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 08:04:08.320 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:04:08.329 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:04:08.330 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:05:17.572 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_10/index_9784_seed_4860_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 08:05:21.790 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 08:05:41.851 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:05:41.901 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:05:43.105 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:07:37.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.24s (1.90min) +2026-06-19 08:07:38.092 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:09:37.937 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:16:37.852 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 29, train_steps: 1160 +2026-06-19 08:16:38.691 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 08:16:38.847 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:16:38.898 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:16:39.169 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:17:45.104 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_10/index_3194_seed_8830_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 08:17:50.010 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 08:18:07.901 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:18:07.910 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:18:07.910 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:19:18.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_10/index_3194_seed_8831_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 08:19:22.371 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 08:19:42.547 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:19:42.881 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:19:43.798 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:21:37.264 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 08:21:37.961 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:23:37.408 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:30:39.546 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 30, train_steps: 1200 +2026-06-19 08:31:15.766 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=30 (missing=0, unexpected=0) +2026-06-19 08:31:18.259 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 08:31:18.453 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:31:18.565 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:31:18.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:32:26.585 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000030/group_10/index_2339_seed_12086_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 08:32:31.598 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 08:32:48.457 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:32:48.465 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:32:48.466 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:33:55.961 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000030/group_10/index_2339_seed_12087_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 08:34:00.781 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 08:34:18.766 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:34:19.980 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:34:21.028 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:36:14.510 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-19 08:36:15.227 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:38:14.174 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:45:14.279 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 31, train_steps: 1240 +2026-06-19 08:45:14.992 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 08:45:15.147 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:45:15.205 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:45:15.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:46:21.399 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_10/index_5895_seed_7611_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 08:46:25.533 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 08:46:43.916 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:46:43.927 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:46:43.927 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:47:50.804 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_10/index_5895_seed_7612_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 08:47:55.752 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 08:48:13.759 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:48:15.288 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:48:16.329 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:50:10.323 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.84s (1.90min) +2026-06-19 08:50:11.023 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:52:11.151 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:59:13.087 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 32, train_steps: 1280 +2026-06-19 08:59:13.775 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 08:59:13.918 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:59:13.954 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:59:14.045 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:00:19.857 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000032/group_10/index_1607_seed_6798_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 09:00:24.716 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 09:00:43.041 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:00:43.050 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:00:43.050 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:01:49.130 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000032/group_10/index_1607_seed_6799_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 09:01:54.135 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 09:02:12.552 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:02:13.436 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:02:14.525 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:04:08.713 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.91s (1.90min) +2026-06-19 09:04:09.421 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:06:10.409 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:13:13.271 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 33, train_steps: 1320 +2026-06-19 09:13:13.952 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 09:13:14.131 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:13:14.242 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:13:14.360 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:14:20.530 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000033/group_10/index_5034_seed_20724_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 09:14:24.478 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 09:14:42.902 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:14:42.912 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:14:42.914 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:15:50.343 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000033/group_10/index_5034_seed_20725_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 09:15:55.430 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 09:16:13.702 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:16:14.822 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:16:15.763 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:18:11.066 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.06s (1.92min) +2026-06-19 09:18:11.786 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:20:11.627 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:27:13.728 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 34, train_steps: 1360 +2026-06-19 09:27:14.445 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 09:27:14.639 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:27:14.752 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:27:14.856 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:28:20.747 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_10/index_7444_seed_3962_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 09:28:25.604 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 09:28:43.443 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:28:43.453 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:28:43.453 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:29:50.027 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_10/index_7444_seed_3963_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 09:29:54.931 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 09:30:13.054 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:30:13.695 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:30:14.862 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:32:08.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.81s (1.90min) +2026-06-19 09:32:09.592 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:34:09.780 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:41:12.390 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 35, train_steps: 1400 +2026-06-19 09:41:13.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 09:41:13.197 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:41:13.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:41:13.329 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:42:19.290 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000035/group_10/index_2371_seed_15034_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 09:42:24.123 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 09:42:42.820 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:42:42.829 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:42:42.829 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:43:48.701 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000035/group_10/index_2371_seed_15035_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 09:43:52.645 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 09:44:11.427 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:44:13.080 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:44:14.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:46:10.686 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 116.56s (1.94min) +2026-06-19 09:46:11.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:48:10.898 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:55:11.959 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 36, train_steps: 1440 +2026-06-19 09:55:12.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 09:55:12.804 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:55:12.880 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:55:12.996 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:56:19.175 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_10/index_4995_seed_11679_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 09:56:24.206 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 09:56:42.862 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:56:42.872 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:56:42.872 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:57:48.194 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_10/index_4995_seed_11680_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 09:57:53.021 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 09:58:11.291 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:58:12.646 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:58:13.775 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:00:08.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-19 10:00:09.672 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:02:10.009 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:09:11.698 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 37, train_steps: 1480 +2026-06-19 10:09:12.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 10:09:12.571 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:09:12.662 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:09:12.812 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:10:18.567 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_10/index_6839_seed_26290_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 10:10:23.334 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 10:10:41.886 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:10:41.895 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:10:41.895 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:11:47.692 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_10/index_6839_seed_26291_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 10:11:51.986 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 10:12:11.510 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:12:13.042 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:12:14.051 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:14:09.030 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.94s (1.92min) +2026-06-19 10:14:09.737 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:16:10.369 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:23:12.348 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 38, train_steps: 1520 +2026-06-19 10:23:13.053 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 10:23:13.267 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:23:13.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:23:13.469 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:24:19.290 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000038/group_10/index_7790_seed_32696_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 10:24:24.346 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 10:24:42.897 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:24:42.906 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:24:42.906 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:25:48.623 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000038/group_10/index_7790_seed_32697_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 10:25:53.568 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 10:26:11.612 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:26:12.667 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:26:13.786 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:28:09.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.16s (1.92min) +2026-06-19 10:28:09.954 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:30:10.743 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:37:11.943 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 39, train_steps: 1560 +2026-06-19 10:37:12.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 10:37:12.868 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:37:12.989 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:37:13.084 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:38:19.232 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000039/group_10/index_6145_seed_22803_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 10:38:23.457 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 10:38:42.477 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:38:42.487 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:38:42.487 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:41:17.567 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000039/group_10/index_6145_seed_22804_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 10:41:22.604 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 10:41:41.903 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:41:42.526 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:41:43.600 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:43:36.471 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.70s (1.88min) +2026-06-19 10:43:37.180 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:45:37.178 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:52:37.860 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 40, train_steps: 1600 +2026-06-19 10:53:14.113 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=40 (missing=0, unexpected=0) +2026-06-19 10:53:15.773 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 10:53:15.926 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:53:15.950 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:53:15.999 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:54:23.711 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_10/index_5160_seed_9956_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 10:54:28.019 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 10:54:46.090 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:54:46.098 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:54:46.098 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:55:52.465 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_10/index_5160_seed_9957_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 10:55:55.559 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 10:56:13.974 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:56:16.410 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:56:17.430 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:58:11.717 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.27s (1.90min) +2026-06-19 10:58:12.416 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:00:11.327 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:07:12.582 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 41, train_steps: 1640 +2026-06-19 11:07:13.291 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 11:07:13.461 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:07:13.541 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:07:13.676 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:08:19.569 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_10/index_9265_seed_19479_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 11:08:22.751 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 11:08:41.905 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:08:41.915 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:08:41.915 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:09:48.507 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_10/index_9265_seed_19480_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 11:09:53.569 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 11:10:11.482 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:10:12.668 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:10:13.822 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:12:08.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.28s (1.90min) +2026-06-19 11:12:09.087 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:14:08.758 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:21:10.881 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 42, train_steps: 1680 +2026-06-19 11:21:11.805 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 11:21:12.008 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:21:12.137 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:21:12.238 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:22:18.117 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000042/group_10/index_4764_seed_23616_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 11:22:22.163 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 11:22:41.618 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:22:41.628 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:22:41.628 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:23:47.591 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000042/group_10/index_4764_seed_23617_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 11:23:50.812 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 11:24:09.431 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:24:12.073 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:24:13.241 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:26:07.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.32s (1.91min) +2026-06-19 11:26:08.279 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:28:06.797 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:35:06.776 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 43, train_steps: 1720 +2026-06-19 11:35:07.458 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 11:35:07.565 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:35:07.609 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:35:07.657 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:36:13.643 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_10/index_96_seed_27840_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 11:36:18.893 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 11:36:37.507 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:36:37.517 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:36:37.517 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:37:46.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_10/index_96_seed_27841_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 11:37:51.457 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 11:38:09.724 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:38:10.299 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:38:11.244 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:40:05.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.38s (1.91min) +2026-06-19 11:40:06.549 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:42:03.525 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:49:04.427 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 44, train_steps: 1760 +2026-06-19 11:49:05.114 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 11:49:05.254 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:49:05.290 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:49:05.387 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:50:11.433 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000044/group_10/index_3851_seed_25214_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 11:50:15.693 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 11:50:34.213 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:50:34.222 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:50:34.223 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:51:41.784 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000044/group_10/index_3851_seed_25215_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 11:51:44.862 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 11:52:03.363 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:52:06.491 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:52:07.512 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:54:01.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.69s (1.89min) +2026-06-19 11:54:01.910 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:56:01.237 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:03:03.943 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 45, train_steps: 1800 +2026-06-19 12:03:04.563 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 12:03:04.788 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:03:04.917 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:03:05.018 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:04:10.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000045/group_10/index_7148_seed_31524_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 12:04:13.961 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 12:04:32.108 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:04:32.117 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:04:32.117 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:05:40.313 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000045/group_10/index_7148_seed_31525_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 12:05:43.495 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 12:06:02.088 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:06:05.347 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:06:06.340 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:08:01.417 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.05s (1.92min) +2026-06-19 12:08:02.101 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:10:01.383 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:17:03.593 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 46, train_steps: 1840 +2026-06-19 12:17:04.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 12:17:04.455 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:17:04.528 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:17:04.628 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:18:10.656 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_10/index_2254_seed_2399_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 12:18:15.782 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 12:18:34.222 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:18:34.231 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:18:34.231 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:19:40.725 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_10/index_2254_seed_2400_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 12:19:45.673 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 12:20:03.975 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:20:05.333 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:20:06.454 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:22:01.733 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.17s (1.92min) +2026-06-19 12:22:02.442 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:24:01.451 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:31:02.921 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 47, train_steps: 1880 +2026-06-19 12:31:03.608 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 12:31:03.814 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:31:03.958 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:31:04.034 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:32:10.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000047/group_10/index_8008_seed_1298_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 12:32:14.006 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 12:32:32.604 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:32:32.614 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:32:32.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:33:39.681 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000047/group_10/index_8008_seed_1299_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 12:33:43.079 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 12:34:01.521 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:34:03.182 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:34:04.360 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:35:58.331 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.84s (1.90min) +2026-06-19 12:35:58.982 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:37:55.893 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:44:57.683 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 48, train_steps: 1920 +2026-06-19 12:44:58.355 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 12:44:58.536 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:44:58.653 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:44:58.781 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:46:04.403 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_10/index_6335_seed_14835_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 12:46:09.507 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 12:46:27.692 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:46:27.701 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:46:27.701 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:47:34.676 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_10/index_6335_seed_14836_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 12:47:38.948 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 12:47:57.606 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:47:58.931 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:47:59.971 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:49:54.620 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-19 12:49:55.342 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:51:52.291 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:58:52.090 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 49, train_steps: 1960 +2026-06-19 12:58:52.782 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 12:58:52.909 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:58:52.946 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:58:53.016 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:59:58.955 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_10/index_8854_seed_28610_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 13:00:02.077 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 13:00:20.457 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:00:20.466 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:00:20.466 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:01:30.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_10/index_8854_seed_28611_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 13:01:35.332 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 13:01:54.133 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:01:54.369 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:01:55.320 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:03:48.446 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.04s (1.88min) +2026-06-19 13:03:49.149 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:05:45.467 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:12:45.264 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 50, train_steps: 2000 +2026-06-19 13:14:20.936 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=50 (missing=0, unexpected=0) +2026-06-19 13:14:23.203 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 13:14:23.363 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:14:23.424 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:14:23.543 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:15:31.339 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_10/index_8252_seed_27573_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 13:15:36.183 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 13:15:54.166 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:15:54.172 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:15:54.172 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:17:00.928 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_10/index_8252_seed_27574_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 13:17:05.085 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 13:17:23.687 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:17:25.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:17:26.217 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:19:20.081 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.84s (1.90min) +2026-06-19 13:19:20.807 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:21:17.527 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:28:18.474 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 51, train_steps: 2040 +2026-06-19 13:28:19.184 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 13:28:19.293 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:28:19.330 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:28:19.377 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:29:25.582 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000051/group_10/index_9068_seed_9390_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 13:29:30.771 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 13:29:48.576 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:29:48.584 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:29:48.584 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:30:54.134 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000051/group_10/index_9068_seed_9391_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 13:30:58.952 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 13:31:17.067 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:31:18.029 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:31:18.998 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:33:13.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.83s (1.90min) +2026-06-19 13:33:13.741 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:35:13.230 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:42:13.608 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 52, train_steps: 2080 +2026-06-19 13:42:14.276 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 13:42:14.397 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:42:14.434 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:42:14.498 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:43:20.901 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_10/index_3560_seed_22559_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 13:43:25.307 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 13:43:43.789 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:43:43.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:43:43.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:44:50.148 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_10/index_3560_seed_22560_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 13:44:53.117 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 13:45:11.476 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:45:15.287 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:45:16.415 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:47:10.481 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.03s (1.90min) +2026-06-19 13:47:11.197 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:49:11.267 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:56:12.601 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 53, train_steps: 2120 +2026-06-19 13:56:13.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 13:56:13.489 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:56:13.593 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:56:13.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:57:19.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_10/index_8135_seed_7520_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 13:57:22.719 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 13:57:41.906 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:57:41.915 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:57:41.916 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:58:49.098 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_10/index_8135_seed_7521_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 13:58:54.286 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 13:59:12.662 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:59:13.130 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:59:14.067 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:01:08.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.41s (1.91min) +2026-06-19 14:01:09.529 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:03:10.474 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:10:12.157 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 54, train_steps: 2160 +2026-06-19 14:10:12.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 14:10:13.076 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:10:13.185 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:10:13.288 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:11:19.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_10/index_8273_seed_25621_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 14:11:23.113 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 14:11:42.274 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:11:42.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:11:42.285 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:12:48.183 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_10/index_8273_seed_25622_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 14:12:51.375 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 14:13:09.649 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:13:12.866 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:13:13.904 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:15:08.493 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.42s (1.91min) +2026-06-19 14:15:09.218 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:17:09.307 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:24:11.898 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 55, train_steps: 2200 +2026-06-19 14:24:12.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 14:24:12.817 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:24:12.933 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:24:12.996 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:25:19.005 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_10/index_4092_seed_11014_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 14:25:22.059 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 14:25:40.156 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:25:40.166 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:25:40.166 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:26:48.352 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_10/index_4092_seed_11015_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 14:26:51.462 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 14:27:09.566 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:27:12.754 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:27:13.940 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:29:09.888 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.82s (1.93min) +2026-06-19 14:29:10.605 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:31:10.764 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:38:12.103 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 56, train_steps: 2240 +2026-06-19 14:38:12.811 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 14:38:12.977 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:38:13.057 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:38:13.160 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:39:19.136 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_10/index_3911_seed_16465_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 14:39:22.400 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 14:39:41.650 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:39:41.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:39:41.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:40:49.150 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_10/index_3911_seed_16466_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 14:40:52.313 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 14:41:10.746 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:41:13.092 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:41:14.104 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:43:08.781 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.66s (1.91min) +2026-06-19 14:43:09.475 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:45:09.718 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:52:11.838 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 57, train_steps: 2280 +2026-06-19 14:52:12.526 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 14:52:12.646 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:52:12.691 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:52:12.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:53:18.822 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_10/index_2988_seed_32818_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 14:53:23.982 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 14:53:42.831 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:53:42.836 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:53:42.837 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:54:48.385 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_10/index_2988_seed_32819_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 14:54:53.269 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 14:55:11.454 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:55:13.057 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:55:13.948 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:57:08.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.79s (1.91min) +2026-06-19 14:57:09.597 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:59:09.881 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:06:12.293 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 58, train_steps: 2320 +2026-06-19 15:06:12.986 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 15:06:13.199 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:06:13.338 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:06:13.429 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:07:19.371 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_10/index_7859_seed_9774_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 15:07:24.244 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 15:07:42.858 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:07:42.867 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:07:42.868 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:08:48.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_10/index_7859_seed_9775_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 15:08:53.520 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 15:09:11.340 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:09:13.062 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:09:13.937 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:11:08.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.71s (1.91min) +2026-06-19 15:11:09.498 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:13:09.811 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:20:12.051 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 59, train_steps: 2360 +2026-06-19 15:20:12.739 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 15:20:12.907 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:20:12.981 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:20:13.129 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:21:19.227 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_10/index_5582_seed_17073_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 15:21:24.077 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 15:21:42.833 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:21:42.842 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:21:42.845 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:22:48.259 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_10/index_5582_seed_17074_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 15:22:53.216 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 15:23:11.160 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:23:12.372 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:23:13.411 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:25:09.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.28s (1.92min) +2026-06-19 15:25:09.743 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:27:10.367 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:34:12.066 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 60, train_steps: 2400 +2026-06-19 15:34:47.777 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=60 (missing=0, unexpected=0) +2026-06-19 15:34:50.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 15:34:50.762 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:34:50.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:34:50.856 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:35:58.343 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_10/index_126_seed_20836_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 15:36:03.180 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 15:36:20.837 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:36:20.846 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:36:20.846 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:37:28.333 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_10/index_126_seed_20837_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 15:37:33.197 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 15:37:51.033 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:37:53.529 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:37:54.604 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:39:47.707 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-19 15:39:48.402 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:41:45.235 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:48:44.569 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 61, train_steps: 2440 +2026-06-19 15:48:45.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 15:48:45.485 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:48:45.608 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:48:45.705 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:49:51.614 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_10/index_5331_seed_24725_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 15:49:56.400 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 15:50:13.969 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:50:13.979 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:50:13.979 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:51:23.472 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_10/index_5331_seed_24726_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 15:51:27.775 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 15:51:46.108 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:51:47.609 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:51:48.622 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:53:41.681 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.04s (1.88min) +2026-06-19 15:53:42.399 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:55:39.328 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:02:39.003 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 62, train_steps: 2480 +2026-06-19 16:02:39.688 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 16:02:39.874 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:02:39.984 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:02:40.084 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:03:45.828 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000062/group_10/index_304_seed_2935_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 16:03:50.921 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 16:04:09.108 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:04:09.118 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:04:09.118 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:05:19.079 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000062/group_10/index_304_seed_2936_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 16:05:23.255 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 16:05:42.494 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:05:44.132 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:05:45.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:07:38.761 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.59s (1.89min) +2026-06-19 16:07:39.471 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:09:38.855 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:16:39.319 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 63, train_steps: 2520 +2026-06-19 16:16:40.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 16:16:40.134 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:16:40.178 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:16:40.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:17:46.864 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_10/index_8945_seed_12695_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 16:17:51.898 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 16:18:09.852 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:18:09.864 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:18:09.864 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:19:19.954 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_10/index_8945_seed_12696_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 16:19:24.338 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 16:19:42.869 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:19:45.197 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:19:46.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:21:39.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.36s (1.89min) +2026-06-19 16:21:40.303 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:23:38.883 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:30:41.058 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 64, train_steps: 2560 +2026-06-19 16:30:41.626 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 16:30:41.776 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:30:41.822 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:30:41.902 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:31:48.702 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_10/index_9484_seed_7053_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 16:31:51.769 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 16:32:09.964 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:32:09.973 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:32:09.974 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:33:21.049 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_10/index_9484_seed_7054_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 16:33:24.117 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 16:33:42.930 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:33:45.471 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:33:46.471 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:35:40.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.63s (1.89min) +2026-06-19 16:35:40.844 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:37:39.400 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:44:41.298 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 65, train_steps: 2600 +2026-06-19 16:44:41.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 16:44:42.181 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:44:42.308 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:44:42.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:45:49.744 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_10/index_6743_seed_3161_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 16:45:54.823 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 16:46:12.531 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:46:12.540 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:46:12.540 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:47:22.841 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_10/index_6743_seed_3162_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 16:47:27.825 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 16:47:45.857 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:47:47.482 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:47:48.419 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:49:42.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.72s (1.90min) +2026-06-19 16:49:42.959 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:51:40.038 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:58:41.280 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 66, train_steps: 2640 +2026-06-19 16:58:41.990 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 16:58:42.173 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:58:42.273 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:58:42.372 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:59:48.940 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_10/index_453_seed_27434_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 16:59:53.906 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 17:00:12.085 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:00:12.094 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:00:12.094 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:01:19.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_10/index_453_seed_27435_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 17:01:24.531 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 17:01:43.036 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:01:43.599 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:01:44.479 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:03:37.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.34s (1.89min) +2026-06-19 17:03:38.681 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:05:38.385 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:12:40.124 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 67, train_steps: 2680 +2026-06-19 17:12:40.909 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 17:12:41.038 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:12:41.271 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:12:41.409 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:13:48.553 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_10/index_7152_seed_29974_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 17:13:53.453 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 17:14:11.658 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:14:11.667 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:14:11.667 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:15:20.537 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_10/index_7152_seed_29975_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 17:15:25.516 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 17:15:44.068 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:15:44.529 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:15:45.557 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:17:38.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 17:17:39.680 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:19:39.115 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:26:39.552 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 68, train_steps: 2720 +2026-06-19 17:26:40.214 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 17:26:40.318 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:26:40.353 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:26:40.420 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:27:47.141 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_10/index_2283_seed_30565_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 17:27:51.948 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 17:28:10.056 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:28:10.065 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:28:10.066 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:29:20.110 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_10/index_2283_seed_30566_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 17:29:24.898 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 17:29:43.313 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:29:43.886 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:29:44.832 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:31:38.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.76s (1.90min) +2026-06-19 17:31:39.421 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:33:38.460 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:40:38.360 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 69, train_steps: 2760 +2026-06-19 17:40:39.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 17:40:39.209 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:40:39.290 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:40:39.400 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:41:45.323 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000069/group_10/index_8622_seed_29659_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 17:41:50.160 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 17:42:08.029 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:42:08.034 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:42:08.035 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:43:17.123 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000069/group_10/index_8622_seed_29660_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 17:43:22.054 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 17:43:40.079 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:43:41.069 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:43:42.022 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:45:35.854 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.55s (1.89min) +2026-06-19 17:45:36.570 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:47:37.117 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:54:38.567 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 70, train_steps: 2800 +2026-06-19 17:55:14.743 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=70 (missing=0, unexpected=0) +2026-06-19 17:55:16.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 17:55:17.126 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:55:17.158 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:55:17.212 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:56:24.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_10/index_7133_seed_8583_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 17:56:29.085 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 17:56:47.279 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:56:47.286 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:56:47.286 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:57:54.592 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_10/index_7133_seed_8584_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 17:57:59.559 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 17:58:17.358 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:58:18.228 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:58:19.342 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:00:13.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.58s (1.89min) +2026-06-19 18:00:13.838 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:02:13.854 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:09:13.890 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 71, train_steps: 2840 +2026-06-19 18:09:14.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 18:09:14.803 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:09:14.919 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:09:15.026 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:10:21.203 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_10/index_7020_seed_24545_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 18:10:26.037 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 18:10:43.764 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:10:43.774 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:10:43.774 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:11:51.930 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_10/index_7020_seed_24546_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 18:11:56.899 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 18:12:14.966 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:12:16.021 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:12:16.916 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:14:12.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.67s (1.93min) +2026-06-19 18:14:13.460 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:16:12.694 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:23:13.825 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 72, train_steps: 2880 +2026-06-19 18:23:14.506 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 18:23:14.695 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:23:14.808 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:23:14.910 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:24:20.912 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_10/index_5447_seed_16209_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 18:24:25.826 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 18:24:43.393 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:24:43.403 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:24:43.404 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:25:51.253 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_10/index_5447_seed_16210_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 18:25:56.224 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 18:26:14.171 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:26:14.711 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:26:15.753 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:28:10.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.45s (1.91min) +2026-06-19 18:28:11.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:30:10.990 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:37:12.780 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 73, train_steps: 2920 +2026-06-19 18:37:13.456 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 18:37:13.630 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:37:13.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:37:13.832 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:38:20.004 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_10/index_4930_seed_30722_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 18:38:24.311 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 18:38:42.812 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:38:42.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:38:42.824 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:39:49.929 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_10/index_4930_seed_30723_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 18:39:52.983 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 18:40:11.230 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:40:13.250 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:40:14.289 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:42:09.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.16s (1.92min) +2026-06-19 18:42:10.285 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:44:10.446 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:51:12.755 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 74, train_steps: 2960 +2026-06-19 18:51:13.468 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 18:51:13.673 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:51:13.798 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:51:13.912 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:52:19.862 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_10/index_9133_seed_23138_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 18:52:22.832 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 18:52:41.593 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:52:41.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:52:41.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:53:49.825 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_10/index_9133_seed_23139_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 18:53:54.812 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 18:54:12.608 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:54:13.982 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:54:14.915 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:56:10.312 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.21s (1.92min) +2026-06-19 18:56:11.032 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:58:11.445 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:05:14.027 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 75, train_steps: 3000 +2026-06-19 19:06:23.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 19:06:23.426 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:06:23.557 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:06:23.642 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:07:29.995 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000075/group_10/index_6396_seed_881_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 19:07:34.889 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 19:07:52.585 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:07:52.594 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:07:52.594 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:08:59.007 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000075/group_10/index_6396_seed_882_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 19:09:03.205 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 19:09:21.896 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:09:24.209 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:09:25.234 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:11:18.397 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.14s (1.89min) +2026-06-19 19:11:19.101 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:13:16.255 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:20:16.322 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 76, train_steps: 3040 +2026-06-19 19:20:16.989 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 19:20:17.162 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:20:17.241 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:20:17.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:21:23.027 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000076/group_10/index_9338_seed_15040_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 19:21:27.617 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 19:21:45.561 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:21:45.570 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:21:45.570 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:22:53.875 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000076/group_10/index_9338_seed_15041_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 19:22:58.901 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 19:23:16.774 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:23:18.455 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:23:19.562 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:25:13.441 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.66s (1.89min) +2026-06-19 19:25:14.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:27:12.735 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:34:13.259 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 77, train_steps: 3080 +2026-06-19 19:34:13.950 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 19:34:14.122 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:34:14.203 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:34:14.346 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:35:20.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_10/index_9666_seed_14033_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 19:35:24.825 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 19:35:43.078 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:35:43.087 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:35:43.087 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:36:49.437 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_10/index_9666_seed_14034_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 19:36:52.487 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 19:37:10.707 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:37:13.074 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:37:14.088 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:39:08.734 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.60s (1.91min) +2026-06-19 19:39:09.441 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:41:09.772 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:48:12.648 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 78, train_steps: 3120 +2026-06-19 19:48:13.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 19:48:13.549 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:48:13.632 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:48:13.736 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:49:19.736 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_10/index_9373_seed_7347_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 19:49:24.835 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 19:49:42.950 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:49:42.958 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:49:42.959 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:50:50.086 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_10/index_9373_seed_7348_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 19:50:55.005 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 19:51:12.702 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:51:13.323 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:51:14.260 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:53:09.267 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-19 19:53:09.976 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:55:09.986 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:02:13.167 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 79, train_steps: 3160 +2026-06-19 20:02:13.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 20:02:13.984 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:02:14.027 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:02:14.075 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:03:20.385 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000079/group_10/index_8424_seed_25857_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 20:03:25.313 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 20:03:42.857 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:03:42.865 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:03:42.865 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:04:50.236 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000079/group_10/index_8424_seed_25858_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 20:04:55.114 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 20:05:12.967 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:05:13.093 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:05:13.986 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:07:09.660 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.53s (1.93min) +2026-06-19 20:07:10.385 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:09:11.035 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:16:13.941 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 80, train_steps: 3200 +2026-06-19 20:16:51.422 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=80 (missing=0, unexpected=0) +2026-06-19 20:16:53.093 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 20:16:53.297 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:16:53.425 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:16:53.528 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:18:02.109 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_10/index_8092_seed_30550_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 20:18:07.027 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 20:18:24.927 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:18:24.936 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:18:24.936 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:19:31.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_10/index_8092_seed_30551_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 20:19:35.934 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 20:19:53.903 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:19:54.752 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:19:55.688 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:21:49.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-19 20:21:50.363 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:23:47.270 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:30:47.423 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 81, train_steps: 3240 +2026-06-19 20:30:48.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 20:30:48.320 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:30:48.408 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:30:48.545 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:31:54.751 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_10/index_4405_seed_25282_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 20:31:59.594 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 20:32:17.713 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:32:17.723 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:32:17.723 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:33:25.123 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_10/index_4405_seed_25283_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 20:33:30.057 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 20:33:48.535 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:33:49.273 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:33:50.159 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:35:44.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.70s (1.89min) +2026-06-19 20:35:44.752 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:37:41.679 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:44:40.002 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 82, train_steps: 3280 +2026-06-19 20:44:40.696 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 20:44:40.816 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:44:40.856 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:44:40.929 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:45:46.928 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_10/index_9876_seed_19539_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 20:45:51.387 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 20:46:09.844 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:46:09.852 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:46:09.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:47:20.110 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_10/index_9876_seed_19540_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 20:47:23.256 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 20:47:42.921 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:47:44.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:47:45.139 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:49:38.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.42s (1.89min) +2026-06-19 20:49:39.275 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:51:38.424 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:58:38.984 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 83, train_steps: 3320 +2026-06-19 20:58:39.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 20:58:40.023 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:58:40.129 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:58:40.244 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:59:46.234 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_10/index_6785_seed_26600_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 20:59:51.395 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 21:00:09.495 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:00:09.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:00:09.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:01:19.854 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_10/index_6785_seed_26601_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 21:01:24.483 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 21:01:43.020 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:01:44.304 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:01:45.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:03:38.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.28s (1.89min) +2026-06-19 21:03:39.340 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:05:38.819 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:12:38.919 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 84, train_steps: 3360 +2026-06-19 21:12:39.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 21:12:39.806 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:12:39.873 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:12:40.012 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:13:46.115 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_10/index_8621_seed_30037_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 21:13:50.307 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 21:14:08.460 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:14:08.469 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:14:08.470 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:15:19.744 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_10/index_8621_seed_30038_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 21:15:23.770 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 21:15:42.854 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:15:43.791 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:15:44.811 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:17:38.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.31s (1.89min) +2026-06-19 21:17:38.864 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:19:38.526 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:26:39.574 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 85, train_steps: 3400 +2026-06-19 21:26:40.168 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 21:26:40.365 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:26:40.484 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:26:40.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:27:47.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_10/index_1714_seed_242_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 21:27:52.367 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 21:28:10.223 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:28:10.228 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:28:10.229 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:29:20.143 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_10/index_1714_seed_243_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 21:29:25.054 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 21:29:43.093 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:29:43.697 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:29:44.632 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:31:38.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-19 21:31:38.988 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:33:38.757 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:40:38.641 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 86, train_steps: 3440 +2026-06-19 21:40:39.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 21:40:39.522 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:40:39.583 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:40:39.715 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:41:46.087 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_10/index_3534_seed_28109_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 21:41:50.896 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 21:42:08.630 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:42:08.639 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:42:08.639 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:43:18.855 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_10/index_3534_seed_28110_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 21:43:22.808 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 21:43:41.889 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:43:42.867 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:43:43.872 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:45:37.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 21:45:37.987 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:47:37.710 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:54:39.052 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 87, train_steps: 3480 +2026-06-19 21:54:39.743 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 21:54:39.947 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:54:40.047 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:54:40.148 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:55:46.019 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000087/group_10/index_5360_seed_31137_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 21:55:51.157 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 21:56:09.137 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:56:09.148 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:56:09.148 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:57:19.094 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000087/group_10/index_5360_seed_31138_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 21:57:24.093 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 21:57:42.887 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:57:43.169 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:57:44.101 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:59:37.604 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.39s (1.89min) +2026-06-19 21:59:38.318 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:01:38.556 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:08:39.980 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 88, train_steps: 3520 +2026-06-19 22:08:40.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 22:08:40.786 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:08:40.818 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:08:40.865 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:09:47.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_10/index_4914_seed_10258_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 22:09:51.905 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 22:10:09.833 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:10:09.843 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:10:09.843 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:11:19.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_10/index_4914_seed_10259_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 22:11:24.355 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 22:11:42.871 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:11:43.755 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:11:44.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:13:38.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.57s (1.89min) +2026-06-19 22:13:39.104 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:15:38.306 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:22:38.699 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 89, train_steps: 3560 +2026-06-19 22:22:39.424 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 22:22:39.531 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:22:39.573 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:22:39.626 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:23:45.830 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_10/index_7127_seed_12253_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 22:23:50.811 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 22:24:08.420 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:24:08.429 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:24:08.430 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:25:19.753 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_10/index_7127_seed_12254_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 22:25:23.955 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 22:25:42.860 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:25:43.683 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:25:44.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:27:38.162 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 22:27:38.887 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:29:38.496 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:36:39.423 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 90, train_steps: 3600 +2026-06-19 22:37:15.395 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=90 (missing=0, unexpected=0) +2026-06-19 22:37:17.307 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 22:37:17.485 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:37:17.590 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:37:17.689 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:38:24.780 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_10/index_4521_seed_8846_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 22:38:27.757 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 22:38:45.795 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:38:45.805 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:38:45.805 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:39:55.000 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_10/index_4521_seed_8847_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 22:40:00.224 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 22:40:17.857 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:40:19.241 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:40:20.216 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:42:14.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.70s (1.89min) +2026-06-19 22:42:14.863 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:44:13.689 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:51:14.848 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 91, train_steps: 3640 +2026-06-19 22:51:15.700 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 22:51:15.906 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:51:16.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:51:16.115 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:52:22.154 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000091/group_10/index_8991_seed_25528_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 22:52:26.903 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 22:52:44.557 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:52:44.566 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:52:44.567 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:53:51.916 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000091/group_10/index_8991_seed_25529_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 22:53:56.154 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 22:54:14.941 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:54:15.937 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:54:16.984 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:56:12.672 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.67s (1.93min) +2026-06-19 22:56:13.389 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:58:13.114 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:05:14.267 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 92, train_steps: 3680 +2026-06-19 23:05:15.117 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 23:05:15.225 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:05:15.279 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:05:15.335 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:06:21.329 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000092/group_10/index_5785_seed_14633_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 23:06:24.471 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 23:06:43.047 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:06:43.054 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:06:43.054 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:07:51.220 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000092/group_10/index_5785_seed_14634_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 23:07:55.519 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 23:08:14.063 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:08:16.058 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:08:17.089 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:10:11.751 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.62s (1.91min) +2026-06-19 23:10:12.471 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:12:11.284 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:19:13.485 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 93, train_steps: 3720 +2026-06-19 23:19:14.152 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 23:19:14.281 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:19:14.316 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:19:14.360 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:20:20.409 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_10/index_272_seed_1483_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 23:20:23.401 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 23:20:43.106 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:20:43.116 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:20:43.117 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:21:50.300 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_10/index_272_seed_1484_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 23:21:53.361 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 23:22:11.426 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:22:14.354 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:22:15.558 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:24:10.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.33s (1.92min) +2026-06-19 23:24:11.708 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:26:11.286 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:33:13.934 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 94, train_steps: 3760 +2026-06-19 23:33:14.640 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 23:33:14.836 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:33:14.942 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:33:15.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:34:21.079 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_10/index_1392_seed_1461_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 23:34:26.181 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 23:34:44.267 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:34:44.275 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:34:44.275 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:35:51.610 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_10/index_1392_seed_1462_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 23:35:56.507 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 23:36:14.298 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:36:16.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:36:17.022 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:38:12.062 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.89s (1.91min) +2026-06-19 23:38:12.784 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:40:11.321 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:47:11.866 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 95, train_steps: 3800 +2026-06-19 23:47:12.588 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-19 23:47:12.688 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:47:12.729 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:47:12.757 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:48:18.757 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_10/index_914_seed_8276_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-19 23:48:22.824 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 23:48:42.467 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:48:42.477 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:48:42.477 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:49:48.409 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_10/index_914_seed_8277_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-19 23:49:53.533 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-19 23:50:11.372 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:50:12.004 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:50:13.164 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:52:08.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.82s (1.91min) +2026-06-19 23:52:08.791 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:54:07.677 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:01:09.847 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 96, train_steps: 3840 +2026-06-20 00:01:10.475 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 00:01:10.685 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:01:10.902 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:01:11.001 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:02:17.218 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_10/index_7475_seed_13413_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 00:02:21.968 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 00:02:40.155 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:02:40.164 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:02:40.165 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:03:47.553 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_10/index_7475_seed_13414_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 00:03:52.452 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 00:04:10.457 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:04:11.783 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:04:12.828 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:06:07.928 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 00:06:08.644 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:08:07.689 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:15:08.500 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 97, train_steps: 3880 +2026-06-20 00:15:09.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 00:15:09.406 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:15:09.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:15:09.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:16:15.523 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_10/index_4653_seed_9016_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 00:16:20.339 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 00:16:38.475 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:16:38.483 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:16:38.483 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:17:47.346 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_10/index_4653_seed_9017_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 00:17:51.364 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 00:18:09.638 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:18:11.238 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:18:12.353 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:20:07.671 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.20s (1.92min) +2026-06-20 00:20:08.401 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:22:07.128 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:29:07.006 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 98, train_steps: 3920 +2026-06-20 00:29:07.693 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 00:29:07.903 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:29:08.026 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:29:08.109 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:30:13.987 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000098/group_10/index_2243_seed_10800_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 00:30:19.022 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 00:30:37.220 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:30:37.226 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:30:37.226 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:31:46.313 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000098/group_10/index_2243_seed_10801_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 00:31:50.426 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 00:32:08.865 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:32:10.419 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:32:11.452 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:34:06.389 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 00:34:07.118 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:36:03.853 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:43:05.346 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 99, train_steps: 3960 +2026-06-20 00:43:06.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 00:43:06.152 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:43:06.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:43:06.263 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:44:12.341 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_10/index_6015_seed_30656_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 00:44:17.470 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 00:44:35.955 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:44:35.963 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:44:35.964 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:45:46.389 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_10/index_6015_seed_30657_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 00:45:50.601 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 00:46:09.600 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:46:09.899 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:46:10.951 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:48:05.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.75s (1.91min) +2026-06-20 00:48:06.441 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:50:03.400 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:57:04.803 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 100, train_steps: 4000 +2026-06-20 00:58:47.851 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=100 (missing=0, unexpected=0) +2026-06-20 00:58:49.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 00:58:49.737 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:58:49.817 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:58:49.937 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:59:57.553 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_10/index_86_seed_8999_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 01:00:02.612 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 01:00:20.580 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:00:20.588 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:00:20.588 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:01:26.880 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_10/index_86_seed_9000_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 01:01:30.838 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 01:01:49.329 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:01:50.303 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:01:51.335 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:03:44.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-20 01:03:45.435 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:05:42.221 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:12:39.868 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 101, train_steps: 4040 +2026-06-20 01:12:40.567 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 01:12:40.667 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:12:40.696 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:12:40.722 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:13:47.840 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_10/index_5828_seed_31693_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 01:13:53.000 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 01:14:12.637 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:14:12.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:14:12.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:15:20.265 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_10/index_5828_seed_31694_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 01:15:24.323 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 01:15:43.314 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:15:43.499 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:15:44.539 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:17:38.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.62s (1.89min) +2026-06-20 01:17:38.932 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:19:38.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:26:37.798 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 102, train_steps: 4080 +2026-06-20 01:26:38.504 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 01:26:38.628 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:26:38.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:26:38.730 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:27:44.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_10/index_3990_seed_583_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 01:27:47.731 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 01:28:06.315 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:28:06.324 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:28:06.325 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:29:17.299 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_10/index_3990_seed_584_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 01:29:22.490 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 01:29:40.377 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:29:41.183 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:29:42.586 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:31:36.383 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-20 01:31:37.098 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:33:37.226 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:40:38.536 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 103, train_steps: 4120 +2026-06-20 01:40:39.244 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 01:40:39.383 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:40:39.419 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:40:39.536 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:41:45.737 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000103/group_10/index_3404_seed_27825_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 01:41:50.641 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 01:42:08.985 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:42:08.994 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:42:08.994 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:43:19.420 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000103/group_10/index_3404_seed_27826_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 01:43:24.335 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 01:43:42.874 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:43:43.334 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:43:44.248 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:45:37.767 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-20 01:45:38.488 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:47:38.266 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:54:37.783 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 104, train_steps: 4160 +2026-06-20 01:54:38.703 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 01:54:38.864 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:54:38.956 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:54:39.079 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:55:45.379 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_10/index_9170_seed_33428_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 01:55:49.519 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 01:56:08.095 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:56:08.104 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:56:08.104 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:57:19.000 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_10/index_9170_seed_33429_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 01:57:21.900 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 01:57:40.254 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:57:42.889 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:57:43.912 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:59:37.314 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-20 01:59:38.031 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:01:38.022 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:08:38.520 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 105, train_steps: 4200 +2026-06-20 02:08:39.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 02:08:39.266 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:08:39.291 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:08:39.336 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:09:45.737 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000105/group_10/index_7285_seed_12405_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 02:09:48.845 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 02:10:07.861 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:10:07.871 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:10:07.874 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:11:19.101 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000105/group_10/index_7285_seed_12406_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 02:11:22.093 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 02:11:41.626 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:11:43.945 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:11:44.954 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:13:38.217 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 02:13:38.926 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:15:38.270 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:22:37.864 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 106, train_steps: 4240 +2026-06-20 02:22:38.565 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 02:22:38.750 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:22:38.839 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:22:38.955 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:23:44.903 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000106/group_10/index_7404_seed_2793_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 02:23:47.857 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 02:24:06.576 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:24:06.585 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:24:06.586 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:25:16.777 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000106/group_10/index_7404_seed_2794_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 02:25:19.781 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 02:25:38.110 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:25:40.847 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:25:41.978 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:27:35.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.90s (1.90min) +2026-06-20 02:27:36.583 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:29:36.814 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:36:37.092 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 107, train_steps: 4280 +2026-06-20 02:36:37.801 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 02:36:37.960 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:36:38.030 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:36:38.166 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:37:44.088 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_10/index_152_seed_11486_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 02:37:49.282 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 02:38:07.174 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:38:07.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:38:07.183 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:39:15.312 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_10/index_152_seed_11487_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 02:39:19.705 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 02:39:38.214 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:39:39.229 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:39:40.458 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:41:35.347 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 02:41:36.072 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:43:35.147 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:50:35.012 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 108, train_steps: 4320 +2026-06-20 02:50:35.726 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 02:50:35.885 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:50:35.936 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:50:36.039 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:51:41.959 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_10/index_4180_seed_26254_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 02:51:45.084 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 02:52:03.619 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:52:03.626 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:52:03.626 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:53:13.170 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_10/index_4180_seed_26255_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 02:53:16.303 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 02:53:35.037 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:53:37.574 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:53:38.607 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:55:33.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.45s (1.91min) +2026-06-20 02:55:33.973 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:57:30.650 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:04:30.700 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 109, train_steps: 4360 +2026-06-20 03:04:31.405 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 03:04:31.579 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:04:31.676 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:04:31.788 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:05:37.549 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000109/group_10/index_6612_seed_432_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 03:05:43.087 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 03:06:01.409 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:06:01.418 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:06:01.419 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:07:09.671 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000109/group_10/index_6612_seed_433_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 03:07:14.037 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 03:07:32.722 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:07:33.407 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:07:34.459 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:09:29.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-20 03:09:30.015 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:11:26.576 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:18:25.325 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 110, train_steps: 4400 +2026-06-20 03:19:01.744 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=110 (missing=0, unexpected=0) +2026-06-20 03:19:03.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 03:19:03.461 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:19:03.515 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:19:03.580 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:20:11.787 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_10/index_6748_seed_22277_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 03:20:14.946 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 03:20:34.275 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:20:34.282 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:20:34.282 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:21:43.373 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_10/index_6748_seed_22278_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 03:21:48.508 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 03:22:07.516 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:22:07.524 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:22:08.550 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:24:03.966 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.41s (1.92min) +2026-06-20 03:24:04.691 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:26:02.475 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:33:05.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 111, train_steps: 4440 +2026-06-20 03:33:06.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 03:33:06.182 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:33:06.224 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:33:06.271 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:34:12.338 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_10/index_8509_seed_17525_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 03:34:17.160 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 03:34:36.062 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:34:36.071 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:34:36.071 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:35:46.718 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_10/index_8509_seed_17526_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 03:35:49.978 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 03:36:08.655 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:36:10.098 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:36:11.150 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:38:06.021 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 03:38:06.732 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:40:04.483 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:47:04.318 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 112, train_steps: 4480 +2026-06-20 03:47:05.043 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 03:47:05.216 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:47:05.290 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:47:05.405 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:48:11.217 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_10/index_4564_seed_22313_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 03:48:14.327 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 03:48:32.754 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:48:32.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:48:32.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:49:40.980 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_10/index_4564_seed_22314_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 03:49:44.125 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 03:50:02.491 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:50:04.983 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:50:06.294 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:52:00.943 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-20 03:52:01.659 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:54:00.965 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:01:04.009 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 113, train_steps: 4520 +2026-06-20 04:01:04.874 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 04:01:05.001 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:01:05.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:01:05.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:02:10.970 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000113/group_10/index_5681_seed_32724_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 04:02:14.082 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 04:02:32.535 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:02:32.544 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:02:32.544 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:03:42.105 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000113/group_10/index_5681_seed_32725_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 04:03:47.288 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 04:04:05.806 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:04:06.611 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:04:07.550 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:06:02.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.15s (1.92min) +2026-06-20 04:06:03.532 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:08:01.584 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:15:03.777 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 114, train_steps: 4560 +2026-06-20 04:15:04.475 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 04:15:04.697 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:15:04.817 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:15:04.934 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:16:10.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000114/group_10/index_9487_seed_31323_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 04:16:14.809 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 04:16:33.458 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:16:33.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:16:33.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:17:40.664 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000114/group_10/index_9487_seed_31324_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 04:17:43.881 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 04:18:02.575 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:18:04.152 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:18:05.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:20:00.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.72s (1.91min) +2026-06-20 04:20:00.895 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:22:00.102 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:29:01.470 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 115, train_steps: 4600 +2026-06-20 04:29:02.193 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 04:29:02.409 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:29:02.528 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:29:02.607 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:30:08.457 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_10/index_2646_seed_7991_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 04:30:11.703 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 04:30:30.032 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:30:30.042 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:30:30.042 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:31:39.057 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_10/index_2646_seed_7992_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 04:31:42.569 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 04:32:00.812 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:32:02.494 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:32:03.513 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:33:58.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.33s (1.91min) +2026-06-20 04:33:58.776 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:35:55.626 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:42:56.470 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 116, train_steps: 4640 +2026-06-20 04:42:57.162 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 04:42:57.366 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:42:57.479 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:42:57.558 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:44:03.675 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_10/index_4551_seed_2408_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 04:44:06.883 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 04:44:25.446 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:44:25.456 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:44:25.456 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:45:34.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_10/index_4551_seed_2409_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 04:45:37.469 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 04:45:55.763 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:45:58.381 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:45:59.416 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:47:53.698 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.25s (1.90min) +2026-06-20 04:47:54.408 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:49:50.891 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:56:50.233 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 117, train_steps: 4680 +2026-06-20 04:56:50.953 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 04:56:51.080 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:56:51.106 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:56:51.153 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:57:57.368 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_10/index_5036_seed_24383_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 04:58:02.528 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 04:58:21.668 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:58:21.677 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:58:21.677 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:59:28.112 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_10/index_5036_seed_24384_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 04:59:32.947 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 04:59:51.106 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:59:51.798 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:59:52.696 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:01:46.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.49s (1.89min) +2026-06-20 05:01:47.078 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:03:44.145 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:10:42.585 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 118, train_steps: 4720 +2026-06-20 05:10:43.295 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 05:10:43.396 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:10:43.434 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:10:43.481 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:11:50.757 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000118/group_10/index_658_seed_14052_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 05:11:55.620 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 05:12:13.784 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:12:13.794 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:12:13.794 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:13:23.142 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000118/group_10/index_658_seed_14053_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 05:13:28.118 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 05:13:46.614 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:13:46.957 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:13:47.864 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:15:41.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.26s (1.89min) +2026-06-20 05:15:42.001 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:17:38.996 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:24:37.901 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 119, train_steps: 4760 +2026-06-20 05:24:38.622 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 05:24:38.762 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:24:38.801 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:24:38.891 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:25:44.645 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_10/index_5869_seed_12390_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 05:25:49.483 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 05:26:07.365 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:26:07.375 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:26:07.375 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:27:15.096 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_10/index_5869_seed_12391_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 05:27:19.340 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 05:27:37.865 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:27:40.119 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:27:41.315 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:29:34.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-20 05:29:35.560 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:31:33.106 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:38:32.148 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 120, train_steps: 4800 +2026-06-20 05:39:08.648 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=120 (missing=0, unexpected=0) +2026-06-20 05:39:10.883 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 05:39:11.042 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:39:11.122 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:39:11.209 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:40:19.287 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000120/group_10/index_2037_seed_8297_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 05:40:22.468 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 05:40:42.889 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:40:42.896 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:40:42.896 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:41:48.680 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000120/group_10/index_2037_seed_8298_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 05:41:51.918 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 05:42:10.630 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:42:12.566 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:42:13.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:44:08.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.96s (1.92min) +2026-06-20 05:44:09.427 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:46:10.089 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:53:12.064 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 121, train_steps: 4840 +2026-06-20 05:53:12.775 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 05:53:12.886 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:53:12.924 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:53:12.973 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:54:19.461 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_10/index_6300_seed_27432_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 05:54:22.605 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 05:54:42.245 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:54:42.254 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:54:42.254 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:55:48.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_10/index_6300_seed_27433_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 05:55:51.653 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 05:56:10.328 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:56:13.156 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:56:14.185 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:58:09.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.97s (1.92min) +2026-06-20 05:58:10.076 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:00:10.577 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:07:11.960 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 122, train_steps: 4880 +2026-06-20 06:07:12.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 06:07:12.821 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:07:12.892 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:07:12.993 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:08:19.065 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_10/index_7601_seed_4029_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 06:08:22.341 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 06:08:41.593 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:08:41.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:08:41.604 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:09:49.265 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_10/index_7601_seed_4030_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 06:09:52.440 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 06:10:11.071 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:10:13.195 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:10:14.222 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:12:08.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.25s (1.90min) +2026-06-20 06:12:09.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:14:10.582 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:21:11.869 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 123, train_steps: 4920 +2026-06-20 06:21:12.561 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 06:21:12.706 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:21:12.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:21:12.831 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:22:18.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_10/index_3632_seed_10966_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 06:22:21.768 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 06:22:40.132 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:22:40.140 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:22:40.141 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:23:48.026 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_10/index_3632_seed_10967_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 06:23:53.316 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 06:24:11.659 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:24:12.071 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:24:13.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:26:08.788 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.47s (1.92min) +2026-06-20 06:26:09.512 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:28:09.836 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:35:11.851 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 124, train_steps: 4960 +2026-06-20 06:35:12.539 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 06:35:12.719 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:35:12.805 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:35:12.905 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:36:18.700 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_10/index_9154_seed_8438_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 06:36:23.540 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 06:36:42.174 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:36:42.183 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:36:42.183 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:37:48.367 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_10/index_9154_seed_8439_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 06:37:52.560 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 06:38:10.871 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:38:12.129 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:38:13.240 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:40:08.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.44s (1.92min) +2026-06-20 06:40:09.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:42:09.776 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:49:11.653 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 125, train_steps: 5000 +2026-06-20 06:50:15.220 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 06:50:15.404 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:50:15.517 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:50:15.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:51:21.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_10/index_4264_seed_24790_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 06:51:24.675 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 06:51:42.980 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:51:42.990 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:51:42.990 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:52:51.371 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_10/index_4264_seed_24791_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 06:52:56.725 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 06:53:15.464 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:53:15.926 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:53:16.965 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:55:11.654 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.68s (1.91min) +2026-06-20 06:55:12.372 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:57:11.765 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:04:11.921 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 126, train_steps: 5040 +2026-06-20 07:04:12.630 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 07:04:12.733 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:04:12.767 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:04:12.796 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:05:18.977 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_10/index_349_seed_27040_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 07:05:22.113 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 07:05:40.146 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:05:40.155 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:05:40.155 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:06:49.437 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_10/index_349_seed_27041_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 07:06:54.636 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 07:07:12.912 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:07:13.988 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:07:15.166 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:09:10.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.08s (1.92min) +2026-06-20 07:09:11.147 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:11:11.166 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:18:11.923 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 127, train_steps: 5080 +2026-06-20 07:18:12.630 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 07:18:12.829 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:18:12.928 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:18:13.022 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:19:18.787 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_10/index_1097_seed_12859_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 07:19:23.591 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 07:19:42.751 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:19:42.760 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:19:42.761 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:20:49.190 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_10/index_1097_seed_12860_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 07:20:53.982 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 07:21:12.119 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:21:13.050 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:21:13.936 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:23:09.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.50s (1.93min) +2026-06-20 07:23:10.327 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:25:10.159 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:32:11.165 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 128, train_steps: 5120 +2026-06-20 07:32:11.885 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 07:32:12.094 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:32:12.206 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:32:12.300 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:33:18.019 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_10/index_6559_seed_5231_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 07:33:22.787 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 07:33:41.978 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:33:41.984 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:33:41.987 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:34:49.315 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_10/index_6559_seed_5232_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 07:34:53.467 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 07:35:12.353 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:35:13.016 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:35:14.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:37:08.641 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.56s (1.91min) +2026-06-20 07:37:09.373 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:39:09.932 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:46:11.992 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 129, train_steps: 5160 +2026-06-20 07:46:12.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 07:46:12.883 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:46:12.964 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:46:13.075 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:47:19.481 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_10/index_2350_seed_27546_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 07:47:22.559 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 07:47:41.950 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:47:41.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:47:41.969 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:48:49.106 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_10/index_2350_seed_27547_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 07:48:52.293 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 07:49:10.729 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:49:13.070 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:49:14.093 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:51:09.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-20 07:51:09.758 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:53:10.305 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:00:11.779 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 130, train_steps: 5200 +2026-06-20 08:00:48.933 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=130 (missing=0, unexpected=0) +2026-06-20 08:00:51.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 08:00:51.378 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:00:51.401 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:00:51.449 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:02:00.417 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_10/index_2413_seed_14735_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 08:02:05.528 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 08:02:23.606 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:02:23.612 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:02:23.612 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:03:29.091 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_10/index_2413_seed_14736_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 08:03:33.454 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 08:03:51.680 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:03:52.509 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:03:53.538 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:05:47.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.72s (1.90min) +2026-06-20 08:05:48.018 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:07:44.847 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:14:45.321 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 131, train_steps: 5240 +2026-06-20 08:14:46.027 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 08:14:46.246 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:14:46.360 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:14:46.437 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:15:52.403 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_10/index_9686_seed_32976_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 08:15:57.496 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 08:16:15.455 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:16:15.463 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:16:15.463 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:17:23.341 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_10/index_9686_seed_32977_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 08:17:27.568 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 08:17:46.205 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:17:46.990 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:17:48.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:19:41.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.38s (1.89min) +2026-06-20 08:19:42.139 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:21:39.467 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:28:37.838 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 132, train_steps: 5280 +2026-06-20 08:28:38.520 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 08:28:38.687 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:28:38.763 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:28:38.859 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:29:45.302 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000132/group_10/index_597_seed_3387_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 08:29:50.476 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 08:30:08.167 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:30:08.175 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:30:08.176 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:31:15.602 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000132/group_10/index_597_seed_3388_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 08:31:19.636 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 08:31:38.028 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:31:40.849 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:31:41.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:33:35.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.44s (1.89min) +2026-06-20 08:33:36.119 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:35:35.264 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:42:34.335 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 133, train_steps: 5320 +2026-06-20 08:42:35.016 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 08:42:35.115 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:42:35.158 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:42:35.204 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:43:41.513 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_10/index_2761_seed_31391_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 08:43:44.654 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 08:44:02.721 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:44:02.731 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:44:02.731 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:45:11.937 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_10/index_2761_seed_31392_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 08:45:15.077 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 08:45:33.491 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:45:35.888 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:45:36.912 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:47:32.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.77s (1.93min) +2026-06-20 08:47:33.450 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:49:30.693 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:56:29.429 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 134, train_steps: 5360 +2026-06-20 08:56:30.114 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 08:56:30.271 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:56:30.328 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:56:30.437 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:57:36.560 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000134/group_10/index_7801_seed_23892_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 08:57:39.760 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 08:58:00.218 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:58:00.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:58:00.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:59:07.546 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000134/group_10/index_7801_seed_23893_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 08:59:12.820 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 08:59:31.039 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:59:32.041 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:59:33.083 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:01:26.319 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 09:01:27.048 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:03:23.807 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:10:23.056 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 135, train_steps: 5400 +2026-06-20 09:10:23.673 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 09:10:23.767 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:10:23.813 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:10:23.860 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:11:30.344 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000135/group_10/index_5296_seed_16232_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 09:11:34.550 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 09:11:52.717 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:11:52.723 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:11:52.723 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:13:00.408 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000135/group_10/index_5296_seed_16233_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 09:13:05.572 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 09:13:23.706 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:13:23.932 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:13:24.970 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:15:18.588 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.61s (1.89min) +2026-06-20 09:15:19.311 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:17:17.102 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:24:16.800 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 136, train_steps: 5440 +2026-06-20 09:24:17.531 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 09:24:17.655 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:24:17.697 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:24:17.768 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:25:23.960 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_10/index_5196_seed_8616_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 09:25:28.897 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 09:25:46.816 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:25:46.825 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:25:46.827 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:26:54.229 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_10/index_5196_seed_8617_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 09:26:58.493 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 09:27:17.535 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:27:18.656 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:27:19.702 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:29:12.959 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.15s (1.89min) +2026-06-20 09:29:13.686 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:31:13.179 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:38:13.219 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 137, train_steps: 5480 +2026-06-20 09:38:13.937 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 09:38:14.091 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:38:14.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:38:14.249 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:39:20.224 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000137/group_10/index_1028_seed_30746_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 09:39:23.282 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 09:39:42.189 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:39:42.200 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:39:42.200 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:40:50.322 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000137/group_10/index_1028_seed_30747_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 09:40:53.629 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 09:41:12.623 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:41:14.371 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:41:15.562 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:43:09.522 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-20 09:43:10.251 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:45:10.505 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:52:12.152 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 138, train_steps: 5520 +2026-06-20 09:52:12.866 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 09:52:13.020 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:52:13.087 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:52:13.180 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:53:19.367 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_10/index_8145_seed_3852_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 09:53:22.493 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 09:53:41.895 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:53:41.905 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:53:41.905 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:54:49.529 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_10/index_8145_seed_3853_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 09:54:52.741 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 09:55:11.162 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:55:13.383 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:55:14.620 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:57:10.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.46s (1.92min) +2026-06-20 09:57:10.921 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:59:11.153 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:06:12.676 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 139, train_steps: 5560 +2026-06-20 10:06:13.362 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 10:06:13.485 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:06:13.526 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:06:13.593 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:07:19.738 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_10/index_6756_seed_27372_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 10:07:24.912 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 10:07:42.872 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:07:42.881 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:07:42.881 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:08:50.409 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_10/index_6756_seed_27373_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 10:08:54.557 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 10:09:13.093 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:09:13.264 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:09:14.306 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:11:08.675 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.20s (1.90min) +2026-06-20 10:11:09.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:13:09.783 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:20:12.675 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 140, train_steps: 5600 +2026-06-20 10:20:50.096 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=140 (missing=0, unexpected=0) +2026-06-20 10:20:51.812 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 10:20:51.939 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:20:51.973 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:20:52.055 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:21:59.680 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000140/group_10/index_4372_seed_8031_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 10:22:04.889 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 10:22:23.097 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:22:23.106 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:22:23.106 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:23:29.017 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000140/group_10/index_4372_seed_8032_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 10:23:33.980 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 10:23:52.018 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:23:53.040 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:23:53.932 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:25:47.600 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-20 10:25:48.313 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:27:45.365 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:34:43.998 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 141, train_steps: 5640 +2026-06-20 10:34:44.690 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 10:34:44.875 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:34:44.977 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:34:45.075 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:35:50.948 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_10/index_1036_seed_17035_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 10:35:55.311 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 10:36:13.797 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:36:13.806 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:36:13.806 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:37:23.082 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_10/index_1036_seed_17036_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 10:37:27.275 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 10:37:45.960 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:37:46.846 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:37:47.898 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:39:41.415 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.51s (1.89min) +2026-06-20 10:39:42.149 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:41:39.464 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:48:38.468 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 142, train_steps: 5680 +2026-06-20 10:48:39.157 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 10:48:39.317 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:48:39.381 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:48:39.464 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:49:45.596 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000142/group_10/index_4839_seed_2253_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 10:49:50.520 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 10:50:08.293 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:50:08.303 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:50:08.303 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:51:19.238 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000142/group_10/index_4839_seed_2254_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 10:51:23.528 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 10:51:42.866 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:51:43.153 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:51:44.187 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:53:37.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-20 10:53:37.851 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:55:37.995 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:02:37.999 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 143, train_steps: 5720 +2026-06-20 11:02:38.630 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 11:02:38.734 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:02:38.770 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:02:38.818 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:03:45.069 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000143/group_10/index_7268_seed_17943_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 11:03:50.013 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 11:04:07.626 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:04:07.635 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:04:07.636 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:05:18.374 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000143/group_10/index_7268_seed_17944_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 11:05:23.472 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 11:05:42.159 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:05:42.169 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:05:43.266 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:07:36.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 11:07:37.313 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:09:36.845 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:16:37.092 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 144, train_steps: 5760 +2026-06-20 11:16:37.800 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 11:16:38.019 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:16:38.138 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:16:38.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:17:44.086 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_10/index_2016_seed_27722_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 11:17:48.238 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 11:18:06.428 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:18:06.437 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:18:06.437 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:19:13.296 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_10/index_2016_seed_27723_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 11:19:16.456 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 11:19:34.811 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:19:38.707 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:19:39.841 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:21:33.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.47s (1.89min) +2026-06-20 11:21:34.087 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:23:30.887 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:30:30.411 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 145, train_steps: 5800 +2026-06-20 11:30:31.098 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 11:30:31.282 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:30:31.382 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:30:31.479 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:31:37.159 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000145/group_10/index_7813_seed_3106_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 11:31:40.299 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 11:32:00.635 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:32:00.654 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:32:00.655 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:33:08.858 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000145/group_10/index_7813_seed_3107_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 11:33:13.958 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 11:33:32.641 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:33:33.207 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:33:34.252 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:35:28.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.89s (1.90min) +2026-06-20 11:35:28.901 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:37:25.684 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:44:23.920 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 146, train_steps: 5840 +2026-06-20 11:44:24.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 11:44:24.726 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:44:24.771 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:44:24.819 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:45:30.894 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_10/index_4204_seed_4685_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 11:45:33.785 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 11:45:51.783 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:45:51.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:45:51.794 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:46:59.241 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_10/index_4204_seed_4686_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 11:47:04.339 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 11:47:22.170 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:47:23.247 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:47:24.201 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:49:19.133 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-20 11:49:19.856 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:51:17.041 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:58:15.916 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 147, train_steps: 5880 +2026-06-20 11:58:16.618 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 11:58:16.822 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:58:16.926 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:58:17.005 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:59:23.507 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_10/index_3032_seed_18063_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 11:59:28.407 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 11:59:46.211 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:59:46.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:59:46.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:00:52.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_10/index_3032_seed_18064_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 12:00:56.892 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 12:01:15.350 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:01:16.232 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:01:17.294 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:03:12.586 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.29s (1.92min) +2026-06-20 12:03:13.294 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:05:12.908 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:12:12.700 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 148, train_steps: 5920 +2026-06-20 12:12:13.409 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 12:12:13.622 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:12:13.723 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:12:13.796 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:13:19.792 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000148/group_10/index_3986_seed_10475_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 12:13:22.814 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 12:13:41.649 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:13:41.659 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:13:41.659 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:14:49.984 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000148/group_10/index_3986_seed_10476_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 12:14:53.153 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 12:15:12.421 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:15:13.144 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:15:14.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:17:09.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.69s (1.91min) +2026-06-20 12:17:09.789 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:19:10.590 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:26:11.380 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 149, train_steps: 5960 +2026-06-20 12:26:12.136 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 12:26:12.336 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:26:12.430 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:26:12.511 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:27:18.398 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000149/group_10/index_5268_seed_5393_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 12:27:21.558 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 12:27:39.980 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:27:39.990 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:27:39.990 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:28:48.769 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000149/group_10/index_5268_seed_5394_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 12:28:53.846 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 12:29:12.104 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:29:13.029 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:29:13.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:31:08.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.72s (1.91min) +2026-06-20 12:31:09.532 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:33:09.828 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:40:10.986 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 150, train_steps: 6000 +2026-06-20 12:41:53.580 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=150 (missing=0, unexpected=0) +2026-06-20 12:41:56.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 12:41:56.190 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:41:56.242 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:41:56.316 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:43:04.062 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_10/index_8994_seed_31947_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 12:43:08.813 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 12:43:26.729 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:43:26.738 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:43:26.738 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:44:32.765 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_10/index_8994_seed_31948_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 12:44:37.516 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 12:44:55.797 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:44:59.744 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:45:00.658 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:46:55.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.70s (1.91min) +2026-06-20 12:46:56.236 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:48:52.990 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:55:51.593 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 151, train_steps: 6040 +2026-06-20 12:55:52.309 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 12:55:52.526 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:55:52.654 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:55:52.711 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:56:58.534 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_10/index_276_seed_30404_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 12:57:03.244 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 12:57:21.278 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:57:21.288 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:57:21.288 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:58:27.344 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_10/index_276_seed_30405_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 12:58:31.361 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 12:58:50.135 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:58:50.721 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:58:51.774 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:00:46.190 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.41s (1.91min) +2026-06-20 13:00:46.908 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:02:43.654 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:09:41.905 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 152, train_steps: 6080 +2026-06-20 13:09:42.597 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 13:09:42.731 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:09:42.768 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:09:42.856 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:10:50.169 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_10/index_4879_seed_7541_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 13:10:55.308 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 13:11:13.430 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:11:13.440 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:11:13.440 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:12:20.482 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_10/index_4879_seed_7542_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 13:12:25.278 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 13:12:43.354 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:12:43.910 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:12:44.835 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:14:38.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-20 13:14:39.492 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:16:38.552 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:23:37.218 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 153, train_steps: 6120 +2026-06-20 13:23:38.094 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 13:23:38.228 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:23:38.266 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:23:38.337 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:24:45.094 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_10/index_1238_seed_5425_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 13:24:49.133 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 13:25:07.484 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:25:07.494 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:25:07.496 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:26:15.103 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_10/index_1238_seed_5426_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 13:26:20.262 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 13:26:38.560 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:26:40.127 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:26:41.341 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:28:34.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.39s (1.89min) +2026-06-20 13:28:35.473 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:30:32.909 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:37:31.701 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 154, train_steps: 6160 +2026-06-20 13:37:32.395 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 13:37:32.602 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:37:32.698 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:37:32.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:38:38.591 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_10/index_4667_seed_24410_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 13:38:43.661 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 13:39:01.740 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:39:01.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:39:01.749 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:40:09.457 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_10/index_4667_seed_24411_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 13:40:14.272 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 13:40:32.482 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:40:33.788 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:40:34.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:42:29.329 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.47s (1.91min) +2026-06-20 13:42:30.062 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:44:27.158 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:51:24.971 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 155, train_steps: 6200 +2026-06-20 13:51:25.703 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 13:51:25.924 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:51:26.039 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:51:26.133 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:52:31.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_10/index_1531_seed_29019_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 13:52:36.710 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 13:52:54.757 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:52:54.766 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:52:54.766 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:54:01.200 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_10/index_1531_seed_29020_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 13:54:06.082 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 13:54:24.159 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:54:24.447 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:54:25.344 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:56:19.129 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.63s (1.89min) +2026-06-20 13:56:19.859 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:58:16.751 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:05:15.342 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 156, train_steps: 6240 +2026-06-20 14:05:16.036 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 14:05:16.235 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:05:16.342 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:05:16.447 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:06:22.915 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_10/index_6862_seed_21516_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 14:06:26.905 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 14:06:45.811 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:06:45.822 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:06:45.822 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:07:51.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_10/index_6862_seed_21517_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 14:07:57.108 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 14:08:15.458 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:08:15.931 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:08:16.846 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:10:11.722 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.74s (1.91min) +2026-06-20 14:10:12.457 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:12:11.105 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:19:12.236 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 157, train_steps: 6280 +2026-06-20 14:19:12.925 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 14:19:13.108 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:19:13.210 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:19:13.303 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:20:19.239 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000157/group_10/index_7935_seed_16861_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 14:20:23.258 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 14:20:42.882 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:20:42.889 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:20:42.889 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:21:48.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000157/group_10/index_7935_seed_16862_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 14:21:53.168 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 14:22:11.673 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:22:12.398 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:22:13.415 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:24:08.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.22s (1.92min) +2026-06-20 14:24:09.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:26:09.906 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:33:10.821 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 158, train_steps: 6320 +2026-06-20 14:33:11.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 14:33:11.667 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:33:11.725 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:33:11.845 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:34:17.576 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_10/index_7997_seed_12891_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 14:34:21.689 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 14:34:40.145 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:34:40.154 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:34:40.154 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:35:47.116 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_10/index_7997_seed_12892_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 14:35:50.172 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 14:36:08.960 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:36:10.935 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:36:12.076 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:38:07.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.88s (1.91min) +2026-06-20 14:38:07.785 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:40:05.261 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:47:04.569 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 159, train_steps: 6360 +2026-06-20 14:47:05.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 14:47:05.412 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:47:05.449 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:47:05.517 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:48:11.488 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_10/index_3718_seed_15327_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 14:48:14.407 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 14:48:32.899 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:48:32.910 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:48:32.910 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:49:40.200 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_10/index_3718_seed_15328_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 14:49:45.364 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 14:50:04.038 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:50:04.048 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:50:05.185 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:51:59.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.86s (1.90min) +2026-06-20 14:51:59.929 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:53:57.323 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:00:59.172 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 160, train_steps: 6400 +2026-06-20 15:01:37.067 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=160 (missing=0, unexpected=0) +2026-06-20 15:01:39.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 15:01:39.443 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:01:39.507 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:01:39.625 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:02:46.846 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_10/index_3321_seed_14243_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 15:02:51.564 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 15:03:09.535 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:03:09.545 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:03:09.545 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:04:18.899 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_10/index_3321_seed_14244_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 15:04:23.744 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 15:04:42.869 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:04:42.879 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:04:43.770 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:06:37.409 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.48s (1.89min) +2026-06-20 15:06:38.136 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:08:38.304 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:15:37.526 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 161, train_steps: 6440 +2026-06-20 15:15:38.206 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 15:15:38.312 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:15:38.355 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:15:38.403 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:16:44.668 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_10/index_6331_seed_7685_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 15:16:49.520 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 15:17:08.813 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:17:08.822 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:17:08.822 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:18:16.510 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_10/index_6331_seed_7686_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 15:18:21.328 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 15:18:39.524 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:18:40.375 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:18:41.666 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:20:35.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-20 15:20:36.332 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:22:36.150 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:29:36.903 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 162, train_steps: 6480 +2026-06-20 15:29:37.569 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 15:29:37.678 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:29:37.717 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:29:37.759 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:30:43.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_10/index_5617_seed_227_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 15:30:48.564 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 15:31:06.511 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:31:06.520 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:31:06.522 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:32:11.847 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_10/index_5617_seed_228_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 15:32:16.705 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 15:32:34.964 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:32:37.444 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:32:38.400 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:34:33.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.48s (1.91min) +2026-06-20 15:34:33.827 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:36:31.319 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:43:31.483 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 163, train_steps: 6520 +2026-06-20 15:43:32.172 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 15:43:32.371 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:43:32.478 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:43:32.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:44:38.530 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000163/group_10/index_6048_seed_8574_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 15:44:43.181 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 15:45:01.997 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:45:02.006 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:45:02.006 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:46:09.578 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000163/group_10/index_6048_seed_8575_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 15:46:12.678 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 15:46:31.318 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:46:33.676 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:46:34.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:48:29.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.47s (1.91min) +2026-06-20 15:48:29.935 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:50:26.541 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:57:25.508 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 164, train_steps: 6560 +2026-06-20 15:57:26.227 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 15:57:26.328 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:57:26.373 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:57:26.420 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:58:32.474 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_10/index_4249_seed_621_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 15:58:35.495 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 15:58:53.670 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:58:53.680 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:58:53.680 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:00:02.460 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_10/index_4249_seed_622_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 16:00:05.520 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 16:00:24.172 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:00:26.402 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:00:27.369 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:02:21.064 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.61s (1.89min) +2026-06-20 16:02:21.793 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:04:18.344 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:11:18.370 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 165, train_steps: 6600 +2026-06-20 16:11:19.088 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 16:11:19.304 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:11:19.422 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:11:19.518 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:12:25.786 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_10/index_8637_seed_6127_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 16:12:29.738 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 16:12:47.869 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:12:47.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:12:47.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:13:54.866 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_10/index_8637_seed_6128_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 16:14:00.007 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 16:14:18.003 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:14:19.106 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:14:20.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:16:14.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.75s (1.90min) +2026-06-20 16:16:14.870 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:18:14.126 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:25:13.923 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 166, train_steps: 6640 +2026-06-20 16:25:14.621 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 16:25:14.802 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:25:14.905 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:25:15.011 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:26:21.221 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_10/index_3947_seed_30094_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 16:26:25.397 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 16:26:43.808 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:26:43.818 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:26:43.818 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:27:50.381 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_10/index_3947_seed_30095_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 16:27:55.584 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 16:28:13.822 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:28:13.831 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:28:14.852 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:30:08.844 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.78s (1.90min) +2026-06-20 16:30:09.574 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:32:10.307 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:39:12.149 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 167, train_steps: 6680 +2026-06-20 16:39:12.849 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 16:39:13.036 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:39:13.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:39:13.215 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:40:19.531 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_10/index_6387_seed_14673_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 16:40:24.301 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 16:40:42.822 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:40:42.830 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:40:42.831 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:41:48.228 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_10/index_6387_seed_14674_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 16:41:52.299 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 16:42:11.061 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:42:12.095 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:42:13.377 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:44:08.745 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.36s (1.92min) +2026-06-20 16:44:09.477 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:46:10.022 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:53:11.455 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 168, train_steps: 6720 +2026-06-20 16:53:12.091 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 16:53:12.243 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:53:12.285 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:53:12.397 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:54:18.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_10/index_8887_seed_8254_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 16:54:21.639 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 16:54:39.874 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:54:39.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:54:39.887 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:55:47.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_10/index_8887_seed_8255_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 16:55:50.929 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 16:56:09.452 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:56:11.767 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:56:12.797 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:58:07.736 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-20 16:58:08.464 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:00:07.581 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:07:09.179 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 169, train_steps: 6760 +2026-06-20 17:07:09.877 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 17:07:10.024 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 17:07:10.084 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:07:10.221 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:08:16.550 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_10/index_9831_seed_22099_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 17:08:19.609 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 17:08:37.918 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:08:37.928 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:08:37.928 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:09:47.680 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_10/index_9831_seed_22100_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 17:09:50.826 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 17:10:09.563 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:10:11.904 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 17:10:12.947 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:12:08.255 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.20s (1.92min) +2026-06-20 17:12:08.969 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:14:09.321 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:21:11.121 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 170, train_steps: 6800 +2026-06-20 17:21:47.596 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=170 (missing=0, unexpected=0) +2026-06-20 17:21:49.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 17:21:49.317 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 17:21:49.375 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:21:49.482 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:22:56.673 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_10/index_6322_seed_3045_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 17:22:59.827 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 17:23:18.231 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:23:18.242 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:23:18.242 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:24:25.697 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_10/index_6322_seed_3046_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 17:24:28.944 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 17:24:47.325 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:24:49.925 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 17:24:50.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:26:44.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.05s (1.88min) +2026-06-20 17:26:44.743 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:28:41.428 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:35:39.742 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 42, sp_rank 0: update step: 171, train_steps: 6840 +2026-06-20 17:35:40.433 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 42, dp_rank 21: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=10, rank_in_group=1 +2026-06-20 17:35:40.655 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 17:35:40.902 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:35:41.006 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:36:47.447 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000171/group_10/index_3912_seed_6526_dp_21_rank_42_sample_2_slot_0.mp4 +2026-06-20 17:36:50.606 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 17:37:09.286 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:37:09.296 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 42 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:37:09.297 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 42 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:38:19.447 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 42: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000171/group_10/index_3912_seed_6527_dp_21_rank_42_sample_3_slot_1.mp4 +2026-06-20 17:38:22.643 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:2 for reward scoring. +2026-06-20 17:38:42.526 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:38:42.537 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 42: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 17:38:43.739 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 42: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:40:36.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 42: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.89s (1.88min) +2026-06-20 17:40:37.378 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 42, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-29 01:31:20.981 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 01:31:21.646 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 01:31:21.647 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 01:53:14.336 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 01:53:15.004 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 01:53:15.007 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 02:22:43.022 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 02:22:43.709 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 02:22:43.712 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 02:48:02.304 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 02:48:02.989 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 02:48:02.989 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:09:53.705 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 03:09:54.373 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:09:54.373 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:33:46.920 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 03:33:47.585 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:33:47.585 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:00:51.060 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 04:00:51.722 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:00:51.724 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:28:57.286 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 04:28:57.955 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:28:57.958 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:54:00.317 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 04:54:00.994 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:54:00.995 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 05:18:37.026 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 05:18:37.698 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 05:18:37.698 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 05:41:12.336 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 05:41:13.015 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 05:41:13.015 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:05:20.675 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 06:05:21.356 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:05:21.357 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:27:33.942 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 06:27:34.620 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:27:34.620 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:50:12.799 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 06:50:13.477 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:50:13.478 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 07:14:09.513 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 07:14:10.176 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 07:14:10.179 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 07:36:27.349 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 07:36:28.025 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 07:36:28.026 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:07:21.852 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 08:07:22.532 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:07:22.533 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:31:44.703 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 08:31:45.380 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:31:45.383 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:01:08.676 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 09:01:09.342 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:01:09.344 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:24:41.766 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 09:24:42.434 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:24:42.435 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:49:18.231 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 09:49:18.910 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:49:18.910 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 10:12:47.172 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 10:12:47.830 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 10:12:47.831 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 10:38:21.105 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 10:38:21.777 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 10:38:21.779 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:04:23.364 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 11:04:24.031 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:04:24.031 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:29:13.703 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 11:29:14.369 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:29:14.370 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:59:16.812 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 11:59:17.475 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:59:17.476 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 12:20:18.697 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 12:20:19.350 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 12:20:19.353 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 12:48:39.134 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 12:48:39.810 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 12:48:39.811 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 13:14:44.208 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 13:14:44.867 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 13:14:44.870 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 13:41:29.703 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 13:41:30.388 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 13:41:30.388 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:07:22.174 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 14:07:22.853 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:07:22.854 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:28:55.397 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 14:28:56.059 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:28:56.062 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:59:01.405 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 14:59:02.090 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:59:02.091 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 15:22:59.558 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 15:23:00.227 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 15:23:00.230 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 15:48:54.034 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 15:48:54.707 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 15:48:54.710 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 16:19:08.082 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 16:19:08.738 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 16:19:08.739 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 16:43:04.141 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 16:43:04.798 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 16:43:04.798 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:11:33.351 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 17:11:34.029 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:11:34.030 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:34:03.585 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 17:34:04.262 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:34:04.263 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:55:29.427 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 17:55:30.101 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:55:30.102 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 18:17:59.429 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 18:18:00.101 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 18:18:00.102 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 18:40:45.406 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 18:40:46.074 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 18:40:46.074 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_45.log b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_45.log new file mode 100644 index 0000000000000000000000000000000000000000..cc2d651f7d19a95d9399c8158bbf8f281d08f291 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_45.log @@ -0,0 +1,1994 @@ +2026-06-19 01:20:19.766 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-19 01:20:19.767 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-19 01:30:17.294 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 01:30:17.407 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:30:17.415 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:31:52.480 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:31:52.482 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:33:45.515 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 01:34:01.005 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 01:35:56.997 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.94s (1.93min) +2026-06-19 01:35:57.688 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 01:37:54.192 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 01:37:57.221 | INFO | hyvideo.utils.grpo_utils:train_one_step:3127 - Rank 45, sp_rank 0: init adaptive KL controller (init=1.000e-07, min=1.000e-07, max=1.000e-05) +2026-06-19 01:45:11.467 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 1, train_steps: 40 +2026-06-19 01:45:12.573 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 01:45:12.678 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:45:12.699 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:46:17.197 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:46:17.200 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:48:10.945 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 01:48:11.851 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 01:50:07.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.70s (1.93min) +2026-06-19 01:50:08.466 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 01:52:06.956 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 01:59:06.753 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 2, train_steps: 80 +2026-06-19 01:59:07.380 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 01:59:07.477 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:59:07.485 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:00:11.619 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:00:11.621 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:02:10.528 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 02:02:11.532 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:04:06.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.88s (1.91min) +2026-06-19 02:04:07.374 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:06:03.757 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:13:04.855 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 3, train_steps: 120 +2026-06-19 02:13:05.465 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 02:13:05.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:13:05.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:14:10.086 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:14:10.088 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:16:04.655 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 02:16:05.823 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:17:59.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.77s (1.90min) +2026-06-19 02:18:00.448 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:19:59.353 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:26:59.879 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 4, train_steps: 160 +2026-06-19 02:27:00.546 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 02:27:00.667 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:27:00.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:28:05.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:28:05.429 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:30:00.316 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 02:30:01.252 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:31:55.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.62s (1.91min) +2026-06-19 02:31:56.651 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:33:53.226 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:40:53.216 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 5, train_steps: 200 +2026-06-19 02:40:53.901 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 02:40:54.081 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:40:54.181 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:41:58.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:41:58.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:43:51.812 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 02:43:52.742 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:45:46.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.50s (1.89min) +2026-06-19 02:45:46.993 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:47:43.354 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:54:40.657 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 6, train_steps: 240 +2026-06-19 02:54:41.360 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 02:54:41.472 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:54:41.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:55:47.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:55:47.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:57:42.949 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 02:57:43.906 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:59:37.354 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.38s (1.89min) +2026-06-19 02:59:38.043 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:01:37.734 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:08:37.471 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 7, train_steps: 280 +2026-06-19 03:08:38.147 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 03:08:38.316 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:08:38.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:09:42.641 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:09:42.646 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:11:38.337 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 03:11:39.427 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:13:34.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.68s (1.91min) +2026-06-19 03:13:34.976 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:15:32.045 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:22:31.144 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 8, train_steps: 320 +2026-06-19 03:22:31.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 03:22:31.981 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:22:32.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:23:37.122 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:23:37.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:25:33.100 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 03:25:33.894 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:27:28.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-19 03:27:29.431 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:29:26.905 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:36:25.646 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 9, train_steps: 360 +2026-06-19 03:36:26.346 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 03:36:26.515 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:36:26.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:37:31.260 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:37:31.263 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:39:26.010 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 03:39:26.818 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:41:20.400 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.36s (1.89min) +2026-06-19 03:41:21.106 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:43:17.559 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:50:16.932 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 10, train_steps: 400 +2026-06-19 03:50:53.200 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=10 (missing=0, unexpected=0) +2026-06-19 03:50:56.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 03:50:56.313 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:50:56.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:52:02.227 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:52:02.231 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:53:57.828 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 03:53:58.621 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:55:52.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.65s (1.89min) +2026-06-19 03:55:53.209 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:57:49.764 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:04:48.441 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 11, train_steps: 440 +2026-06-19 04:04:49.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 04:04:49.239 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:04:49.291 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:05:53.666 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:05:53.672 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:07:48.853 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 04:07:49.645 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:09:43.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.25s (1.89min) +2026-06-19 04:09:43.797 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:11:40.269 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:18:37.282 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 12, train_steps: 480 +2026-06-19 04:18:37.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 04:18:38.082 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:18:38.120 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:19:42.902 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:19:42.903 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:21:37.180 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 04:21:37.964 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:23:33.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.19s (1.92min) +2026-06-19 04:23:34.070 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:25:31.243 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:32:29.817 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 13, train_steps: 520 +2026-06-19 04:32:30.512 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 04:32:30.649 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:32:30.693 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:33:35.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:33:35.359 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:35:30.313 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 04:35:31.103 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:37:24.102 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.77s (1.88min) +2026-06-19 04:37:24.791 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:39:21.077 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:46:18.586 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 14, train_steps: 560 +2026-06-19 04:46:19.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 04:46:19.422 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:46:19.529 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:47:24.265 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:47:24.269 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:49:17.445 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 04:49:18.247 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:51:12.963 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.49s (1.91min) +2026-06-19 04:51:13.650 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:53:12.995 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:00:11.825 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 15, train_steps: 600 +2026-06-19 05:00:12.525 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 05:00:12.643 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:00:12.681 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:01:17.459 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:01:17.462 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:03:11.349 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 05:03:12.457 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:05:07.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.64s (1.91min) +2026-06-19 05:05:08.060 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:07:05.975 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:14:04.894 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 16, train_steps: 640 +2026-06-19 05:14:05.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 05:14:05.784 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:14:05.932 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:15:10.270 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:15:10.273 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:17:06.988 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 05:17:07.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:19:03.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.02s (1.92min) +2026-06-19 05:19:03.741 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:21:01.962 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:28:03.315 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 17, train_steps: 680 +2026-06-19 05:28:04.006 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 05:28:04.175 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:28:04.255 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:29:09.091 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:29:09.094 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:31:03.926 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 05:31:04.840 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:32:58.689 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.48s (1.89min) +2026-06-19 05:32:59.387 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:34:56.266 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:41:58.094 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 18, train_steps: 720 +2026-06-19 05:41:58.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 05:41:58.894 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:41:58.936 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:43:03.551 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:43:03.554 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:45:00.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 05:45:00.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:46:55.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.88s (1.90min) +2026-06-19 05:46:55.797 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:48:52.331 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:55:50.497 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 19, train_steps: 760 +2026-06-19 05:55:51.178 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 05:55:51.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:55:51.453 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:56:55.841 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:56:55.849 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:58:49.068 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 05:58:49.873 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:00:43.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.10s (1.89min) +2026-06-19 06:00:43.876 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:02:40.342 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:09:39.753 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 20, train_steps: 800 +2026-06-19 06:10:16.192 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=20 (missing=0, unexpected=0) +2026-06-19 06:10:18.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 06:10:18.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:10:18.932 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:11:24.756 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:11:24.760 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:13:19.050 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 06:13:19.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:15:13.224 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.90s (1.88min) +2026-06-19 06:15:13.910 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:17:12.687 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:24:12.555 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 21, train_steps: 840 +2026-06-19 06:24:13.270 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 06:24:13.412 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:24:13.466 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:25:17.846 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:25:17.849 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:27:11.261 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 06:27:12.132 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:29:06.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.51s (1.91min) +2026-06-19 06:29:07.668 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:31:04.578 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:38:04.277 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 22, train_steps: 880 +2026-06-19 06:38:04.947 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 06:38:05.076 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:38:05.117 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:39:09.527 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:39:09.534 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:41:03.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 06:41:04.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:42:58.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.99s (1.90min) +2026-06-19 06:42:59.165 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:44:56.505 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:51:58.166 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 23, train_steps: 920 +2026-06-19 06:51:58.842 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 06:51:58.960 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:51:58.995 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:53:03.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:53:03.369 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:54:59.789 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 06:55:00.597 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:56:54.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.11s (1.90min) +2026-06-19 06:56:55.622 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:58:52.321 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:05:51.312 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 24, train_steps: 960 +2026-06-19 07:05:52.004 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 07:05:52.181 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:05:52.281 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:06:56.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:06:56.709 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:08:52.217 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 07:08:52.994 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:10:46.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-19 07:10:46.871 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:12:44.155 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:19:42.042 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 25, train_steps: 1000 +2026-06-19 07:20:52.104 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 07:20:52.221 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:20:52.255 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:21:56.851 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:21:56.859 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:23:52.666 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 07:23:53.478 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:25:46.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.81s (1.88min) +2026-06-19 07:25:47.205 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:27:44.382 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:34:43.681 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 26, train_steps: 1040 +2026-06-19 07:34:44.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 07:34:44.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:34:44.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:35:49.306 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:35:49.309 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:37:45.443 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 07:37:46.259 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:39:39.639 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.14s (1.89min) +2026-06-19 07:39:40.348 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:41:39.073 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:48:39.680 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 27, train_steps: 1080 +2026-06-19 07:48:40.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 07:48:40.501 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:48:40.592 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:49:45.416 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:49:45.416 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:51:43.285 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 07:51:44.085 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:53:38.298 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.99s (1.90min) +2026-06-19 07:53:39.007 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:55:38.422 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:02:38.171 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 28, train_steps: 1120 +2026-06-19 08:02:38.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 08:02:39.046 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:02:39.127 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:03:43.728 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:03:43.731 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:05:41.901 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 08:05:42.911 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:07:37.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.24s (1.90min) +2026-06-19 08:07:38.091 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:09:37.934 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:16:37.844 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 29, train_steps: 1160 +2026-06-19 08:16:38.692 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 08:16:38.868 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:16:39.133 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:17:43.550 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:17:43.552 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:19:42.881 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 08:19:43.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:21:37.264 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 08:21:37.961 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:23:37.407 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:30:39.544 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 30, train_steps: 1200 +2026-06-19 08:31:15.231 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=30 (missing=0, unexpected=0) +2026-06-19 08:31:18.258 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 08:31:18.347 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:31:18.404 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:32:24.912 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:32:24.915 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:34:19.980 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 08:34:20.917 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:36:14.510 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-19 08:36:15.227 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:38:14.173 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:45:14.282 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 31, train_steps: 1240 +2026-06-19 08:45:14.994 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 08:45:15.204 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:45:15.340 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:46:19.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:46:19.607 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:48:15.288 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 08:48:16.248 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:50:10.323 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.84s (1.90min) +2026-06-19 08:50:11.023 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:52:11.150 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:59:13.090 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 32, train_steps: 1280 +2026-06-19 08:59:13.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 08:59:13.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:59:14.063 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:00:18.420 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:00:18.422 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:02:13.436 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 09:02:14.618 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:04:08.713 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.91s (1.90min) +2026-06-19 09:04:09.421 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:06:10.411 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:13:13.273 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 33, train_steps: 1320 +2026-06-19 09:13:13.952 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 09:13:14.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:13:14.095 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:14:18.739 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:14:18.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:16:14.822 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 09:16:15.834 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:18:11.066 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.06s (1.92min) +2026-06-19 09:18:11.785 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:20:11.629 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:27:13.811 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 34, train_steps: 1360 +2026-06-19 09:27:14.444 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 09:27:14.623 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:27:14.719 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:28:19.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:28:19.127 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:30:13.695 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 09:30:14.794 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:32:08.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.81s (1.90min) +2026-06-19 09:32:09.592 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:34:09.780 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:41:12.404 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 35, train_steps: 1400 +2026-06-19 09:41:13.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 09:41:13.226 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:41:13.297 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:42:17.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:42:17.720 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:44:13.080 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 09:44:13.914 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:46:10.686 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 116.56s (1.94min) +2026-06-19 09:46:11.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:48:10.896 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:55:11.958 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 36, train_steps: 1440 +2026-06-19 09:55:12.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 09:55:12.856 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:55:12.981 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:56:17.575 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:56:17.575 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:58:12.646 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 09:58:13.653 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:00:08.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-19 10:00:09.671 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:02:10.008 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:09:11.703 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 37, train_steps: 1480 +2026-06-19 10:09:12.395 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 10:09:12.537 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:09:12.583 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:10:16.859 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:10:16.863 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:12:13.042 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 10:12:13.842 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:14:09.030 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.94s (1.92min) +2026-06-19 10:14:09.737 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:16:10.368 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:23:12.351 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 38, train_steps: 1520 +2026-06-19 10:23:13.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 10:23:13.204 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:23:13.265 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:24:17.726 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:24:17.731 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:26:12.667 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 10:26:13.684 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:28:09.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.16s (1.92min) +2026-06-19 10:28:09.954 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:30:10.743 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:37:11.946 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 39, train_steps: 1560 +2026-06-19 10:37:12.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 10:37:12.830 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:37:12.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:38:17.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:38:17.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:41:42.526 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 10:41:43.691 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:43:36.471 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.70s (1.88min) +2026-06-19 10:43:37.180 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:45:37.179 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:52:37.853 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 40, train_steps: 1600 +2026-06-19 10:53:14.108 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=40 (missing=0, unexpected=0) +2026-06-19 10:53:15.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 10:53:15.915 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:53:15.934 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:54:21.980 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:54:21.984 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:56:16.410 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 10:56:17.402 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:58:11.717 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.27s (1.90min) +2026-06-19 10:58:12.416 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:00:11.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:07:12.578 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 41, train_steps: 1640 +2026-06-19 11:07:13.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 11:07:13.501 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:07:13.640 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:08:17.792 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:08:17.795 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:10:12.668 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 11:10:13.758 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:12:08.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.28s (1.90min) +2026-06-19 11:12:09.087 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:14:08.757 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:21:10.886 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 42, train_steps: 1680 +2026-06-19 11:21:11.810 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 11:21:12.012 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:21:12.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:22:16.539 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:22:16.539 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:24:12.073 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 11:24:13.025 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:26:07.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.32s (1.91min) +2026-06-19 11:26:08.278 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:28:06.795 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:35:06.765 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 43, train_steps: 1720 +2026-06-19 11:35:07.458 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 11:35:07.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:35:07.792 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:36:12.087 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:36:12.090 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:38:10.299 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 11:38:11.141 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:40:05.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.38s (1.91min) +2026-06-19 11:40:06.549 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:42:03.525 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:49:04.430 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 44, train_steps: 1760 +2026-06-19 11:49:05.114 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 11:49:05.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:49:05.328 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:50:09.849 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:50:09.860 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:52:06.491 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 11:52:07.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:54:01.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.69s (1.89min) +2026-06-19 11:54:01.909 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:56:01.237 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:03:03.865 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 45, train_steps: 1800 +2026-06-19 12:03:04.557 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 12:03:04.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:03:04.778 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:04:09.129 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:04:09.130 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:06:05.347 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 12:06:06.300 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:08:01.417 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.05s (1.92min) +2026-06-19 12:08:02.101 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:10:01.383 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:17:03.587 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 46, train_steps: 1840 +2026-06-19 12:17:04.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 12:17:04.402 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:17:04.444 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:18:08.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:18:08.938 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:20:05.333 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 12:20:06.499 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:22:01.733 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.17s (1.92min) +2026-06-19 12:22:02.442 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:24:01.453 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:31:02.922 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 47, train_steps: 1880 +2026-06-19 12:31:03.608 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 12:31:03.802 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:31:03.926 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:32:07.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:32:07.972 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:34:03.182 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 12:34:04.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:35:58.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-19 12:35:58.982 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:37:55.893 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:44:57.679 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 48, train_steps: 1920 +2026-06-19 12:44:58.355 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 12:44:58.521 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:44:58.610 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:46:02.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:46:02.818 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:47:58.931 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 12:47:59.921 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:49:54.620 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-19 12:49:55.342 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:51:52.291 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:58:52.086 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 49, train_steps: 1960 +2026-06-19 12:58:52.783 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 12:58:52.989 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:58:53.089 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:59:57.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:59:57.442 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:01:54.369 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 13:01:55.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:03:48.446 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.04s (1.88min) +2026-06-19 13:03:49.149 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:05:45.469 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:12:45.267 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 50, train_steps: 2000 +2026-06-19 13:14:21.474 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=50 (missing=0, unexpected=0) +2026-06-19 13:14:23.203 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 13:14:23.322 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:14:23.353 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:15:29.744 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:15:29.747 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:17:25.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 13:17:26.027 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:19:20.081 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.84s (1.90min) +2026-06-19 13:19:20.807 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:21:17.527 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:28:18.465 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 51, train_steps: 2040 +2026-06-19 13:28:19.185 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 13:28:19.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:28:19.434 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:29:23.987 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:29:23.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:31:18.029 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 13:31:19.040 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:33:13.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.83s (1.90min) +2026-06-19 13:33:13.741 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:35:13.231 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:42:13.612 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 52, train_steps: 2080 +2026-06-19 13:42:14.277 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 13:42:14.476 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:42:14.571 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:43:19.138 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:43:19.141 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:45:15.287 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 13:45:16.376 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:47:10.481 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.03s (1.90min) +2026-06-19 13:47:11.196 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:49:11.267 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:56:12.598 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 53, train_steps: 2120 +2026-06-19 13:56:13.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 13:56:13.463 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:56:13.520 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:57:17.855 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:57:17.858 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:59:13.130 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 13:59:14.097 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:01:08.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.41s (1.91min) +2026-06-19 14:01:09.529 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:03:10.475 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:10:12.164 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 54, train_steps: 2160 +2026-06-19 14:10:12.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 14:10:13.018 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:10:13.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:11:17.491 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:11:17.492 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:13:12.866 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 14:13:13.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:15:08.493 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.42s (1.91min) +2026-06-19 14:15:09.218 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:17:09.306 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:24:11.899 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 55, train_steps: 2200 +2026-06-19 14:24:12.602 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 14:24:12.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:24:12.761 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:25:17.388 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:25:17.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:27:12.754 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 14:27:13.730 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:29:09.888 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.82s (1.93min) +2026-06-19 14:29:10.604 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:31:10.764 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:38:12.113 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 56, train_steps: 2240 +2026-06-19 14:38:12.811 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 14:38:12.946 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:38:12.981 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:39:17.362 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:39:17.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:41:13.092 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 14:41:13.901 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:43:08.781 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.66s (1.91min) +2026-06-19 14:43:09.475 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:45:09.717 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:52:11.838 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 57, train_steps: 2280 +2026-06-19 14:52:12.526 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 14:52:12.679 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:52:12.743 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:53:17.216 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:53:17.218 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:55:13.057 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 14:55:13.872 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:57:08.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.79s (1.91min) +2026-06-19 14:57:09.597 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:59:09.882 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:06:12.292 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 58, train_steps: 2320 +2026-06-19 15:06:12.985 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 15:06:13.133 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:06:13.191 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:07:17.808 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:07:17.811 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:09:13.062 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 15:09:13.868 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:11:08.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.71s (1.91min) +2026-06-19 15:11:09.498 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:13:09.810 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:20:12.055 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 59, train_steps: 2360 +2026-06-19 15:20:12.739 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 15:20:12.923 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:20:13.020 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:21:17.597 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:21:17.600 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:23:12.372 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 15:23:13.402 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:25:09.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.28s (1.92min) +2026-06-19 15:25:09.743 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:27:10.367 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:34:12.067 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 60, train_steps: 2400 +2026-06-19 15:34:48.035 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=60 (missing=0, unexpected=0) +2026-06-19 15:34:50.615 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 15:34:50.855 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:34:50.937 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:35:56.808 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:35:56.811 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:37:53.529 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 15:37:54.544 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:39:47.707 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-19 15:39:48.402 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:41:45.234 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:48:44.576 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 61, train_steps: 2440 +2026-06-19 15:48:45.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 15:48:45.498 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:48:45.625 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:49:50.027 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:49:50.030 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:51:47.609 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 15:51:48.411 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:53:41.681 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.04s (1.88min) +2026-06-19 15:53:42.399 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:55:39.327 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:02:39.006 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 62, train_steps: 2480 +2026-06-19 16:02:39.688 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 16:02:39.902 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:02:40.019 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:03:44.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:03:44.240 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:05:44.132 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 16:05:44.954 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:07:38.761 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.59s (1.89min) +2026-06-19 16:07:39.471 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:09:38.854 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:16:39.317 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 63, train_steps: 2520 +2026-06-19 16:16:40.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 16:16:40.261 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:16:40.386 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:17:44.859 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:17:44.865 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:19:45.197 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 16:19:46.185 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:21:39.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.36s (1.89min) +2026-06-19 16:21:40.303 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:23:38.884 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:30:41.057 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 64, train_steps: 2560 +2026-06-19 16:30:41.625 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 16:30:41.743 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:30:41.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:31:47.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:31:47.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:33:45.471 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 16:33:46.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:35:40.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.63s (1.89min) +2026-06-19 16:35:40.844 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:37:39.400 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:44:41.294 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 65, train_steps: 2600 +2026-06-19 16:44:41.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 16:44:42.096 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:44:42.122 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:45:47.562 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:45:47.563 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:47:47.482 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 16:47:48.458 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:49:42.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.72s (1.90min) +2026-06-19 16:49:42.958 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:51:40.038 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:58:41.282 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 66, train_steps: 2640 +2026-06-19 16:58:41.989 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 16:58:42.163 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:58:42.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:59:47.372 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:59:47.372 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:01:43.599 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 17:01:44.461 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:03:37.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.34s (1.89min) +2026-06-19 17:03:38.682 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:05:38.382 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:12:40.121 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 67, train_steps: 2680 +2026-06-19 17:12:40.909 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 17:12:41.271 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:12:41.387 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:13:46.762 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:13:46.762 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:15:44.529 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 17:15:45.513 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:17:38.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 17:17:39.679 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:19:39.114 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:26:39.545 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 68, train_steps: 2720 +2026-06-19 17:26:40.217 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 17:26:40.422 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:26:40.564 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:27:45.379 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:27:45.382 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:29:43.886 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 17:29:44.880 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:31:38.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.76s (1.90min) +2026-06-19 17:31:39.420 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:33:38.461 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:40:38.360 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 69, train_steps: 2760 +2026-06-19 17:40:39.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 17:40:39.132 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:40:39.163 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:41:43.666 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:41:43.667 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:43:41.069 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 17:43:42.266 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:45:35.854 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.55s (1.89min) +2026-06-19 17:45:36.570 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:47:37.118 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:54:38.571 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 70, train_steps: 2800 +2026-06-19 17:55:14.910 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=70 (missing=0, unexpected=0) +2026-06-19 17:55:16.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 17:55:17.159 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:55:17.228 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:56:23.459 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:56:23.462 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:58:18.229 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 17:58:19.425 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:00:13.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.58s (1.89min) +2026-06-19 18:00:13.838 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:02:13.855 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:09:13.890 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 71, train_steps: 2840 +2026-06-19 18:09:14.615 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 18:09:14.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:09:14.834 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:10:19.458 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:10:19.462 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:12:16.021 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 18:12:16.893 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:14:12.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.67s (1.93min) +2026-06-19 18:14:13.460 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:16:12.693 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:23:13.825 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 72, train_steps: 2880 +2026-06-19 18:23:14.506 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 18:23:14.710 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:23:14.848 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:24:19.198 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:24:19.201 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:26:14.711 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 18:26:15.674 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:28:10.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.45s (1.91min) +2026-06-19 18:28:11.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:30:10.989 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:37:12.780 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 73, train_steps: 2920 +2026-06-19 18:37:13.457 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 18:37:13.659 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:37:13.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:38:18.329 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:38:18.332 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:40:13.249 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 18:40:14.055 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:42:09.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.16s (1.92min) +2026-06-19 18:42:10.284 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:44:10.444 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:51:12.752 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 74, train_steps: 2960 +2026-06-19 18:51:13.468 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 18:51:13.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:51:13.760 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:52:18.326 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:52:18.326 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:54:13.982 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 18:54:14.946 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:56:10.312 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.21s (1.92min) +2026-06-19 18:56:11.031 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:58:11.445 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:05:14.017 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 75, train_steps: 3000 +2026-06-19 19:06:23.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 19:06:23.342 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:06:23.398 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:07:28.112 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:07:28.117 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:09:24.209 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 19:09:25.043 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:11:18.397 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.14s (1.89min) +2026-06-19 19:11:19.100 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:13:16.253 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:20:16.381 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 76, train_steps: 3040 +2026-06-19 19:20:16.989 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 19:20:17.127 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:20:17.153 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:21:21.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:21:21.359 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:23:18.455 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 19:23:19.599 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:25:13.441 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.66s (1.89min) +2026-06-19 19:25:14.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:27:12.736 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:34:13.267 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 77, train_steps: 3080 +2026-06-19 19:34:13.950 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 19:34:14.143 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:34:14.239 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:35:18.555 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:35:18.560 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:37:13.074 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 19:37:14.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:39:08.734 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.60s (1.91min) +2026-06-19 19:39:09.441 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:41:09.770 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:48:12.646 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 78, train_steps: 3120 +2026-06-19 19:48:13.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 19:48:13.488 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:48:13.520 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:49:18.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:49:18.102 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:51:13.323 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 19:51:14.148 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:53:09.267 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-19 19:53:09.976 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:55:09.985 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:02:13.168 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 79, train_steps: 3160 +2026-06-19 20:02:13.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 20:02:14.038 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:02:14.094 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:03:18.564 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:03:18.569 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:05:13.093 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 20:05:13.898 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:07:09.660 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.53s (1.93min) +2026-06-19 20:07:10.385 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:09:11.033 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:16:13.935 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 80, train_steps: 3200 +2026-06-19 20:16:51.452 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=80 (missing=0, unexpected=0) +2026-06-19 20:16:53.093 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 20:16:53.281 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:16:53.389 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:18:00.247 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:18:00.250 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:19:54.752 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 20:19:55.714 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:21:49.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-19 20:21:50.363 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:23:47.272 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:30:47.422 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 81, train_steps: 3240 +2026-06-19 20:30:48.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 20:30:48.250 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:30:48.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:31:53.004 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:31:53.005 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:33:49.273 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 20:33:50.252 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:35:44.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.70s (1.89min) +2026-06-19 20:35:44.752 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:37:41.680 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:44:40.003 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 82, train_steps: 3280 +2026-06-19 20:44:40.702 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 20:44:40.903 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:44:41.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:45:45.310 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:45:45.313 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:47:44.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 20:47:45.096 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:49:38.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.42s (1.89min) +2026-06-19 20:49:39.275 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:51:38.425 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:58:38.984 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 83, train_steps: 3320 +2026-06-19 20:58:39.739 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 20:58:39.954 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:58:39.997 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:59:44.285 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:59:44.288 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:01:44.304 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 21:01:45.278 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:03:38.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.28s (1.89min) +2026-06-19 21:03:39.340 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:05:38.821 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:12:38.925 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 84, train_steps: 3360 +2026-06-19 21:12:39.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 21:12:39.752 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:12:39.796 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:13:44.464 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:13:44.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:15:43.791 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 21:15:44.768 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:17:38.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.31s (1.89min) +2026-06-19 21:17:38.864 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:19:38.526 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:26:39.497 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 85, train_steps: 3400 +2026-06-19 21:26:40.169 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 21:26:40.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:26:40.466 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:27:45.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:27:45.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:29:43.697 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 21:29:44.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:31:38.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-19 21:31:38.988 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:33:38.759 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:40:38.650 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 86, train_steps: 3440 +2026-06-19 21:40:39.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 21:40:39.510 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:40:39.545 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:41:44.399 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:41:44.403 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:43:42.867 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 21:43:43.839 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:45:37.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 21:45:37.987 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:47:37.709 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:54:39.049 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 87, train_steps: 3480 +2026-06-19 21:54:39.742 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 21:54:39.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:54:39.919 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:55:44.191 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:55:44.191 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:57:43.169 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 21:57:44.163 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:59:37.604 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.39s (1.89min) +2026-06-19 21:59:38.318 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:01:38.556 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:08:39.987 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 88, train_steps: 3520 +2026-06-19 22:08:40.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 22:08:40.820 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:08:40.887 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:09:45.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:09:45.649 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:11:43.755 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 22:11:44.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:13:38.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.57s (1.89min) +2026-06-19 22:13:39.103 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:15:38.304 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:22:38.763 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 89, train_steps: 3560 +2026-06-19 22:22:39.425 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 22:22:39.639 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:22:39.773 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:23:44.264 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:23:44.267 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:25:43.683 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 22:25:44.566 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:27:38.162 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 22:27:38.887 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:29:38.497 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:36:39.420 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 90, train_steps: 3600 +2026-06-19 22:37:15.531 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=90 (missing=0, unexpected=0) +2026-06-19 22:37:17.307 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 22:37:17.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:37:17.476 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:38:23.009 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:38:23.013 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:40:19.241 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 22:40:20.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:42:14.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.70s (1.89min) +2026-06-19 22:42:14.863 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:44:13.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:51:14.853 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 91, train_steps: 3640 +2026-06-19 22:51:15.699 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 22:51:15.852 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:51:15.916 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:52:20.523 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:52:20.526 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:54:15.937 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 22:54:16.819 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:56:12.672 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.67s (1.93min) +2026-06-19 22:56:13.388 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:58:13.110 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:05:14.261 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 92, train_steps: 3680 +2026-06-19 23:05:15.117 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 23:05:15.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:05:15.355 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:06:19.757 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:06:19.762 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:08:16.058 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 23:08:16.886 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:10:11.751 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.62s (1.91min) +2026-06-19 23:10:12.471 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:12:11.282 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:19:13.487 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 93, train_steps: 3720 +2026-06-19 23:19:14.153 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 23:19:14.339 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:19:14.414 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:20:18.752 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:20:18.755 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:22:14.354 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 23:22:15.354 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:24:10.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.33s (1.92min) +2026-06-19 23:24:11.708 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:26:11.285 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:33:13.933 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 94, train_steps: 3760 +2026-06-19 23:33:14.640 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 23:33:14.803 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:33:14.885 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:34:19.244 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:34:19.247 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:36:16.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 23:36:16.944 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:38:12.062 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.89s (1.91min) +2026-06-19 23:38:12.784 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:40:11.320 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:47:11.876 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 95, train_steps: 3800 +2026-06-19 23:47:12.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-19 23:47:12.757 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:47:12.829 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:48:17.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:48:17.148 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:50:12.004 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-19 23:50:13.197 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:52:08.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.82s (1.91min) +2026-06-19 23:52:08.791 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:54:07.679 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:01:09.776 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 96, train_steps: 3840 +2026-06-20 00:01:10.474 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 00:01:10.580 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:01:10.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:02:15.425 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:02:15.433 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:04:11.783 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 00:04:12.916 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:06:07.928 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 00:06:08.644 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:08:07.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:15:08.501 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 97, train_steps: 3880 +2026-06-20 00:15:09.218 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 00:15:09.358 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:15:09.393 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:16:13.667 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:16:13.671 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:18:11.238 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 00:18:12.140 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:20:07.671 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.20s (1.92min) +2026-06-20 00:20:08.401 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:22:07.126 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:29:07.002 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 98, train_steps: 3920 +2026-06-20 00:29:07.691 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 00:29:07.835 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:29:07.874 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:30:12.427 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:30:12.430 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:32:10.419 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 00:32:11.243 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:34:06.389 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 00:34:07.118 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:36:03.851 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:43:05.411 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 99, train_steps: 3960 +2026-06-20 00:43:06.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 00:43:06.241 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:43:06.335 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:44:10.845 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:44:10.848 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:46:09.899 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 00:46:10.749 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:48:05.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.75s (1.91min) +2026-06-20 00:48:06.440 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:50:03.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:57:04.799 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 100, train_steps: 4000 +2026-06-20 00:58:47.402 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=100 (missing=0, unexpected=0) +2026-06-20 00:58:49.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 00:58:49.663 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:58:49.725 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:59:55.757 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:59:55.761 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:01:50.303 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 01:01:51.131 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:03:44.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-20 01:03:45.434 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:05:42.219 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:12:39.865 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 101, train_steps: 4040 +2026-06-20 01:12:40.568 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 01:12:40.757 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:12:40.849 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:13:46.154 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:13:46.154 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:15:43.499 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 01:15:44.347 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:17:38.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.62s (1.89min) +2026-06-20 01:17:38.931 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:19:38.325 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:26:37.804 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 102, train_steps: 4080 +2026-06-20 01:26:38.504 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 01:26:38.660 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:26:38.711 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:27:43.045 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:27:43.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:29:41.183 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 01:29:42.553 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:31:36.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-20 01:31:37.098 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:33:37.226 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:40:38.539 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 103, train_steps: 4120 +2026-06-20 01:40:39.244 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 01:40:39.412 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:40:39.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:41:44.167 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:41:44.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:43:43.334 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 01:43:44.340 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:45:37.767 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-20 01:45:38.488 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:47:38.268 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:54:37.790 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 104, train_steps: 4160 +2026-06-20 01:54:38.703 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 01:54:38.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:54:38.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:55:43.656 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:55:43.656 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:57:42.889 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 01:57:43.887 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:59:37.314 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-20 01:59:38.031 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:01:38.022 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:08:38.450 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 105, train_steps: 4200 +2026-06-20 02:08:39.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 02:08:39.337 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:08:39.446 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:09:43.801 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:09:43.803 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:11:43.945 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 02:11:44.919 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:13:38.217 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 02:13:38.926 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:15:38.270 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:22:37.856 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 106, train_steps: 4240 +2026-06-20 02:22:38.564 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 02:22:38.671 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:22:38.701 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:23:43.240 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:23:43.243 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:25:40.847 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 02:25:41.946 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:27:35.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.90s (1.90min) +2026-06-20 02:27:36.583 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:29:36.812 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:36:37.091 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 107, train_steps: 4280 +2026-06-20 02:36:37.802 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 02:36:37.997 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:36:38.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:37:42.212 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:37:42.215 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:39:39.229 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 02:39:40.226 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:41:35.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 02:41:36.072 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:43:35.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:50:35.012 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 108, train_steps: 4320 +2026-06-20 02:50:35.728 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 02:50:35.924 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:50:36.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:51:40.534 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:51:40.537 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:53:37.574 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 02:53:38.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:55:33.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.45s (1.91min) +2026-06-20 02:55:33.973 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:57:30.649 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:04:30.697 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 109, train_steps: 4360 +2026-06-20 03:04:31.405 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 03:04:31.544 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:04:31.578 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:05:35.983 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:05:35.986 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:07:33.407 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 03:07:34.233 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:09:29.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-20 03:09:30.014 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:11:26.576 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:18:25.329 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 110, train_steps: 4400 +2026-06-20 03:19:01.502 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=110 (missing=0, unexpected=0) +2026-06-20 03:19:03.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 03:19:03.373 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:19:03.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:20:10.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:20:10.074 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:22:07.524 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 03:22:08.513 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:24:03.966 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.41s (1.92min) +2026-06-20 03:24:04.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:26:02.476 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:33:05.396 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 111, train_steps: 4440 +2026-06-20 03:33:06.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 03:33:06.235 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:33:06.289 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:34:10.723 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:34:10.729 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:36:10.098 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 03:36:11.096 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:38:06.021 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 03:38:06.732 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:40:04.484 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:47:04.309 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 112, train_steps: 4480 +2026-06-20 03:47:05.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 03:47:05.231 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:47:05.325 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:48:09.578 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:48:09.581 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:50:04.983 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 03:50:06.257 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:52:00.943 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-20 03:52:01.658 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:54:00.964 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:01:04.007 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 113, train_steps: 4520 +2026-06-20 04:01:04.874 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 04:01:05.015 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:01:05.060 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:02:09.371 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:02:09.374 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:04:06.611 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 04:04:07.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:06:02.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.15s (1.92min) +2026-06-20 04:06:03.532 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:08:01.584 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:15:03.779 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 114, train_steps: 4560 +2026-06-20 04:15:04.474 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 04:15:04.660 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:15:04.760 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:16:09.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:16:09.027 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:18:04.152 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 04:18:05.160 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:20:00.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.72s (1.91min) +2026-06-20 04:20:00.895 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:22:00.100 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:29:01.468 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 115, train_steps: 4600 +2026-06-20 04:29:02.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 04:29:02.376 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:29:02.456 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:30:06.669 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:30:06.672 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:32:02.494 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 04:32:03.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:33:58.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.33s (1.91min) +2026-06-20 04:33:58.776 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:35:55.624 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:42:56.469 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 116, train_steps: 4640 +2026-06-20 04:42:57.160 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 04:42:57.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:42:57.461 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:44:01.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:44:01.824 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:45:58.380 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 04:45:59.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:47:53.698 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.25s (1.90min) +2026-06-20 04:47:54.408 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:49:50.889 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:56:50.227 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 117, train_steps: 4680 +2026-06-20 04:56:50.960 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 04:56:51.154 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:56:51.285 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:57:55.524 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:57:55.527 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:59:51.798 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 04:59:52.625 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:01:46.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.49s (1.89min) +2026-06-20 05:01:47.078 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:03:44.144 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:10:42.590 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 118, train_steps: 4720 +2026-06-20 05:10:43.295 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 05:10:43.431 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:10:43.465 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:11:48.972 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:11:48.973 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:13:46.957 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 05:13:47.800 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:15:41.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.26s (1.89min) +2026-06-20 05:15:42.001 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:17:38.995 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:24:37.900 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 119, train_steps: 4760 +2026-06-20 05:24:38.622 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 05:24:38.774 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:24:38.836 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:25:43.184 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:25:43.186 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:27:40.119 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 05:27:41.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:29:34.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-20 05:29:35.560 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:31:33.107 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:38:32.143 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 120, train_steps: 4800 +2026-06-20 05:39:09.432 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=120 (missing=0, unexpected=0) +2026-06-20 05:39:10.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 05:39:10.990 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:39:11.022 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:40:17.654 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:40:17.658 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:42:12.566 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 05:42:13.551 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:44:08.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.96s (1.92min) +2026-06-20 05:44:09.427 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:46:10.089 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:53:12.053 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 121, train_steps: 4840 +2026-06-20 05:53:12.775 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 05:53:12.923 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:53:12.954 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:54:17.820 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:54:17.822 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:56:13.156 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 05:56:14.160 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:58:09.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.97s (1.92min) +2026-06-20 05:58:10.076 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:00:10.575 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:07:11.955 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 122, train_steps: 4880 +2026-06-20 06:07:12.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 06:07:12.749 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:07:12.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:08:17.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:08:17.236 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:10:13.195 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 06:10:14.020 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:12:08.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.25s (1.90min) +2026-06-20 06:12:09.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:14:10.580 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:21:11.866 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 123, train_steps: 4920 +2026-06-20 06:21:12.561 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 06:21:12.731 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:21:12.814 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:22:16.914 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:22:16.914 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:24:12.071 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 06:24:13.096 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:26:08.788 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.47s (1.92min) +2026-06-20 06:26:09.512 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:28:09.834 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:35:11.845 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 124, train_steps: 4960 +2026-06-20 06:35:12.545 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 06:35:12.735 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:35:12.841 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:36:16.783 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:36:16.786 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:38:12.129 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 06:38:13.030 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:40:08.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.44s (1.92min) +2026-06-20 06:40:09.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:42:09.773 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:49:11.648 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 125, train_steps: 5000 +2026-06-20 06:50:15.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 06:50:15.361 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:50:15.426 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:51:19.726 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:51:19.729 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:53:15.926 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 06:53:16.901 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:55:11.654 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.68s (1.91min) +2026-06-20 06:55:12.372 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:57:11.766 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:04:11.919 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 126, train_steps: 5040 +2026-06-20 07:04:12.631 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 07:04:12.813 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:04:12.910 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:05:17.510 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:05:17.514 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:07:13.988 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 07:07:15.195 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:09:10.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.08s (1.92min) +2026-06-20 07:09:11.147 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:11:11.168 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:18:11.929 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 127, train_steps: 5080 +2026-06-20 07:18:12.629 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 07:18:12.735 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:18:12.771 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:19:17.149 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:19:17.152 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:21:13.050 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 07:21:13.944 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:23:09.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.50s (1.93min) +2026-06-20 07:23:10.327 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:25:10.158 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:32:11.169 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 128, train_steps: 5120 +2026-06-20 07:32:11.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 07:32:11.997 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:32:12.037 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:33:16.185 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:33:16.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:35:13.016 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 07:35:14.012 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:37:08.641 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.56s (1.91min) +2026-06-20 07:37:09.373 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:39:09.932 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:46:11.986 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 129, train_steps: 5160 +2026-06-20 07:46:12.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 07:46:12.868 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:46:12.926 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:47:17.581 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:47:17.584 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:49:13.070 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 07:49:14.049 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:51:09.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-20 07:51:09.759 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:53:10.306 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:00:11.780 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 130, train_steps: 5200 +2026-06-20 08:00:49.032 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=130 (missing=0, unexpected=0) +2026-06-20 08:00:51.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 08:00:51.419 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:00:51.485 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:01:58.744 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:01:58.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:03:52.509 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 08:03:53.387 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:05:47.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.72s (1.90min) +2026-06-20 08:05:48.018 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:07:44.847 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:14:45.316 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 131, train_steps: 5240 +2026-06-20 08:14:46.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 08:14:46.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:14:46.186 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:15:50.887 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:15:50.890 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:17:46.990 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 08:17:47.986 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:19:41.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.38s (1.89min) +2026-06-20 08:19:42.139 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:21:39.468 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:28:37.839 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 132, train_steps: 5280 +2026-06-20 08:28:38.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 08:28:38.658 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:28:38.690 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:29:43.557 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:29:43.560 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:31:40.849 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 08:31:41.961 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:33:35.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.44s (1.89min) +2026-06-20 08:33:36.119 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:35:35.265 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:42:34.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 133, train_steps: 5320 +2026-06-20 08:42:35.017 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 08:42:35.172 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:42:35.221 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:43:39.865 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:43:39.870 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:45:35.888 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 08:45:36.858 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:47:32.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.77s (1.93min) +2026-06-20 08:47:33.450 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:49:30.694 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:56:29.431 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 134, train_steps: 5360 +2026-06-20 08:56:30.113 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 08:56:30.215 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:56:30.260 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:57:34.747 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:57:34.750 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:59:32.041 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 08:59:33.020 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:01:26.319 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 09:01:27.049 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:03:23.808 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:10:22.982 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 135, train_steps: 5400 +2026-06-20 09:10:23.675 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 09:10:23.841 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:10:23.917 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:11:28.463 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:11:28.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:13:23.932 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 09:13:24.928 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:15:18.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.61s (1.89min) +2026-06-20 09:15:19.311 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:17:17.102 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:24:16.798 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 136, train_steps: 5440 +2026-06-20 09:24:17.531 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 09:24:17.637 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:24:17.680 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:25:22.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:25:22.272 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:27:18.656 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 09:27:19.659 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:29:12.959 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.15s (1.89min) +2026-06-20 09:29:13.687 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:31:13.177 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:38:13.213 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 137, train_steps: 5480 +2026-06-20 09:38:13.936 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 09:38:14.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:38:14.096 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:39:18.546 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:39:18.548 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:41:14.371 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 09:41:15.358 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:43:09.522 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-20 09:43:10.251 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:45:10.504 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:52:12.149 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 138, train_steps: 5520 +2026-06-20 09:52:12.866 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 09:52:13.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:52:13.104 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:53:17.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:53:17.752 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:55:13.383 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 09:55:14.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:57:10.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.46s (1.92min) +2026-06-20 09:57:10.921 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:59:11.152 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:06:12.672 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 139, train_steps: 5560 +2026-06-20 10:06:13.363 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 10:06:13.500 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:06:13.540 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:07:18.043 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:07:18.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:09:13.264 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 10:09:14.083 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:11:08.675 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.20s (1.90min) +2026-06-20 10:11:09.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:13:09.781 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:20:12.683 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 140, train_steps: 5600 +2026-06-20 10:20:49.522 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=140 (missing=0, unexpected=0) +2026-06-20 10:20:51.812 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 10:20:51.902 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:20:51.941 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:21:58.045 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:21:58.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:23:53.040 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 10:23:53.875 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:25:47.600 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-20 10:25:48.313 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:27:45.364 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:34:44.000 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 141, train_steps: 5640 +2026-06-20 10:34:44.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 10:34:44.785 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:34:44.833 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:35:49.364 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:35:49.368 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:37:46.846 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 10:37:47.849 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:39:41.415 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.51s (1.89min) +2026-06-20 10:39:42.149 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:41:39.465 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:48:38.463 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 142, train_steps: 5680 +2026-06-20 10:48:39.159 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 10:48:39.359 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:48:39.432 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:49:44.155 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:49:44.155 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:51:43.153 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 10:51:44.124 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:53:37.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-20 10:53:37.851 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:55:37.996 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:02:37.914 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 143, train_steps: 5720 +2026-06-20 11:02:38.632 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 11:02:38.784 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:02:38.834 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:03:43.565 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:03:43.644 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:05:42.169 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 11:05:43.198 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:07:36.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 11:07:37.313 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:09:36.843 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:16:37.098 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 144, train_steps: 5760 +2026-06-20 11:16:37.798 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 11:16:37.907 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:16:37.946 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:17:42.360 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:17:42.363 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:19:38.707 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 11:19:39.625 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:21:33.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.47s (1.89min) +2026-06-20 11:21:34.087 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:23:30.885 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:30:30.407 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 145, train_steps: 5800 +2026-06-20 11:30:31.097 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 11:30:31.263 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:30:31.344 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:31:35.656 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:31:35.658 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:33:33.207 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 11:33:34.042 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:35:28.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.89s (1.90min) +2026-06-20 11:35:28.900 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:37:25.682 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:44:23.919 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 146, train_steps: 5840 +2026-06-20 11:44:24.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 11:44:24.743 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:44:24.786 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:45:29.316 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:45:29.325 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:47:23.247 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 11:47:24.234 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:49:19.134 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-20 11:49:19.857 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:51:17.043 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:58:15.924 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 147, train_steps: 5880 +2026-06-20 11:58:16.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 11:58:16.720 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:58:16.762 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:59:21.465 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:59:21.468 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:01:16.232 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 12:01:17.235 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:03:12.586 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.29s (1.92min) +2026-06-20 12:03:13.294 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:05:12.906 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:12:12.694 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 148, train_steps: 5920 +2026-06-20 12:12:13.404 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 12:12:13.512 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:12:13.546 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:13:18.139 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:13:18.140 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:15:13.144 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 12:15:14.153 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:17:09.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.69s (1.91min) +2026-06-20 12:17:09.789 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:19:10.590 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:26:11.387 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 149, train_steps: 5960 +2026-06-20 12:26:12.134 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 12:26:12.262 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:26:12.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:27:16.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:27:16.802 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:29:13.029 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 12:29:14.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:31:08.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.72s (1.91min) +2026-06-20 12:31:09.532 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:33:09.830 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:40:10.988 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 150, train_steps: 6000 +2026-06-20 12:41:53.657 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=150 (missing=0, unexpected=0) +2026-06-20 12:41:56.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 12:41:56.305 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:41:56.409 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:43:02.454 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:43:02.459 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:44:59.744 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 12:45:00.737 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:46:55.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.70s (1.91min) +2026-06-20 12:46:56.236 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:48:52.991 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:55:51.600 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 151, train_steps: 6040 +2026-06-20 12:55:52.307 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 12:55:52.465 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:55:52.518 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:56:56.864 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:56:56.868 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:58:50.721 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 12:58:51.706 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:00:46.190 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.41s (1.91min) +2026-06-20 13:00:46.908 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:02:43.655 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:09:41.900 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 152, train_steps: 6080 +2026-06-20 13:09:42.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 13:09:42.766 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:09:42.839 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:10:48.525 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:10:48.525 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:12:43.910 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 13:12:44.896 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:14:38.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-20 13:14:39.492 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:16:38.553 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:23:37.216 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 153, train_steps: 6120 +2026-06-20 13:23:38.094 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 13:23:38.262 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:23:38.315 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:24:43.318 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:24:43.318 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:26:40.127 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 13:26:41.254 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:28:34.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.39s (1.89min) +2026-06-20 13:28:35.473 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:30:32.910 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:37:31.693 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 154, train_steps: 6160 +2026-06-20 13:37:32.394 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 13:37:32.572 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:37:32.646 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:38:36.936 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:38:36.936 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:40:33.789 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 13:40:34.792 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:42:29.329 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.47s (1.91min) +2026-06-20 13:42:30.062 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:44:27.160 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:51:24.973 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 155, train_steps: 6200 +2026-06-20 13:51:25.702 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 13:51:25.897 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:51:25.990 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:52:30.312 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:52:30.318 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:54:24.447 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 13:54:25.445 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:56:19.129 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.63s (1.89min) +2026-06-20 13:56:19.859 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:58:16.753 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:05:15.340 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 156, train_steps: 6240 +2026-06-20 14:05:16.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 14:05:16.140 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:05:16.177 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:06:21.116 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:06:21.117 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:08:15.931 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 14:08:16.940 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:10:11.722 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.74s (1.91min) +2026-06-20 14:10:12.457 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:12:11.106 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:19:12.237 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 157, train_steps: 6280 +2026-06-20 14:19:12.924 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 14:19:13.081 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:19:13.135 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:20:17.650 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:20:17.653 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:22:12.398 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 14:22:13.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:24:08.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.22s (1.92min) +2026-06-20 14:24:09.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:26:09.908 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:33:10.825 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 158, train_steps: 6320 +2026-06-20 14:33:11.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 14:33:11.718 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:33:11.829 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:34:15.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:34:15.864 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:36:10.935 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 14:36:12.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:38:07.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.88s (1.91min) +2026-06-20 14:38:07.785 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:40:05.260 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:47:04.564 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 159, train_steps: 6360 +2026-06-20 14:47:05.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 14:47:05.425 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:47:05.464 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:48:09.812 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:48:09.819 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:50:04.048 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 14:50:05.045 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:51:59.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.86s (1.90min) +2026-06-20 14:51:59.929 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:53:57.322 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:00:59.176 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 160, train_steps: 6400 +2026-06-20 15:01:36.535 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=160 (missing=0, unexpected=0) +2026-06-20 15:01:39.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 15:01:39.414 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:01:39.456 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:02:45.220 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:02:45.221 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:04:42.879 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 15:04:43.695 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:06:37.409 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.48s (1.89min) +2026-06-20 15:06:38.136 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:08:38.301 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:15:37.522 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 161, train_steps: 6440 +2026-06-20 15:15:38.207 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 15:15:38.331 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:15:38.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:16:42.689 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:16:42.692 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:18:40.375 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 15:18:41.389 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:20:35.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-20 15:20:36.332 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:22:36.148 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:29:36.891 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 162, train_steps: 6480 +2026-06-20 15:29:37.569 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 15:29:37.713 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:29:37.744 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:30:42.060 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:30:42.063 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:32:37.444 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 15:32:38.277 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:34:33.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.48s (1.91min) +2026-06-20 15:34:33.827 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:36:31.318 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:43:31.478 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 163, train_steps: 6520 +2026-06-20 15:43:32.171 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 15:43:32.322 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:43:32.379 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:44:36.945 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:44:36.949 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:46:33.676 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 15:46:34.508 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:48:29.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.47s (1.91min) +2026-06-20 15:48:29.934 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:50:26.539 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:57:25.506 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 164, train_steps: 6560 +2026-06-20 15:57:26.227 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 15:57:26.347 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:57:26.387 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:58:30.762 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:58:30.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:00:26.402 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 16:00:27.217 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:02:21.064 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.61s (1.89min) +2026-06-20 16:02:21.793 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:04:18.343 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:11:18.368 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 165, train_steps: 6600 +2026-06-20 16:11:19.087 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 16:11:19.243 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:11:19.294 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:12:23.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:12:23.883 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:14:19.106 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 16:14:20.253 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:16:14.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.75s (1.90min) +2026-06-20 16:16:14.870 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:18:14.128 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:25:13.928 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 166, train_steps: 6640 +2026-06-20 16:25:14.619 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 16:25:14.722 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:25:14.761 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:26:19.421 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:26:19.425 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:28:13.831 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 16:28:14.934 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:30:08.844 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.78s (1.90min) +2026-06-20 16:30:09.574 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:32:10.310 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:39:12.151 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 167, train_steps: 6680 +2026-06-20 16:39:12.849 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 16:39:13.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:39:13.106 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:40:17.656 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:40:17.663 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:42:12.095 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 16:42:13.326 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:44:08.745 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.36s (1.92min) +2026-06-20 16:44:09.477 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:46:10.023 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:53:11.378 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 168, train_steps: 6720 +2026-06-20 16:53:12.089 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 16:53:12.210 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:53:12.252 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:54:16.707 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:54:16.710 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:56:11.767 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 16:56:12.763 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:58:07.736 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-20 16:58:08.464 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:00:07.581 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:07:09.177 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 169, train_steps: 6760 +2026-06-20 17:07:09.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 17:07:10.046 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:07:10.110 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:08:14.780 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:08:14.786 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:10:11.904 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 17:10:12.892 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:12:08.255 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.20s (1.92min) +2026-06-20 17:12:08.969 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:14:09.322 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:21:11.045 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 170, train_steps: 6800 +2026-06-20 17:21:47.177 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=170 (missing=0, unexpected=0) +2026-06-20 17:21:49.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 17:21:49.379 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:21:49.502 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:22:54.949 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:22:54.952 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:24:49.925 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 17:24:50.922 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:26:44.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.05s (1.88min) +2026-06-20 17:26:44.743 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:28:41.428 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:35:39.739 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 45, sp_rank 0: update step: 171, train_steps: 6840 +2026-06-20 17:35:40.431 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 45, dp_rank 22: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=11, rank_in_group=0 +2026-06-20 17:35:40.567 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:35:40.613 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:36:45.774 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 45 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:36:45.777 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 45 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:38:42.537 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 45: Using local metric names without global sync: ['avg'] +2026-06-20 17:38:43.710 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 45: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:40:36.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 45: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.89s (1.88min) +2026-06-20 17:40:37.378 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 45, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-29 01:28:58.409 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 01:28:58.410 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 01:51:27.847 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 01:51:27.848 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 02:20:50.797 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 02:20:50.798 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 02:45:59.525 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 02:45:59.527 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:07:53.652 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:07:53.653 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:31:52.394 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:31:52.395 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:59:04.105 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:59:04.106 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:27:01.582 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:27:01.585 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:52:07.185 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:52:07.186 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 05:16:39.682 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 05:16:39.684 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 05:39:28.279 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 05:39:28.279 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:02:48.333 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:02:48.334 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:25:38.867 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:25:38.868 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:48:26.589 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:48:26.592 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 07:12:16.760 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 07:12:16.761 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 07:34:14.665 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 07:34:14.666 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:05:09.648 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:05:09.651 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:29:49.894 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:29:49.896 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:59:02.237 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:59:02.237 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:22:53.736 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:22:53.737 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:47:19.833 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:47:19.833 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 10:10:29.095 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 10:10:29.095 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 10:36:27.959 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 10:36:27.960 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:02:30.420 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:02:30.421 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:27:31.560 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:27:31.560 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:56:52.733 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:56:52.734 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 12:18:29.850 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 12:18:29.852 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 12:46:40.203 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 12:46:40.204 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 13:12:54.772 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 13:12:54.774 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 13:38:57.830 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 13:38:57.831 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:04:55.409 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:04:55.410 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:27:06.326 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:27:06.327 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:56:30.028 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:56:30.029 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 15:21:05.883 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 15:21:05.884 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 15:46:59.963 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 15:46:59.964 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 16:17:02.826 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 16:17:02.826 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 16:41:03.545 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 16:41:03.545 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:09:23.592 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:09:23.593 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:32:09.720 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:32:09.727 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:53:40.560 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:53:40.563 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 18:16:06.567 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 18:16:06.568 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 18:38:52.296 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 18:38:52.296 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_53.log b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_53.log new file mode 100644 index 0000000000000000000000000000000000000000..85d00f96d96b5006d9e323acb06b27afe50f9c69 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_53.log @@ -0,0 +1,1994 @@ +2026-06-19 01:23:02.669 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-19 01:23:02.670 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-19 01:30:17.294 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 01:30:17.401 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:30:17.405 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:31:51.998 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:31:52.000 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:33:45.514 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 01:34:00.984 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 01:35:56.997 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.94s (1.93min) +2026-06-19 01:35:57.694 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 01:37:54.193 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 01:37:57.225 | INFO | hyvideo.utils.grpo_utils:train_one_step:3127 - Rank 53, sp_rank 0: init adaptive KL controller (init=1.000e-07, min=1.000e-07, max=1.000e-05) +2026-06-19 01:45:11.462 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 1, train_steps: 40 +2026-06-19 01:45:12.575 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 01:45:12.736 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:45:12.802 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:46:17.907 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:46:17.907 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:48:10.944 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 01:48:11.836 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 01:50:07.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.70s (1.93min) +2026-06-19 01:50:08.465 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 01:52:06.956 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 01:59:06.755 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 2, train_steps: 80 +2026-06-19 01:59:07.380 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 01:59:07.476 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:59:07.484 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:00:11.695 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:00:11.697 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:02:10.528 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 02:02:11.517 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:04:06.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.88s (1.91min) +2026-06-19 02:04:07.374 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:06:03.755 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:13:04.851 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 3, train_steps: 120 +2026-06-19 02:13:05.466 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 02:13:05.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:13:05.689 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:14:10.090 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:14:10.092 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:16:04.655 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 02:16:05.665 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:17:59.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.77s (1.90min) +2026-06-19 02:18:00.448 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:19:59.351 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:26:59.875 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 4, train_steps: 160 +2026-06-19 02:27:00.546 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 02:27:00.664 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:27:00.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:28:05.330 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:28:05.342 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:30:00.316 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 02:30:01.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:31:55.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.62s (1.91min) +2026-06-19 02:31:56.652 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:33:53.225 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:40:53.226 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 5, train_steps: 200 +2026-06-19 02:40:53.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 02:40:54.051 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:40:54.119 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:41:58.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:41:58.774 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:43:51.811 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 02:43:52.660 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:45:46.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.50s (1.89min) +2026-06-19 02:45:46.993 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:47:43.354 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:54:40.650 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 6, train_steps: 240 +2026-06-19 02:54:41.361 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 02:54:41.539 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:54:41.877 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:55:46.617 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:55:46.621 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:57:42.949 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 02:57:43.754 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:59:37.353 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.38s (1.89min) +2026-06-19 02:59:38.043 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:01:37.732 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:08:37.474 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 7, train_steps: 280 +2026-06-19 03:08:38.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 03:08:38.235 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:08:38.277 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:09:42.707 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:09:42.710 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:11:38.337 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 03:11:39.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:13:34.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.68s (1.91min) +2026-06-19 03:13:34.976 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:15:32.044 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:22:31.145 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 8, train_steps: 320 +2026-06-19 03:22:31.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 03:22:32.002 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:22:32.105 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:23:37.139 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:23:37.141 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:25:33.100 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 03:25:33.889 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:27:28.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-19 03:27:29.431 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:29:26.905 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:36:25.643 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 9, train_steps: 360 +2026-06-19 03:36:26.346 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 03:36:26.458 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:36:26.497 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:37:31.062 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:37:31.066 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:39:26.010 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 03:39:26.809 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:41:20.400 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.36s (1.89min) +2026-06-19 03:41:21.105 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:43:17.561 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:50:16.925 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 10, train_steps: 400 +2026-06-19 03:50:52.778 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=10 (missing=0, unexpected=0) +2026-06-19 03:50:56.181 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 03:50:56.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:50:56.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:52:02.226 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:52:02.229 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:53:57.828 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 03:53:58.625 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:55:52.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.65s (1.89min) +2026-06-19 03:55:53.208 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:57:49.766 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:04:48.380 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 11, train_steps: 440 +2026-06-19 04:04:49.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 04:04:49.203 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:04:49.239 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:05:53.683 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:05:53.690 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:07:48.852 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 04:07:49.803 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:09:43.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.25s (1.89min) +2026-06-19 04:09:43.797 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:11:40.269 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:18:37.285 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 12, train_steps: 480 +2026-06-19 04:18:37.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 04:18:38.122 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:18:38.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:19:42.672 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:19:42.672 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:21:37.180 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 04:21:37.970 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:23:33.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.19s (1.92min) +2026-06-19 04:23:34.070 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:25:31.244 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:32:29.810 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 13, train_steps: 520 +2026-06-19 04:32:30.512 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 04:32:30.620 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:32:30.666 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:33:35.112 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:33:35.117 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:35:30.312 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 04:35:31.098 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:37:24.102 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.77s (1.88min) +2026-06-19 04:37:24.791 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:39:21.077 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:46:18.581 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 14, train_steps: 560 +2026-06-19 04:46:19.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 04:46:19.406 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:46:19.483 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:47:24.546 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:47:24.549 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:49:17.445 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 04:49:18.232 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:51:12.963 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.49s (1.91min) +2026-06-19 04:51:13.650 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:53:12.995 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:00:11.824 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 15, train_steps: 600 +2026-06-19 05:00:12.525 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 05:00:12.688 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:00:12.763 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:01:18.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:01:18.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:03:11.349 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 05:03:12.446 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:05:07.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.64s (1.91min) +2026-06-19 05:05:08.060 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:07:05.978 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:14:04.901 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 16, train_steps: 640 +2026-06-19 05:14:05.573 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 05:14:05.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:14:05.820 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:15:10.255 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:15:10.258 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:17:06.988 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 05:17:07.921 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:19:03.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.02s (1.92min) +2026-06-19 05:19:03.741 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:21:01.963 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:28:03.335 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 17, train_steps: 680 +2026-06-19 05:28:04.006 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 05:28:04.123 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:28:04.166 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:29:08.809 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:29:08.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:31:03.926 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 05:31:04.987 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:32:58.689 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.48s (1.89min) +2026-06-19 05:32:59.386 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:34:56.268 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:41:58.106 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 18, train_steps: 720 +2026-06-19 05:41:58.777 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 05:41:58.921 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:41:58.958 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:43:03.710 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:43:03.714 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:45:00.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 05:45:01.065 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:46:55.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.88s (1.90min) +2026-06-19 05:46:55.797 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:48:52.332 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:55:50.502 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 19, train_steps: 760 +2026-06-19 05:55:51.178 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 05:55:51.338 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:55:51.409 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:56:55.827 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:56:55.830 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:58:49.068 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 05:58:50.014 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:00:43.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.10s (1.89min) +2026-06-19 06:00:43.876 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:02:40.343 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:09:39.758 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 20, train_steps: 800 +2026-06-19 06:10:16.133 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=20 (missing=0, unexpected=0) +2026-06-19 06:10:18.747 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 06:10:18.863 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:10:18.913 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:11:25.120 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:11:25.124 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:13:19.050 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 06:13:20.114 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:15:13.224 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.90s (1.88min) +2026-06-19 06:15:13.910 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:17:12.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:24:12.562 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 21, train_steps: 840 +2026-06-19 06:24:13.270 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 06:24:13.427 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:24:13.494 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:25:18.561 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:25:18.561 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:27:11.261 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 06:27:12.455 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:29:06.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.51s (1.91min) +2026-06-19 06:29:07.667 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:31:04.581 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:38:04.282 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 22, train_steps: 880 +2026-06-19 06:38:04.947 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 06:38:05.121 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:38:05.227 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:39:09.501 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:39:09.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:41:03.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 06:41:04.274 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:42:58.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.99s (1.90min) +2026-06-19 06:42:59.165 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:44:56.507 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:51:58.189 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 23, train_steps: 920 +2026-06-19 06:51:58.842 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 06:51:58.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:51:58.989 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:53:03.358 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:53:03.361 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:54:59.789 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 06:55:00.641 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:56:54.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.11s (1.90min) +2026-06-19 06:56:55.622 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:58:52.321 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:05:51.317 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 24, train_steps: 960 +2026-06-19 07:05:52.005 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 07:05:52.216 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:05:52.329 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:06:56.770 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:06:56.770 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:08:52.217 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 07:08:52.993 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:10:46.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-19 07:10:46.871 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:12:44.156 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:19:41.965 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 25, train_steps: 1000 +2026-06-19 07:20:52.104 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 07:20:52.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:20:52.285 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:21:56.869 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:21:56.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:23:52.666 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 07:23:53.640 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:25:46.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.81s (1.88min) +2026-06-19 07:25:47.205 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:27:44.384 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:34:43.688 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 26, train_steps: 1040 +2026-06-19 07:34:44.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 07:34:44.537 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:34:44.636 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:35:48.978 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:35:48.982 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:37:45.443 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 07:37:46.416 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:39:39.638 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.14s (1.89min) +2026-06-19 07:39:40.348 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:41:39.074 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:48:39.737 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 27, train_steps: 1080 +2026-06-19 07:48:40.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 07:48:40.500 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:48:40.569 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:49:45.103 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:49:45.104 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:51:43.285 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 07:51:44.079 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:53:38.298 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.99s (1.90min) +2026-06-19 07:53:39.007 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:55:38.424 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:02:38.177 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 28, train_steps: 1120 +2026-06-19 08:02:38.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 08:02:38.997 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:02:39.029 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:03:43.713 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:03:43.718 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:05:41.901 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 08:05:43.054 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:07:37.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.24s (1.90min) +2026-06-19 08:07:38.091 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:09:37.937 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:16:37.856 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 29, train_steps: 1160 +2026-06-19 08:16:38.691 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 08:16:38.841 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:16:38.891 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:17:43.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:17:43.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:19:42.881 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 08:19:43.829 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:21:37.264 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 08:21:37.961 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:23:37.409 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:30:39.552 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 30, train_steps: 1200 +2026-06-19 08:31:15.564 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=30 (missing=0, unexpected=0) +2026-06-19 08:31:18.258 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 08:31:18.437 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:31:18.545 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:32:25.074 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:32:25.077 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:34:19.980 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 08:34:21.097 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:36:14.510 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-19 08:36:15.226 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:38:14.176 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:45:14.279 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 31, train_steps: 1240 +2026-06-19 08:45:14.992 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 08:45:15.143 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:45:15.202 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:46:19.810 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:46:19.813 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:48:15.288 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 08:48:16.241 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:50:10.322 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.84s (1.90min) +2026-06-19 08:50:11.023 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:52:11.151 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:59:13.091 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 32, train_steps: 1280 +2026-06-19 08:59:13.775 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 08:59:13.890 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:59:13.932 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:00:19.108 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:00:19.108 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:02:13.436 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 09:02:14.443 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:04:08.713 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.91s (1.90min) +2026-06-19 09:04:09.421 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:06:10.409 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:13:13.271 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 33, train_steps: 1320 +2026-06-19 09:13:13.953 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 09:13:14.153 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:13:14.271 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:14:19.168 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:14:19.169 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:16:14.822 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 09:16:15.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:18:11.066 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.06s (1.92min) +2026-06-19 09:18:11.785 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:20:11.628 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:27:13.723 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 34, train_steps: 1360 +2026-06-19 09:27:14.445 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 09:27:14.621 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:27:14.714 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:28:19.528 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:28:19.531 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:30:13.695 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 09:30:14.773 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:32:08.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.81s (1.90min) +2026-06-19 09:32:09.592 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:34:09.780 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:41:12.391 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 35, train_steps: 1400 +2026-06-19 09:41:13.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 09:41:13.216 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:41:13.286 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:42:18.563 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:42:18.563 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:44:13.080 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 09:44:14.063 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:46:10.686 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 116.56s (1.94min) +2026-06-19 09:46:11.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:48:10.898 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:55:11.966 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 36, train_steps: 1440 +2026-06-19 09:55:12.653 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 09:55:12.849 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:55:12.977 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:56:18.235 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:56:18.236 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:58:12.646 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 09:58:13.830 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:00:08.966 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-19 10:00:09.671 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:02:10.011 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:09:11.712 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 37, train_steps: 1480 +2026-06-19 10:09:12.395 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 10:09:12.490 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:09:12.543 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:10:17.851 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:10:17.855 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:12:13.042 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 10:12:13.896 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:14:09.029 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.94s (1.92min) +2026-06-19 10:14:09.737 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:16:10.370 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:23:12.343 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 38, train_steps: 1520 +2026-06-19 10:23:13.053 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 10:23:13.244 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:23:13.371 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:24:18.645 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:24:18.646 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:26:12.667 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 10:26:13.823 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:28:09.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.16s (1.92min) +2026-06-19 10:28:09.955 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:30:10.746 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:37:12.065 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 39, train_steps: 1560 +2026-06-19 10:37:12.646 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 10:37:12.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:37:12.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:38:18.654 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:38:18.654 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:41:42.526 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 10:41:43.679 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:43:36.471 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.70s (1.88min) +2026-06-19 10:43:37.180 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:45:37.179 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:52:37.858 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 40, train_steps: 1600 +2026-06-19 10:53:14.100 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=40 (missing=0, unexpected=0) +2026-06-19 10:53:15.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 10:53:15.875 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:53:15.914 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:54:21.898 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:54:21.898 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:56:16.410 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 10:56:17.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:58:11.717 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.27s (1.90min) +2026-06-19 10:58:12.415 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:00:11.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:07:12.576 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 41, train_steps: 1640 +2026-06-19 11:07:13.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 11:07:13.503 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:07:13.625 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:08:18.646 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:08:18.646 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:10:12.668 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 11:10:13.862 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:12:08.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.28s (1.90min) +2026-06-19 11:12:09.087 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:14:08.759 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:21:10.881 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 42, train_steps: 1680 +2026-06-19 11:21:11.804 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 11:21:11.965 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:21:12.022 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:22:17.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:22:17.318 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:24:12.073 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 11:24:13.017 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:26:07.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.32s (1.91min) +2026-06-19 11:26:08.278 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:28:06.795 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:35:06.766 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 43, train_steps: 1720 +2026-06-19 11:35:07.460 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 11:35:07.568 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:35:07.614 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:36:12.006 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:36:12.010 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:38:10.299 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 11:38:11.133 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:40:05.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.38s (1.91min) +2026-06-19 11:40:06.548 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:42:03.525 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:49:04.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 44, train_steps: 1760 +2026-06-19 11:49:05.113 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 11:49:05.266 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:49:05.324 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:50:09.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:50:09.870 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:52:06.491 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 11:52:07.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:54:01.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.69s (1.89min) +2026-06-19 11:54:01.909 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:56:01.235 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:03:03.867 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 45, train_steps: 1800 +2026-06-19 12:03:04.558 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 12:03:04.777 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:03:04.903 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:04:09.228 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:04:09.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:06:05.347 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 12:06:06.142 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:08:01.417 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.05s (1.92min) +2026-06-19 12:08:02.101 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:10:01.382 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:17:03.584 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 46, train_steps: 1840 +2026-06-19 12:17:04.281 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 12:17:04.484 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:17:04.582 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:18:09.088 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:18:09.088 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:20:05.333 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 12:20:06.340 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:22:01.733 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.17s (1.92min) +2026-06-19 12:22:02.441 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:24:01.451 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:31:02.918 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 47, train_steps: 1880 +2026-06-19 12:31:03.607 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 12:31:03.805 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:31:03.934 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:32:08.281 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:32:08.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:34:03.182 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 12:34:04.148 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:35:58.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-19 12:35:58.982 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:37:55.893 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:44:57.685 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 48, train_steps: 1920 +2026-06-19 12:44:58.355 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 12:44:58.502 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:44:58.560 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:46:03.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:46:03.131 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:47:58.930 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 12:47:59.917 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:49:54.620 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-19 12:49:55.342 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:51:52.292 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:58:52.095 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 49, train_steps: 1960 +2026-06-19 12:58:52.782 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 12:58:52.901 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:58:52.939 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:59:57.534 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:59:57.537 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:01:54.369 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 13:01:55.217 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:03:48.446 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.04s (1.88min) +2026-06-19 13:03:49.148 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:05:45.468 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:12:45.266 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 50, train_steps: 2000 +2026-06-19 13:14:21.641 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=50 (missing=0, unexpected=0) +2026-06-19 13:14:23.203 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 13:14:23.354 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:14:23.430 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:15:30.361 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:15:30.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:17:25.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 13:17:26.194 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:19:20.081 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.84s (1.90min) +2026-06-19 13:19:20.807 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:21:17.528 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:28:18.473 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 51, train_steps: 2040 +2026-06-19 13:28:19.184 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 13:28:19.309 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:28:19.340 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:29:23.892 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:29:23.895 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:31:18.028 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 13:31:19.054 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:33:13.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.83s (1.90min) +2026-06-19 13:33:13.741 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:35:13.231 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:42:13.623 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 52, train_steps: 2080 +2026-06-19 13:42:14.276 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 13:42:14.414 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:42:14.458 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:43:19.602 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:43:19.606 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:45:15.287 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 13:45:16.373 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:47:10.481 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.03s (1.90min) +2026-06-19 13:47:11.196 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:49:11.267 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:56:12.606 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 53, train_steps: 2120 +2026-06-19 13:56:13.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 13:56:13.492 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:56:13.599 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:57:18.979 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:57:18.980 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:59:13.130 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 13:59:14.091 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:01:08.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.41s (1.91min) +2026-06-19 14:01:09.529 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:03:10.475 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:10:12.172 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 54, train_steps: 2160 +2026-06-19 14:10:12.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 14:10:13.045 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:10:13.126 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:11:18.310 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:11:18.310 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:13:12.866 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 14:13:13.731 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:15:08.493 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.42s (1.91min) +2026-06-19 14:15:09.217 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:17:09.307 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:24:11.901 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 55, train_steps: 2200 +2026-06-19 14:24:12.602 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 14:24:12.732 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:24:12.769 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:25:18.218 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:25:18.218 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:27:12.754 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 14:27:13.719 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:29:09.888 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.82s (1.93min) +2026-06-19 14:29:10.604 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:31:10.764 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:38:12.106 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 56, train_steps: 2240 +2026-06-19 14:38:12.811 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 14:38:12.970 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:38:13.028 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:39:18.184 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:39:18.187 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:41:13.092 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 14:41:14.063 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:43:08.781 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.66s (1.91min) +2026-06-19 14:43:09.475 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:45:09.720 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:52:11.853 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 57, train_steps: 2280 +2026-06-19 14:52:12.526 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 14:52:12.734 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:52:12.854 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:53:17.994 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:53:17.995 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:55:13.057 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 14:55:13.938 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:57:08.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.79s (1.91min) +2026-06-19 14:57:09.597 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:59:09.881 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:06:12.283 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 58, train_steps: 2320 +2026-06-19 15:06:12.987 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 15:06:13.195 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:06:13.320 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:07:18.621 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:07:18.621 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:09:13.062 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 15:09:13.869 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:11:08.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.71s (1.91min) +2026-06-19 15:11:09.498 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:13:09.810 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:20:12.047 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 59, train_steps: 2360 +2026-06-19 15:20:12.739 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 15:20:12.937 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:20:13.085 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:21:18.122 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:21:18.123 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:23:12.372 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 15:23:13.328 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:25:09.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.28s (1.92min) +2026-06-19 15:25:09.743 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:27:10.368 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:34:12.063 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 60, train_steps: 2400 +2026-06-19 15:34:48.002 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=60 (missing=0, unexpected=0) +2026-06-19 15:34:50.615 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 15:34:50.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:34:50.766 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:35:56.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:35:56.885 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:37:53.529 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 15:37:54.706 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:39:47.707 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-19 15:39:48.402 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:41:45.236 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:48:44.584 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 61, train_steps: 2440 +2026-06-19 15:48:45.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 15:48:45.401 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:48:45.433 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:49:50.141 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:49:50.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:51:47.609 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 15:51:48.463 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:53:41.681 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.04s (1.88min) +2026-06-19 15:53:42.400 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:55:39.328 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:02:38.995 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 62, train_steps: 2480 +2026-06-19 16:02:39.688 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 16:02:39.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:02:39.994 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:03:44.426 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:03:44.429 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:05:44.132 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 16:05:44.936 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:07:38.761 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.59s (1.89min) +2026-06-19 16:07:39.470 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:09:38.854 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:16:39.310 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 63, train_steps: 2520 +2026-06-19 16:16:40.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 16:16:40.226 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:16:40.326 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:17:44.887 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:17:44.890 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:19:45.197 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 16:19:46.010 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:21:39.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.36s (1.89min) +2026-06-19 16:21:40.302 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:23:38.883 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:30:40.928 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 64, train_steps: 2560 +2026-06-19 16:30:41.626 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 16:30:41.807 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:30:41.889 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:31:46.526 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:31:46.528 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:33:45.471 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 16:33:46.276 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:35:40.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.63s (1.89min) +2026-06-19 16:35:40.844 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:37:39.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:44:41.292 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 65, train_steps: 2600 +2026-06-19 16:44:41.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 16:44:42.152 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:44:42.259 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:45:46.980 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:45:46.983 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:47:47.482 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 16:47:48.289 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:49:42.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.72s (1.90min) +2026-06-19 16:49:42.958 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:51:40.037 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:58:41.275 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 66, train_steps: 2640 +2026-06-19 16:58:41.990 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 16:58:42.121 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:58:42.166 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:59:46.525 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:59:46.526 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:01:43.599 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 17:01:44.412 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:03:37.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.34s (1.89min) +2026-06-19 17:03:38.682 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:05:38.384 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:12:40.116 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 67, train_steps: 2680 +2026-06-19 17:12:40.908 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 17:12:41.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:12:41.444 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:13:46.029 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:13:46.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:15:44.529 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 17:15:45.339 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:17:38.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 17:17:39.678 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:19:39.113 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:26:39.543 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 68, train_steps: 2720 +2026-06-19 17:26:40.215 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 17:26:40.410 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:26:40.538 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:27:45.373 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:27:45.376 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:29:43.886 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 17:29:44.876 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:31:38.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.76s (1.90min) +2026-06-19 17:31:39.421 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:33:38.461 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:40:38.373 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 69, train_steps: 2760 +2026-06-19 17:40:39.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 17:40:39.174 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:40:39.205 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:41:43.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:41:43.621 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:43:41.069 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 17:43:42.126 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:45:35.854 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.55s (1.89min) +2026-06-19 17:45:36.569 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:47:37.118 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:54:38.583 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 70, train_steps: 2800 +2026-06-19 17:55:14.948 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=70 (missing=0, unexpected=0) +2026-06-19 17:55:16.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 17:55:17.112 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:55:17.159 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:56:23.640 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:56:23.642 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:58:18.228 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 17:58:19.328 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:00:13.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.58s (1.89min) +2026-06-19 18:00:13.838 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:02:13.854 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:09:13.889 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 71, train_steps: 2840 +2026-06-19 18:09:14.615 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 18:09:14.782 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:09:14.869 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:10:19.585 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:10:19.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:12:16.021 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 18:12:16.841 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:14:12.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.67s (1.93min) +2026-06-19 18:14:13.460 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:16:12.693 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:23:13.823 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 72, train_steps: 2880 +2026-06-19 18:23:14.506 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 18:23:14.699 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:23:14.813 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:24:19.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:24:19.618 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:26:14.711 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 18:26:15.673 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:28:10.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.45s (1.91min) +2026-06-19 18:28:11.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:30:10.990 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:37:12.781 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 73, train_steps: 2920 +2026-06-19 18:37:13.456 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 18:37:13.650 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:37:13.766 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:38:19.113 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:38:19.113 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:40:13.249 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 18:40:14.064 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:42:09.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.16s (1.92min) +2026-06-19 18:42:10.284 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:44:10.445 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:51:12.748 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 74, train_steps: 2960 +2026-06-19 18:51:13.468 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 18:51:13.682 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:51:13.830 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:52:19.233 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:52:19.234 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:54:13.982 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 18:54:14.805 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:56:10.312 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.21s (1.92min) +2026-06-19 18:56:11.031 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:58:11.444 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:05:14.005 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 75, train_steps: 3000 +2026-06-19 19:06:23.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 19:06:23.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:06:23.523 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:07:28.127 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:07:28.133 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:09:24.209 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 19:09:25.191 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:11:18.397 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.14s (1.89min) +2026-06-19 19:11:19.100 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:13:16.255 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:20:16.338 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 76, train_steps: 3040 +2026-06-19 19:20:16.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 19:20:17.110 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:20:17.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:21:21.501 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:21:21.504 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:23:18.455 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 19:23:19.610 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:25:13.441 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.66s (1.89min) +2026-06-19 19:25:14.145 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:27:12.736 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:34:13.273 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 77, train_steps: 3080 +2026-06-19 19:34:13.949 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 19:34:14.045 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:34:14.091 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:35:19.289 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:35:19.289 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:37:13.073 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 19:37:13.956 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:39:08.734 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.60s (1.91min) +2026-06-19 19:39:09.440 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:41:09.770 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:48:12.642 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 78, train_steps: 3120 +2026-06-19 19:48:13.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 19:48:13.486 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:48:13.518 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:49:19.165 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:49:19.166 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:51:13.323 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 19:51:14.143 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:53:09.267 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-19 19:53:09.975 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:55:09.986 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:02:13.169 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 79, train_steps: 3160 +2026-06-19 20:02:13.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 20:02:14.061 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:02:14.163 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:03:19.152 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:03:19.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:05:13.093 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 20:05:13.902 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:07:09.660 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.53s (1.93min) +2026-06-19 20:07:10.385 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:09:11.034 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:16:13.931 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 80, train_steps: 3200 +2026-06-19 20:16:50.588 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=80 (missing=0, unexpected=0) +2026-06-19 20:16:53.092 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 20:16:53.198 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:16:53.258 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:18:00.453 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:18:00.456 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:19:54.752 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 20:19:55.570 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:21:49.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-19 20:21:50.363 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:23:47.271 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:30:47.441 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 81, train_steps: 3240 +2026-06-19 20:30:48.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 20:30:48.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:30:48.489 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:31:53.249 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:31:53.252 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:33:49.273 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 20:33:50.262 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:35:44.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.70s (1.89min) +2026-06-19 20:35:44.752 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:37:41.680 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:44:40.001 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 82, train_steps: 3280 +2026-06-19 20:44:40.696 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 20:44:40.877 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:44:40.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:45:45.426 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:45:45.430 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:47:44.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 20:47:45.108 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:49:38.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.42s (1.89min) +2026-06-19 20:49:39.275 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:51:38.425 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:58:38.985 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 83, train_steps: 3320 +2026-06-19 20:58:39.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 20:58:40.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:58:40.170 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:59:44.544 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:59:44.547 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:01:44.303 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 21:01:45.113 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:03:38.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.28s (1.89min) +2026-06-19 21:03:39.339 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:05:38.820 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:12:38.921 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 84, train_steps: 3360 +2026-06-19 21:12:39.649 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 21:12:39.828 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:12:39.928 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:13:44.715 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:13:44.715 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:15:43.791 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 21:15:44.607 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:17:38.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.31s (1.89min) +2026-06-19 21:17:38.863 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:19:38.525 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:26:39.493 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 85, train_steps: 3400 +2026-06-19 21:26:40.168 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 21:26:40.320 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:26:40.367 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:27:45.269 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:27:45.272 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:29:43.697 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 21:29:44.513 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:31:38.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-19 21:31:38.988 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:33:38.757 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:40:38.643 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 86, train_steps: 3440 +2026-06-19 21:40:39.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 21:40:39.578 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:40:39.711 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:41:44.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:41:44.628 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:43:42.867 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 21:43:43.851 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:45:37.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 21:45:37.987 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:47:37.712 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:54:39.057 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 87, train_steps: 3480 +2026-06-19 21:54:39.743 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 21:54:39.865 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:54:39.908 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:55:44.291 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:55:44.294 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:57:43.169 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 21:57:44.153 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:59:37.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.39s (1.89min) +2026-06-19 21:59:38.317 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:01:38.557 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:08:39.991 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 88, train_steps: 3520 +2026-06-19 22:08:40.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 22:08:40.804 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:08:40.844 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:09:45.424 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:09:45.425 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:11:43.755 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 22:11:44.747 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:13:38.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.57s (1.89min) +2026-06-19 22:13:39.103 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:15:38.307 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:22:38.705 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 89, train_steps: 3560 +2026-06-19 22:22:39.424 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 22:22:39.604 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:22:39.719 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:23:44.209 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:23:44.213 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:25:43.683 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 22:25:44.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:27:38.162 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 22:27:38.886 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:29:38.496 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:36:39.435 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 90, train_steps: 3600 +2026-06-19 22:37:15.344 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=90 (missing=0, unexpected=0) +2026-06-19 22:37:17.307 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 22:37:17.537 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:37:17.650 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:38:23.270 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:38:23.273 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:40:19.240 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 22:40:20.251 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:42:14.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.70s (1.89min) +2026-06-19 22:42:14.862 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:44:13.691 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:51:14.858 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 91, train_steps: 3640 +2026-06-19 22:51:15.700 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 22:51:15.876 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:51:15.949 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:52:20.508 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:52:20.511 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:54:15.937 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 22:54:16.923 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:56:12.672 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.67s (1.93min) +2026-06-19 22:56:13.388 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:58:13.113 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:05:14.275 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 92, train_steps: 3680 +2026-06-19 23:05:15.118 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 23:05:15.346 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:05:15.475 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:06:19.956 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:06:19.959 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:08:16.058 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 23:08:17.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:10:11.751 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.62s (1.91min) +2026-06-19 23:10:12.471 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:12:11.285 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:19:13.487 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 93, train_steps: 3720 +2026-06-19 23:19:14.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 23:19:14.259 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:19:14.288 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:20:19.374 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:20:19.377 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:22:14.354 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 23:22:15.527 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:24:10.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.33s (1.92min) +2026-06-19 23:24:11.709 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:26:11.287 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:33:13.947 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 94, train_steps: 3760 +2026-06-19 23:33:14.640 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 23:33:14.792 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:33:14.858 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:34:19.747 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:34:19.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:36:16.135 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 23:36:17.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:38:12.062 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.89s (1.91min) +2026-06-19 23:38:12.784 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:40:11.322 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:47:11.880 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 95, train_steps: 3800 +2026-06-19 23:47:12.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-19 23:47:12.747 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:47:12.813 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:48:17.811 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:48:17.811 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:50:12.004 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-19 23:50:13.198 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:52:08.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.82s (1.91min) +2026-06-19 23:52:08.791 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:54:07.678 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:01:09.782 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 96, train_steps: 3840 +2026-06-20 00:01:10.475 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 00:01:10.635 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:01:10.787 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:02:15.692 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:02:15.694 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:04:11.783 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 00:04:13.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:06:07.928 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 00:06:08.644 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:08:07.691 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:15:08.511 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 97, train_steps: 3880 +2026-06-20 00:15:09.218 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 00:15:09.314 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:15:09.357 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:16:13.895 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:16:13.895 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:18:11.238 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 00:18:12.418 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:20:07.671 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.20s (1.92min) +2026-06-20 00:20:08.400 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:22:07.128 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:29:07.019 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 98, train_steps: 3920 +2026-06-20 00:29:07.692 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 00:29:07.848 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:29:07.903 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:30:12.464 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:30:12.472 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:32:10.419 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 00:32:11.541 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:34:06.389 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 00:34:07.117 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:36:03.853 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:43:05.351 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 99, train_steps: 3960 +2026-06-20 00:43:06.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 00:43:06.121 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:43:06.158 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:44:10.771 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:44:10.774 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:46:09.899 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 00:46:10.832 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:48:05.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.75s (1.91min) +2026-06-20 00:48:06.440 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:50:03.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:57:04.795 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 100, train_steps: 4000 +2026-06-20 00:58:47.542 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=100 (missing=0, unexpected=0) +2026-06-20 00:58:49.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 00:58:49.667 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:58:49.727 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:59:55.752 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:59:55.756 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:01:50.303 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 01:01:51.122 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:03:44.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-20 01:03:45.434 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:05:42.220 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:12:39.860 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 101, train_steps: 4040 +2026-06-20 01:12:40.567 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 01:12:40.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:12:40.761 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:13:45.760 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:13:45.763 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:15:43.499 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 01:15:44.320 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:17:38.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.62s (1.89min) +2026-06-20 01:17:38.931 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:19:38.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:26:37.819 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 102, train_steps: 4080 +2026-06-20 01:26:38.503 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 01:26:38.620 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:26:38.651 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:27:43.034 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:27:43.037 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:29:41.183 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 01:29:42.481 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:31:36.383 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-20 01:31:37.097 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:33:37.227 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:40:38.548 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 103, train_steps: 4120 +2026-06-20 01:40:39.250 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 01:40:39.455 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:40:39.590 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:41:44.141 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:41:44.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:43:43.334 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 01:43:44.338 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:45:37.767 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-20 01:45:38.487 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:47:38.267 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:54:37.794 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 104, train_steps: 4160 +2026-06-20 01:54:38.703 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 01:54:38.889 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:54:38.970 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:55:43.684 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:55:43.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:57:42.889 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 01:57:43.889 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:59:37.314 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-20 01:59:38.031 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:01:38.022 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:08:38.457 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 105, train_steps: 4200 +2026-06-20 02:08:39.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 02:08:39.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:08:39.286 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:09:44.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:09:44.028 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:11:43.945 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 02:11:44.933 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:13:38.217 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 02:13:38.925 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:15:38.271 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:22:37.867 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 106, train_steps: 4240 +2026-06-20 02:22:38.564 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 02:22:38.742 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:22:38.809 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:23:43.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:23:43.431 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:25:40.847 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 02:25:41.845 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:27:35.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.90s (1.90min) +2026-06-20 02:27:36.583 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:29:36.814 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:36:37.099 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 107, train_steps: 4280 +2026-06-20 02:36:37.801 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 02:36:37.950 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:36:38.015 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:37:42.562 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:37:42.566 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:39:39.229 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 02:39:40.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:41:35.347 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 02:41:36.072 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:43:35.149 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:50:35.009 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 108, train_steps: 4320 +2026-06-20 02:50:35.727 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 02:50:35.916 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:50:36.012 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:51:40.682 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:51:40.685 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:53:37.573 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 02:53:38.554 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:55:33.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.45s (1.91min) +2026-06-20 02:55:33.972 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:57:30.650 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:04:30.707 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 109, train_steps: 4360 +2026-06-20 03:04:31.405 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 03:04:31.557 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:04:31.623 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:05:35.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:05:35.970 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:07:33.407 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 03:07:34.397 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:09:29.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-20 03:09:30.014 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:11:26.577 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:18:25.336 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 110, train_steps: 4400 +2026-06-20 03:19:01.641 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=110 (missing=0, unexpected=0) +2026-06-20 03:19:03.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 03:19:03.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:19:03.434 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:20:09.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:20:09.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:22:07.524 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 03:22:08.492 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:24:03.966 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.41s (1.92min) +2026-06-20 03:24:04.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:26:02.474 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:33:05.389 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 111, train_steps: 4440 +2026-06-20 03:33:06.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 03:33:06.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:33:06.323 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:34:10.839 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:34:10.842 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:36:10.098 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 03:36:11.141 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:38:06.021 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 03:38:06.732 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:40:04.481 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:47:04.305 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 112, train_steps: 4480 +2026-06-20 03:47:05.043 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 03:47:05.183 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:47:05.215 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:48:09.685 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:48:09.686 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:50:04.983 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 03:50:06.091 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:52:00.943 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-20 03:52:01.657 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:54:00.965 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:01:04.001 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 113, train_steps: 4520 +2026-06-20 04:01:04.875 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 04:01:05.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:01:05.166 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:02:09.405 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:02:09.409 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:04:06.611 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 04:04:07.606 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:06:02.814 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.15s (1.92min) +2026-06-20 04:06:03.532 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:08:01.585 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:15:03.787 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 114, train_steps: 4560 +2026-06-20 04:15:04.474 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 04:15:04.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:15:04.784 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:16:09.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:16:09.183 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:18:04.152 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 04:18:05.314 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:20:00.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.72s (1.91min) +2026-06-20 04:20:00.895 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:22:00.102 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:29:01.484 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 115, train_steps: 4600 +2026-06-20 04:29:02.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 04:29:02.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:29:02.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:30:06.904 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:30:06.907 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:32:02.494 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 04:32:03.357 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:33:58.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.33s (1.91min) +2026-06-20 04:33:58.775 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:35:55.627 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:42:56.478 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 116, train_steps: 4640 +2026-06-20 04:42:57.159 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 04:42:57.296 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:42:57.338 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:44:02.181 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:44:02.181 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:45:58.380 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 04:45:59.374 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:47:53.698 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.25s (1.90min) +2026-06-20 04:47:54.407 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:49:50.891 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:56:50.240 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 117, train_steps: 4680 +2026-06-20 04:56:50.952 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 04:56:51.051 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:56:51.083 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:57:55.705 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:57:55.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:59:51.798 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 04:59:52.695 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:01:46.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.49s (1.89min) +2026-06-20 05:01:47.078 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:03:44.145 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:10:42.583 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 118, train_steps: 4720 +2026-06-20 05:10:43.296 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 05:10:43.469 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:10:43.551 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:11:48.127 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:11:48.130 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:13:46.957 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 05:13:47.787 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:15:41.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.26s (1.89min) +2026-06-20 05:15:42.000 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:17:38.996 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:24:37.899 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 119, train_steps: 4760 +2026-06-20 05:24:38.622 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 05:24:38.743 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:24:38.787 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:25:43.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:25:43.373 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:27:40.119 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 05:27:41.094 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:29:34.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-20 05:29:35.559 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:31:33.107 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:38:32.157 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 120, train_steps: 4800 +2026-06-20 05:39:09.156 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=120 (missing=0, unexpected=0) +2026-06-20 05:39:10.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 05:39:11.016 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:39:11.062 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:40:18.047 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:40:18.051 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:42:12.566 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 05:42:13.553 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:44:08.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.96s (1.92min) +2026-06-20 05:44:09.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:46:10.088 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:53:12.071 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 121, train_steps: 4840 +2026-06-20 05:53:12.775 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 05:53:12.870 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:53:12.912 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:54:18.600 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:54:18.600 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:56:13.156 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 05:56:14.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:58:09.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.97s (1.92min) +2026-06-20 05:58:10.076 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:00:10.577 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:07:11.964 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 122, train_steps: 4880 +2026-06-20 06:07:12.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 06:07:12.811 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:07:12.880 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:08:18.164 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:08:18.164 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:10:13.195 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 06:10:14.183 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:12:08.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.25s (1.90min) +2026-06-20 06:12:09.327 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:14:10.583 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:21:11.885 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 123, train_steps: 4920 +2026-06-20 06:21:12.560 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 06:21:12.678 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:21:12.717 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:22:17.751 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:22:17.752 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:24:12.071 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 06:24:13.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:26:08.788 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.47s (1.92min) +2026-06-20 06:26:09.511 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:28:09.836 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:35:11.875 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 124, train_steps: 4960 +2026-06-20 06:35:12.538 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 06:35:12.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:35:12.681 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:36:17.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:36:17.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:38:12.129 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 06:38:13.089 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:40:08.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.44s (1.92min) +2026-06-20 06:40:09.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:42:09.775 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:49:11.656 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 125, train_steps: 5000 +2026-06-20 06:50:15.220 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 06:50:15.425 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:50:15.556 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:51:20.310 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:51:20.311 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:53:15.926 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 06:53:16.800 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:55:11.654 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.68s (1.91min) +2026-06-20 06:55:12.372 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:57:11.764 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:04:11.919 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 126, train_steps: 5040 +2026-06-20 07:04:12.631 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 07:04:12.805 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:04:12.886 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:05:18.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:05:18.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:07:13.988 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 07:07:15.038 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:09:10.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.08s (1.92min) +2026-06-20 07:09:11.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:11:11.166 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:18:11.927 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 127, train_steps: 5080 +2026-06-20 07:18:12.629 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 07:18:12.797 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:18:12.870 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:19:17.779 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:19:17.779 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:21:13.050 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 07:21:13.874 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:23:09.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.50s (1.93min) +2026-06-20 07:23:10.327 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:25:10.159 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:32:11.180 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 128, train_steps: 5120 +2026-06-20 07:32:11.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 07:32:12.025 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:32:12.063 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:33:17.027 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:33:17.027 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:35:13.016 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 07:35:13.839 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:37:08.641 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.56s (1.91min) +2026-06-20 07:37:09.373 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:39:09.932 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:46:11.989 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 129, train_steps: 5160 +2026-06-20 07:46:12.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 07:46:12.891 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:46:13.003 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:47:18.538 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:47:18.538 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:49:13.070 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 07:49:13.889 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:51:09.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-20 07:51:09.758 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:53:10.305 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:00:11.780 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 130, train_steps: 5200 +2026-06-20 08:00:49.274 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=130 (missing=0, unexpected=0) +2026-06-20 08:00:51.244 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 08:00:51.352 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:00:51.392 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:01:58.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:01:58.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:03:52.509 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 08:03:53.310 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:05:47.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.72s (1.90min) +2026-06-20 08:05:48.018 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:07:44.846 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:14:45.318 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 131, train_steps: 5240 +2026-06-20 08:14:46.026 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 08:14:46.223 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:14:46.323 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:15:50.956 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:15:50.959 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:17:46.989 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 08:17:47.817 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:19:41.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.38s (1.89min) +2026-06-20 08:19:42.138 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:21:39.467 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:28:37.825 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 132, train_steps: 5280 +2026-06-20 08:28:38.520 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 08:28:38.698 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:28:38.791 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:29:43.350 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:29:43.353 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:31:40.849 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 08:31:41.693 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:33:35.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.44s (1.89min) +2026-06-20 08:33:36.119 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:35:35.264 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:42:34.330 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 133, train_steps: 5320 +2026-06-20 08:42:35.016 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 08:42:35.113 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:42:35.155 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:43:40.015 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:43:40.016 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:45:35.887 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 08:45:36.699 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:47:32.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.77s (1.93min) +2026-06-20 08:47:33.449 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:49:30.693 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:56:29.424 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 134, train_steps: 5360 +2026-06-20 08:56:30.115 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 08:56:30.275 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:56:30.334 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:57:34.873 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:57:34.877 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:59:32.040 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 08:59:32.851 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:01:26.319 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 09:01:27.048 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:03:23.806 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:10:22.982 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 135, train_steps: 5400 +2026-06-20 09:10:23.673 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 09:10:23.809 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:10:23.850 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:11:28.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:11:28.663 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:13:23.932 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 09:13:24.750 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:15:18.588 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.61s (1.89min) +2026-06-20 09:15:19.310 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:17:17.100 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:24:16.798 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 136, train_steps: 5440 +2026-06-20 09:24:17.531 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 09:24:17.653 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:24:17.694 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:25:22.405 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:25:22.408 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:27:18.656 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 09:27:19.482 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:29:12.958 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.15s (1.89min) +2026-06-20 09:29:13.686 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:31:13.178 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:38:13.224 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 137, train_steps: 5480 +2026-06-20 09:38:13.936 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 09:38:14.109 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:38:14.183 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:39:19.160 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:39:19.161 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:41:14.371 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 09:41:15.532 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:43:09.522 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-20 09:43:10.251 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:45:10.505 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:52:12.161 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 138, train_steps: 5520 +2026-06-20 09:52:12.866 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 09:52:13.002 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:52:13.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:53:18.569 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:53:18.569 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:55:13.383 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 09:55:14.462 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:57:10.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.46s (1.92min) +2026-06-20 09:57:10.921 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:59:11.152 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:06:12.667 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 139, train_steps: 5560 +2026-06-20 10:06:13.363 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 10:06:13.481 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:06:13.523 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:07:18.693 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:07:18.693 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:09:13.264 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 10:09:14.071 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:11:08.675 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.20s (1.90min) +2026-06-20 10:11:09.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:13:09.782 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:20:12.686 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 140, train_steps: 5600 +2026-06-20 10:20:49.827 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=140 (missing=0, unexpected=0) +2026-06-20 10:20:51.812 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 10:20:51.913 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:20:51.949 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:21:58.291 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:21:58.294 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:23:53.040 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 10:23:53.850 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:25:47.600 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-20 10:25:48.313 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:27:45.366 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:34:44.008 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 141, train_steps: 5640 +2026-06-20 10:34:44.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 10:34:44.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:34:44.837 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:35:49.306 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:35:49.307 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:37:46.846 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 10:37:47.746 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:39:41.415 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.51s (1.89min) +2026-06-20 10:39:42.148 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:41:39.463 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:48:38.462 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 142, train_steps: 5680 +2026-06-20 10:48:39.157 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 10:48:39.307 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:48:39.361 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:49:43.998 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:49:44.001 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:51:43.153 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 10:51:43.974 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:53:37.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-20 10:53:37.850 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:55:37.995 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:02:37.920 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 143, train_steps: 5720 +2026-06-20 11:02:38.630 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 11:02:38.759 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:02:38.791 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:03:43.553 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:03:43.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:05:42.169 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 11:05:43.142 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:07:36.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 11:07:37.312 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:09:36.844 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:16:37.160 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 144, train_steps: 5760 +2026-06-20 11:16:37.798 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 11:16:37.933 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:16:37.955 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:17:42.487 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:17:42.490 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:19:38.707 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 11:19:39.637 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:21:33.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.47s (1.89min) +2026-06-20 11:21:34.087 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:23:30.890 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:30:30.410 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 145, train_steps: 5800 +2026-06-20 11:30:31.096 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 11:30:31.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:30:31.261 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:31:35.714 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:31:35.718 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:33:33.207 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 11:33:34.196 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:35:28.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.89s (1.90min) +2026-06-20 11:35:28.900 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:37:25.683 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:44:23.912 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 146, train_steps: 5840 +2026-06-20 11:44:24.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 11:44:24.773 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:44:24.833 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:45:29.272 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:45:29.274 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:47:23.247 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 11:47:24.068 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:49:19.133 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-20 11:49:19.856 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:51:17.043 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:58:15.932 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 147, train_steps: 5880 +2026-06-20 11:58:16.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 11:58:16.714 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:58:16.759 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:59:21.704 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:59:21.788 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:01:16.232 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 12:01:17.247 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:03:12.586 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.29s (1.92min) +2026-06-20 12:03:13.293 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:05:12.907 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:12:12.704 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 148, train_steps: 5920 +2026-06-20 12:12:13.406 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 12:12:13.572 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:12:13.623 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:13:19.057 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:13:19.057 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:15:13.144 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 12:15:14.004 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:17:09.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.69s (1.91min) +2026-06-20 12:17:09.788 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:19:10.589 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:26:11.456 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 149, train_steps: 5960 +2026-06-20 12:26:12.134 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 12:26:12.320 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:26:12.391 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:27:17.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:27:17.979 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:29:13.029 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 12:29:13.851 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:31:08.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.72s (1.91min) +2026-06-20 12:31:09.532 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:33:09.829 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:40:10.986 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 150, train_steps: 6000 +2026-06-20 12:41:53.251 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=150 (missing=0, unexpected=0) +2026-06-20 12:41:56.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 12:41:56.157 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:41:56.212 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:43:02.247 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:43:02.247 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:44:59.744 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 12:45:00.581 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:46:55.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.70s (1.91min) +2026-06-20 12:46:56.236 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:48:52.990 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:55:51.601 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 151, train_steps: 6040 +2026-06-20 12:55:52.306 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 12:55:52.457 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:55:52.494 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:56:56.902 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:56:56.905 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:58:50.721 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 12:58:51.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:00:46.189 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.41s (1.91min) +2026-06-20 13:00:46.908 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:02:43.655 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:09:41.917 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 152, train_steps: 6080 +2026-06-20 13:09:42.597 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 13:09:42.754 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:09:42.810 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:10:47.551 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:10:47.552 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:12:43.910 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 13:12:44.904 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:14:38.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-20 13:14:39.492 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:16:38.553 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:23:37.221 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 153, train_steps: 6120 +2026-06-20 13:23:38.091 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 13:23:38.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:23:38.251 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:24:43.463 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:24:43.466 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:26:40.127 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 13:26:41.305 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:28:34.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.39s (1.89min) +2026-06-20 13:28:35.473 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:30:32.908 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:37:31.695 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 154, train_steps: 6160 +2026-06-20 13:37:32.394 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 13:37:32.563 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:37:32.623 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:38:37.034 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:38:37.038 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:40:33.788 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 13:40:34.625 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:42:29.329 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.47s (1.91min) +2026-06-20 13:42:30.062 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:44:27.160 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:51:24.980 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 155, train_steps: 6200 +2026-06-20 13:51:25.701 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 13:51:25.892 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:51:25.987 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:52:30.496 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:52:30.499 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:54:24.447 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 13:54:25.443 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:56:19.129 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.63s (1.89min) +2026-06-20 13:56:19.859 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:58:16.752 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:05:15.351 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 156, train_steps: 6240 +2026-06-20 14:05:16.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 14:05:16.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:05:16.189 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:06:21.093 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:06:21.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:08:15.931 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 14:08:16.927 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:10:11.722 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.74s (1.91min) +2026-06-20 14:10:12.457 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:12:11.106 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:19:12.247 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 157, train_steps: 6280 +2026-06-20 14:19:12.925 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 14:19:13.115 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:19:13.214 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:20:18.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:20:18.391 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:22:12.398 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 14:22:13.502 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:24:08.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.22s (1.92min) +2026-06-20 14:24:09.689 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:26:09.908 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:33:10.839 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 158, train_steps: 6320 +2026-06-20 14:33:11.517 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 14:33:11.673 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:33:11.729 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:34:16.922 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:34:16.922 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:36:10.935 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 14:36:12.135 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:38:07.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.88s (1.91min) +2026-06-20 14:38:07.785 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:40:05.260 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:47:04.576 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 159, train_steps: 6360 +2026-06-20 14:47:05.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 14:47:05.411 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:47:05.446 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:48:09.995 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:48:09.999 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:50:04.048 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 14:50:05.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:51:59.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.86s (1.90min) +2026-06-20 14:51:59.929 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:53:57.324 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:00:59.186 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 160, train_steps: 6400 +2026-06-20 15:01:37.183 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=160 (missing=0, unexpected=0) +2026-06-20 15:01:39.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 15:01:39.475 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:01:39.572 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:02:45.311 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:02:45.314 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:04:42.879 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 15:04:43.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:06:37.408 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.48s (1.89min) +2026-06-20 15:06:38.136 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:08:38.302 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:15:37.521 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 161, train_steps: 6440 +2026-06-20 15:15:38.205 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 15:15:38.320 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:15:38.359 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:16:42.852 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:16:42.852 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:18:40.375 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 15:18:41.381 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:20:35.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-20 15:20:36.332 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:22:36.151 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:29:36.893 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 162, train_steps: 6480 +2026-06-20 15:29:37.569 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 15:29:37.692 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:29:37.726 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:30:42.281 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:30:42.282 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:32:37.444 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 15:32:38.272 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:34:33.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.48s (1.91min) +2026-06-20 15:34:33.827 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:36:31.318 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:43:31.485 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 163, train_steps: 6520 +2026-06-20 15:43:32.170 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 15:43:32.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:43:32.310 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:44:36.992 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:44:36.995 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:46:33.676 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 15:46:34.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:48:29.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.47s (1.91min) +2026-06-20 15:48:29.934 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:50:26.541 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:57:25.530 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 164, train_steps: 6560 +2026-06-20 15:57:26.227 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 15:57:26.324 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:57:26.368 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:58:30.757 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:58:30.760 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:00:26.401 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 16:00:27.286 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:02:21.064 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.61s (1.89min) +2026-06-20 16:02:21.793 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:04:18.345 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:11:18.368 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 165, train_steps: 6600 +2026-06-20 16:11:19.087 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 16:11:19.265 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:11:19.333 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:12:24.226 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:12:24.226 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:14:19.105 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 16:14:20.082 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:16:14.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.75s (1.90min) +2026-06-20 16:16:14.870 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:18:14.127 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:25:13.923 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 166, train_steps: 6640 +2026-06-20 16:25:14.621 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 16:25:14.775 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:25:14.833 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:26:19.839 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:26:19.842 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:28:13.831 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 16:28:14.767 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:30:08.844 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.78s (1.90min) +2026-06-20 16:30:09.574 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:32:10.308 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:39:12.158 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 167, train_steps: 6680 +2026-06-20 16:39:12.849 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 16:39:13.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:39:13.110 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:40:18.600 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:40:18.600 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:42:12.094 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 16:42:13.162 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:44:08.745 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.36s (1.92min) +2026-06-20 16:44:09.477 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:46:10.021 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:53:11.391 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 168, train_steps: 6720 +2026-06-20 16:53:12.088 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 16:53:12.187 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:53:12.234 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:54:17.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:54:17.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:56:11.767 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 16:56:12.808 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:58:07.736 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-20 16:58:08.464 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:00:07.580 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:07:09.174 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 169, train_steps: 6760 +2026-06-20 17:07:09.877 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 17:07:09.990 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:07:10.054 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:08:14.960 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:08:14.963 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:10:11.903 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 17:10:12.729 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:12:08.254 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.20s (1.92min) +2026-06-20 17:12:08.969 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:14:09.321 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:21:11.044 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 170, train_steps: 6800 +2026-06-20 17:21:47.216 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=170 (missing=0, unexpected=0) +2026-06-20 17:21:49.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 17:21:49.397 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:21:49.509 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:22:54.899 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:22:54.902 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:24:49.925 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 17:24:50.749 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:26:44.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.05s (1.88min) +2026-06-20 17:26:44.742 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:28:41.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:35:39.735 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 53, sp_rank 0: update step: 171, train_steps: 6840 +2026-06-20 17:35:40.432 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 53, dp_rank 26: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=0 +2026-06-20 17:35:40.635 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:35:40.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:36:45.642 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 53 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:36:45.646 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 53 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:38:42.537 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 53: Using local metric names without global sync: ['avg'] +2026-06-20 17:38:43.556 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 53: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:40:36.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 53: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.89s (1.88min) +2026-06-20 17:40:37.378 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 53, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-29 01:28:46.726 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 01:28:46.726 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 01:51:26.586 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 01:51:26.587 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 02:20:50.861 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 02:20:50.863 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 02:46:33.747 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 02:46:33.748 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:07:22.196 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:07:22.199 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:32:26.693 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:32:26.696 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:59:04.288 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:59:04.290 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:26:25.042 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:26:25.043 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:52:07.360 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:52:07.363 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 05:15:26.706 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 05:15:26.707 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 05:38:50.112 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 05:38:50.113 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:02:40.577 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:02:40.579 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:25:38.534 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:25:38.537 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:49:05.726 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:49:05.726 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 07:12:19.711 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 07:12:19.714 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 07:36:15.346 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 07:36:15.347 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:07:30.448 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:07:30.449 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:30:05.998 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:30:05.999 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:59:24.252 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:59:24.254 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:24:19.439 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:24:19.439 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:47:48.021 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:47:48.022 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 10:10:21.149 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 10:10:21.151 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 10:36:09.756 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 10:36:09.756 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:03:38.715 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:03:38.717 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:27:30.671 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:27:30.672 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:55:23.057 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:55:23.058 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 12:20:25.684 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 12:20:25.684 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 12:48:26.968 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 12:48:26.968 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 13:12:55.512 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 13:12:55.515 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 13:35:41.828 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 13:35:41.828 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:05:35.825 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:05:35.826 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:26:55.160 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:26:55.161 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:49:38.468 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:49:38.468 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 15:20:22.750 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 15:20:22.751 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 15:47:25.748 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 15:47:25.749 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 16:15:41.348 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 16:15:41.349 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 16:40:28.895 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 16:40:28.895 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:09:35.766 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:09:35.767 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:32:40.646 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:32:40.648 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:54:06.055 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:54:06.056 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 18:16:51.658 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 18:16:51.660 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 18:39:28.473 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 18:39:28.474 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_54.log b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_54.log new file mode 100644 index 0000000000000000000000000000000000000000..781c68d6f6389393288648da3ea6e6cc4b66170b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_54.log @@ -0,0 +1,3241 @@ +2026-06-19 01:25:11.789 | INFO | __main__:__init__:933 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-19 01:25:12.441 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-19 01:25:12.443 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-19 01:30:17.294 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 01:30:31.663 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 01:30:31.664 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:30:31.665 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:31:52.733 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_13/index_6493_seed_27064_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 01:31:57.500 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 01:32:14.890 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:32:14.899 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:32:14.901 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:33:22.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_13/index_6493_seed_27065_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 01:33:26.678 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 01:33:45.094 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:33:45.514 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 01:34:00.989 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 01:35:56.997 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.94s (1.93min) +2026-06-19 01:35:57.694 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 01:37:54.186 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 01:37:57.227 | INFO | hyvideo.utils.grpo_utils:train_one_step:3127 - Rank 54, sp_rank 0: init adaptive KL controller (init=1.000e-07, min=1.000e-07, max=1.000e-05) +2026-06-19 01:45:11.949 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 1, train_steps: 40 +2026-06-19 01:45:12.575 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 01:45:12.735 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 01:45:12.792 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:45:12.833 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:46:19.762 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000001/group_13/index_587_seed_16701_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 01:46:22.946 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 01:46:39.943 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:46:39.951 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:46:39.951 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:47:47.275 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000001/group_13/index_587_seed_16702_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 01:47:50.495 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 01:48:07.638 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:48:10.944 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 01:48:12.090 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 01:50:07.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.70s (1.93min) +2026-06-19 01:50:08.466 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 01:52:06.957 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 01:59:06.766 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 2, train_steps: 80 +2026-06-19 01:59:07.381 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 01:59:07.512 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 01:59:07.535 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:59:07.600 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:00:13.021 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000002/group_13/index_3343_seed_16091_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 02:00:16.195 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:00:33.229 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:00:33.238 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:00:33.238 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:01:46.207 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000002/group_13/index_3343_seed_16092_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 02:01:51.175 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:02:09.280 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:02:10.528 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:02:11.842 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:04:06.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.88s (1.91min) +2026-06-19 02:04:07.374 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:06:03.755 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:13:04.859 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 3, train_steps: 120 +2026-06-19 02:13:05.465 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 02:13:05.571 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:13:05.580 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:13:05.590 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:14:10.994 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_13/index_1654_seed_2811_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 02:14:16.010 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:14:33.892 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:14:33.901 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:14:33.901 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:15:40.412 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_13/index_1654_seed_2812_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 02:15:44.533 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:16:02.101 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:16:04.655 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:16:05.796 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:17:59.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.77s (1.90min) +2026-06-19 02:18:00.448 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:19:59.352 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:26:59.887 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 4, train_steps: 160 +2026-06-19 02:27:00.547 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 02:27:00.748 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:27:00.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:27:00.945 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:28:06.760 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000004/group_13/index_9378_seed_21400_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 02:28:09.947 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:28:28.709 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:28:28.718 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:28:28.718 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:29:35.160 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000004/group_13/index_9378_seed_21401_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 02:29:40.159 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:29:58.442 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:30:00.316 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:30:01.318 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:31:55.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.62s (1.91min) +2026-06-19 02:31:56.653 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:33:53.226 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:40:53.228 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 5, train_steps: 200 +2026-06-19 02:40:53.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 02:40:54.065 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:40:54.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:40:54.262 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:42:00.189 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_13/index_2163_seed_5758_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 02:42:04.193 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:42:21.437 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:42:21.446 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:42:21.446 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:43:28.524 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_13/index_2163_seed_5759_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 02:43:31.744 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:43:49.047 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:43:51.812 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:43:52.681 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:45:46.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.50s (1.89min) +2026-06-19 02:45:46.994 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:47:43.354 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:54:40.665 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 6, train_steps: 240 +2026-06-19 02:54:41.360 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 02:54:41.472 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:54:41.504 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:54:41.602 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:55:48.064 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_13/index_2074_seed_2223_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 02:55:51.341 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:56:08.608 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:56:08.618 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:56:08.618 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:57:19.536 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_13/index_2074_seed_2224_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 02:57:24.549 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:57:42.190 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:57:42.949 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:57:43.946 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:59:37.353 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.38s (1.89min) +2026-06-19 02:59:38.043 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:01:37.732 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:08:37.484 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 7, train_steps: 280 +2026-06-19 03:08:38.147 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 03:08:38.273 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:08:38.310 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:08:38.399 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:09:43.992 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_13/index_6829_seed_25619_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 03:09:48.725 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:10:06.151 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:10:06.160 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:10:06.162 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:11:14.709 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_13/index_6829_seed_25620_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 03:11:19.505 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:11:37.179 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:11:38.337 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:11:39.460 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:13:34.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.68s (1.91min) +2026-06-19 03:13:34.976 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:15:32.044 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:22:31.154 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 8, train_steps: 320 +2026-06-19 03:22:31.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 03:22:32.019 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:22:32.139 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:22:32.233 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:23:38.523 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000008/group_13/index_8866_seed_20342_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 03:23:43.342 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:24:00.954 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:24:00.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:24:00.963 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:25:09.051 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000008/group_13/index_8866_seed_20343_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 03:25:14.029 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:25:31.906 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:25:33.100 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:25:34.063 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:27:28.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-19 03:27:29.431 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:29:26.906 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:36:25.652 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 9, train_steps: 360 +2026-06-19 03:36:26.346 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 03:36:26.520 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:36:26.618 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:36:26.714 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:37:32.558 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_13/index_7625_seed_31624_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 03:37:36.459 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:37:53.978 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:37:53.987 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:37:53.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:39:02.022 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_13/index_7625_seed_31625_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 03:39:06.966 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:39:25.051 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:39:26.010 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:39:26.996 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:41:20.400 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.36s (1.89min) +2026-06-19 03:41:21.105 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:43:17.561 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:50:16.937 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 10, train_steps: 400 +2026-06-19 03:50:53.265 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=10 (missing=0, unexpected=0) +2026-06-19 03:50:56.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 03:50:56.352 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:50:56.459 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:50:56.536 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:52:03.392 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_13/index_3387_seed_27192_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 03:52:08.179 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:52:24.723 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:52:24.732 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:52:24.732 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:53:32.499 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_13/index_3387_seed_27193_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 03:53:36.420 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:53:52.737 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:53:57.828 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:53:58.756 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:55:52.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.65s (1.89min) +2026-06-19 03:55:53.209 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:57:49.766 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:04:48.386 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 11, train_steps: 440 +2026-06-19 04:04:49.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 04:04:49.253 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:04:49.340 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:04:49.440 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:05:55.066 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_13/index_1969_seed_26017_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 04:05:58.372 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:06:16.114 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:06:16.122 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:06:16.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:07:24.555 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_13/index_1969_seed_26018_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 04:07:27.845 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:07:45.640 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:07:48.852 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:07:49.732 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:09:43.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.25s (1.89min) +2026-06-19 04:09:43.797 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:11:40.270 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:18:37.301 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 12, train_steps: 480 +2026-06-19 04:18:37.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 04:18:38.091 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:18:38.130 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:18:38.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:19:44.288 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_13/index_2917_seed_14637_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 04:19:47.554 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:20:05.417 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:20:05.427 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:20:05.427 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:21:12.907 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_13/index_2917_seed_14638_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 04:21:17.113 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:21:34.967 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:21:37.180 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:21:38.107 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:23:33.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.19s (1.92min) +2026-06-19 04:23:34.070 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:25:31.245 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:32:29.827 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 13, train_steps: 520 +2026-06-19 04:32:30.513 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 04:32:30.678 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:32:30.756 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:32:30.868 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:33:36.651 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_13/index_7170_seed_29638_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 04:33:41.727 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:33:59.618 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:33:59.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:33:59.628 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:35:06.127 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_13/index_7170_seed_29639_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 04:35:11.263 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:35:30.256 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:35:30.312 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:35:31.196 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:37:24.102 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.77s (1.88min) +2026-06-19 04:37:24.792 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:39:21.078 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:46:18.592 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 14, train_steps: 560 +2026-06-19 04:46:19.247 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 04:46:19.435 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:46:19.542 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:46:19.662 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:47:25.651 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_13/index_2349_seed_30174_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 04:47:28.803 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:47:46.314 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:47:46.323 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:47:46.323 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:48:54.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_13/index_2349_seed_30175_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 04:48:57.467 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:49:15.121 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:49:17.445 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:49:18.367 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:51:12.963 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.49s (1.91min) +2026-06-19 04:51:13.650 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:53:12.996 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:00:11.841 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 15, train_steps: 600 +2026-06-19 05:00:12.525 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 05:00:12.658 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:00:12.691 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:00:12.782 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:01:19.765 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000015/group_13/index_7266_seed_2780_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 05:01:23.771 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:01:41.748 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:01:41.754 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:01:41.754 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:02:47.657 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000015/group_13/index_7266_seed_2781_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 05:02:52.477 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:03:10.516 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:03:11.349 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:03:12.689 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:05:07.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.64s (1.91min) +2026-06-19 05:05:08.060 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:07:05.977 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:14:04.895 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 16, train_steps: 640 +2026-06-19 05:14:05.573 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 05:14:05.686 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:14:05.733 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:14:05.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:15:11.714 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_13/index_5589_seed_23270_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 05:15:16.780 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:15:34.751 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:15:34.759 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:15:34.759 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:16:42.985 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_13/index_5589_seed_23271_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 05:16:47.860 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:17:05.972 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:17:06.988 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:17:07.960 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:19:03.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.02s (1.92min) +2026-06-19 05:19:03.742 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:21:01.963 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:28:03.330 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 17, train_steps: 680 +2026-06-19 05:28:04.007 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 05:28:04.191 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:28:04.288 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:28:04.389 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:29:09.985 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_13/index_173_seed_694_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 05:29:15.100 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:29:34.278 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:29:34.288 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:29:34.289 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:30:39.611 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_13/index_173_seed_695_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 05:30:44.694 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:31:03.916 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:31:03.926 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:31:05.027 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:32:58.688 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.48s (1.89min) +2026-06-19 05:32:59.387 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:34:56.267 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:41:58.104 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 18, train_steps: 720 +2026-06-19 05:41:58.777 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 05:41:58.998 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:41:59.120 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:41:59.221 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:43:04.828 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_13/index_1580_seed_6548_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 05:43:09.061 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:43:26.540 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:43:26.548 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:43:26.549 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:44:35.489 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_13/index_1580_seed_6549_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 05:44:40.561 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:44:58.559 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:45:00.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:45:01.209 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:46:55.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.88s (1.90min) +2026-06-19 05:46:55.797 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:48:52.332 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:55:50.498 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 19, train_steps: 760 +2026-06-19 05:55:51.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 05:55:51.379 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:55:51.496 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:55:51.586 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:56:57.122 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_13/index_8612_seed_31124_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 05:57:01.880 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:57:19.820 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:57:19.829 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:57:19.829 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:58:25.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_13/index_8612_seed_31125_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 05:58:29.424 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:58:47.185 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:58:49.068 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:58:49.986 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:00:43.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.10s (1.89min) +2026-06-19 06:00:43.876 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:02:40.343 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:09:39.763 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 20, train_steps: 800 +2026-06-19 06:10:15.939 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=20 (missing=0, unexpected=0) +2026-06-19 06:10:18.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 06:10:18.963 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:10:19.051 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:10:19.098 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:11:26.459 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_13/index_4883_seed_5498_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 06:11:29.637 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:11:45.960 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:11:45.970 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:11:45.970 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:12:54.649 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_13/index_4883_seed_5499_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 06:12:59.669 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:13:16.719 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:13:19.050 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:13:20.167 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:15:13.224 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.90s (1.88min) +2026-06-19 06:15:13.910 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:17:12.689 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:24:12.564 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 21, train_steps: 840 +2026-06-19 06:24:13.271 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 06:24:13.466 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:24:13.563 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:24:13.655 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:25:19.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_13/index_3040_seed_28885_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 06:25:24.792 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:25:42.826 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:25:42.835 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:25:42.835 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:26:47.524 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_13/index_3040_seed_28886_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 06:26:51.786 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:27:09.462 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:27:11.261 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:27:12.441 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:29:06.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.51s (1.91min) +2026-06-19 06:29:07.668 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:31:04.580 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:38:04.281 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 22, train_steps: 880 +2026-06-19 06:38:04.947 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 06:38:05.059 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:38:05.100 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:38:05.169 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:39:11.053 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_13/index_7742_seed_32496_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 06:39:16.164 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:39:33.873 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:39:33.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:39:33.885 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:40:39.831 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_13/index_7742_seed_32497_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 06:40:44.771 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:41:02.741 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:41:03.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:41:04.299 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:42:58.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.99s (1.90min) +2026-06-19 06:42:59.165 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:44:56.505 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:51:58.182 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 23, train_steps: 920 +2026-06-19 06:51:58.843 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 06:51:59.030 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:51:59.161 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:51:59.242 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:53:04.876 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000023/group_13/index_1830_seed_1574_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 06:53:09.712 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:53:27.478 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:53:27.488 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:53:27.489 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:54:35.216 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000023/group_13/index_1830_seed_1575_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 06:54:40.261 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:54:58.172 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:54:59.789 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:55:00.733 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:56:54.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.11s (1.90min) +2026-06-19 06:56:55.623 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:58:52.322 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:05:51.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 24, train_steps: 960 +2026-06-19 07:05:52.004 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 07:05:52.159 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:05:52.216 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:05:52.347 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:06:58.290 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_13/index_3619_seed_3548_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 07:07:03.415 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:07:21.364 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:07:21.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:07:21.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:08:27.528 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_13/index_3619_seed_3549_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 07:08:32.696 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:08:50.635 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:08:52.217 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:08:53.195 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:10:46.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-19 07:10:46.871 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:12:44.156 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:19:41.978 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 25, train_steps: 1000 +2026-06-19 07:20:52.105 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 07:20:52.288 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:20:52.398 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:20:52.497 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:21:58.468 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_13/index_9319_seed_17721_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 07:22:02.401 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:22:20.553 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:22:20.562 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:22:20.562 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:23:28.892 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_13/index_9319_seed_17722_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 07:23:34.136 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:23:52.610 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:23:52.666 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:23:53.694 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:25:46.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.81s (1.88min) +2026-06-19 07:25:47.206 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:27:44.383 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:34:43.683 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 26, train_steps: 1040 +2026-06-19 07:34:44.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 07:34:44.477 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:34:44.513 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:34:44.582 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:35:50.276 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_13/index_3139_seed_4508_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 07:35:55.205 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:36:13.168 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:36:13.175 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:36:13.175 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:37:20.737 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_13/index_3139_seed_4509_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 07:37:24.923 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:37:42.947 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:37:45.443 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:37:46.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:39:39.638 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.14s (1.89min) +2026-06-19 07:39:40.349 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:41:39.074 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:48:39.682 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 27, train_steps: 1080 +2026-06-19 07:48:40.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 07:48:40.516 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:48:40.609 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:48:40.709 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:49:46.762 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_13/index_6674_seed_22581_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 07:49:51.838 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:50:09.600 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:50:09.608 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:50:09.608 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:51:19.438 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_13/index_6674_seed_22582_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 07:51:23.589 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:51:41.619 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:51:43.285 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:51:44.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:53:38.298 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.99s (1.90min) +2026-06-19 07:53:39.007 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:55:38.423 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:02:38.183 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 28, train_steps: 1120 +2026-06-19 08:02:38.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 08:02:38.977 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:02:39.015 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:02:39.063 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:03:45.301 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_13/index_9980_seed_29622_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 08:03:48.418 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:04:06.294 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:04:06.303 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:04:06.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:05:17.454 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_13/index_9980_seed_29623_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 08:05:20.637 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:05:38.782 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:05:41.901 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:05:42.986 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:07:37.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.24s (1.90min) +2026-06-19 08:07:38.091 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:09:37.936 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:16:37.861 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 29, train_steps: 1160 +2026-06-19 08:16:38.691 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 08:16:38.826 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:16:38.859 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:16:39.026 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:17:44.649 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_13/index_7481_seed_1688_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 08:17:47.765 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:18:05.809 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:18:05.818 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:18:05.818 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:19:18.383 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_13/index_7481_seed_1689_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 08:19:23.588 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:19:42.483 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:19:42.881 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:19:43.877 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:21:37.264 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 08:21:37.961 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:23:37.408 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:30:39.550 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 30, train_steps: 1200 +2026-06-19 08:31:15.513 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=30 (missing=0, unexpected=0) +2026-06-19 08:31:18.259 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 08:31:18.468 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:31:18.563 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:31:18.659 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:32:26.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000030/group_13/index_876_seed_5368_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 08:32:30.702 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:32:48.568 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:32:48.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:32:48.578 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:33:55.850 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000030/group_13/index_876_seed_5369_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 08:33:59.163 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:34:17.164 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:34:19.980 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:34:21.071 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:36:14.510 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-19 08:36:15.227 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:38:14.176 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:45:14.284 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 31, train_steps: 1240 +2026-06-19 08:45:14.994 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 08:45:15.191 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:45:15.323 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:45:15.417 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:46:21.353 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_13/index_6_seed_25487_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 08:46:24.408 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:46:42.207 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:46:42.216 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:46:42.216 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:47:50.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_13/index_6_seed_25488_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 08:47:56.099 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:48:14.519 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:48:15.288 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:48:16.460 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:50:10.323 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.84s (1.90min) +2026-06-19 08:50:11.024 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:52:11.151 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:59:13.098 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 32, train_steps: 1280 +2026-06-19 08:59:13.775 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 08:59:13.910 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:59:13.949 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:59:14.038 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:00:20.583 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000032/group_13/index_874_seed_26115_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 09:00:25.535 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:00:43.669 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:00:43.677 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:00:43.678 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:01:48.629 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000032/group_13/index_874_seed_26116_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 09:01:53.482 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:02:12.639 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:02:13.436 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:02:14.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:04:08.713 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.91s (1.90min) +2026-06-19 09:04:09.421 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:06:10.410 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:13:13.280 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 33, train_steps: 1320 +2026-06-19 09:13:13.953 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 09:13:14.142 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:13:14.253 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:13:14.369 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:14:20.545 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000033/group_13/index_3436_seed_10814_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 09:14:25.308 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:14:43.394 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:14:43.404 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:14:43.404 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:15:49.783 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000033/group_13/index_3436_seed_10815_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 09:15:54.579 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:16:13.234 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:16:14.822 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:16:15.858 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:18:11.066 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.06s (1.92min) +2026-06-19 09:18:11.786 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:20:11.628 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:27:13.735 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 34, train_steps: 1360 +2026-06-19 09:27:14.445 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 09:27:14.634 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:27:14.731 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:27:14.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:28:21.185 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_13/index_5339_seed_9496_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 09:28:25.568 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:28:43.841 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:28:43.850 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:28:43.850 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:29:49.902 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_13/index_5339_seed_9497_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 09:29:54.724 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:30:13.170 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:30:13.695 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:30:14.961 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:32:08.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.81s (1.90min) +2026-06-19 09:32:09.592 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:34:09.780 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:41:12.408 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 35, train_steps: 1400 +2026-06-19 09:41:13.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 09:41:13.228 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:41:13.320 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:41:13.430 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:42:20.021 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000035/group_13/index_4255_seed_27252_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 09:42:24.240 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:42:42.193 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:42:42.203 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:42:42.203 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:43:48.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000035/group_13/index_4255_seed_27253_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 09:43:53.840 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:44:13.031 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:44:13.080 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:44:14.090 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:46:10.686 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 116.56s (1.94min) +2026-06-19 09:46:11.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:48:10.898 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:55:11.961 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 36, train_steps: 1440 +2026-06-19 09:55:12.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 09:55:12.799 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:55:12.856 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:55:12.993 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:56:19.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_13/index_2741_seed_15282_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 09:56:24.654 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:56:42.472 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:56:42.481 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:56:42.482 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:57:47.947 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_13/index_2741_seed_15283_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 09:57:52.176 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:58:12.089 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:58:12.646 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:58:13.792 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:00:08.966 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-19 10:00:09.672 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:02:10.010 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:09:11.714 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 37, train_steps: 1480 +2026-06-19 10:09:12.398 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 10:09:12.619 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:09:12.756 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:09:12.839 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:10:19.308 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_13/index_2580_seed_12045_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 10:10:22.376 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:10:41.688 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:10:41.696 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:10:41.697 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:11:47.650 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_13/index_2580_seed_12046_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 10:11:52.793 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:12:13.034 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:12:13.042 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:12:14.041 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:14:09.029 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.94s (1.92min) +2026-06-19 10:14:09.738 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:16:10.369 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:23:12.353 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 38, train_steps: 1520 +2026-06-19 10:23:13.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 10:23:13.167 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:23:13.206 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:23:13.279 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:24:20.275 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000038/group_13/index_7824_seed_3053_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 10:24:24.238 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:24:42.196 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:24:42.206 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:24:42.206 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:25:48.630 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000038/group_13/index_7824_seed_3054_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 10:25:51.745 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:26:09.686 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:26:12.667 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:26:13.803 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:28:09.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.16s (1.92min) +2026-06-19 10:28:09.956 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:30:10.744 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:37:12.065 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 39, train_steps: 1560 +2026-06-19 10:37:12.646 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 10:37:12.798 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:37:12.849 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:37:12.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:38:19.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000039/group_13/index_5221_seed_2867_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 10:38:24.944 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:38:43.327 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:38:43.336 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:38:43.337 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:41:17.621 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000039/group_13/index_5221_seed_2868_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 10:41:21.661 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:41:40.287 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:41:42.526 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:41:43.618 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:43:36.471 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.70s (1.88min) +2026-06-19 10:43:37.180 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:45:37.178 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:52:37.872 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 40, train_steps: 1600 +2026-06-19 10:53:13.837 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=40 (missing=0, unexpected=0) +2026-06-19 10:53:15.773 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 10:53:15.950 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:53:16.042 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:53:16.115 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:54:23.540 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_13/index_6865_seed_8412_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 10:54:26.587 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:54:45.727 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:54:45.735 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:54:45.735 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:55:52.327 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_13/index_6865_seed_8413_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 10:55:55.646 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:56:15.476 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:56:16.410 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:56:17.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:58:11.717 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.27s (1.90min) +2026-06-19 10:58:12.416 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:00:11.327 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:07:12.586 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 41, train_steps: 1640 +2026-06-19 11:07:13.291 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 11:07:13.465 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:07:13.548 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:07:13.680 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:08:20.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_13/index_5658_seed_4763_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 11:08:23.301 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:08:41.077 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:08:41.086 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:08:41.086 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:09:48.322 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_13/index_5658_seed_4764_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 11:09:51.460 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:10:09.459 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:10:12.668 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:10:13.787 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:12:08.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.28s (1.90min) +2026-06-19 11:12:09.087 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:14:08.759 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:21:10.891 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 42, train_steps: 1680 +2026-06-19 11:21:11.805 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 11:21:12.015 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:21:12.143 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:21:12.244 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:22:18.948 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000042/group_13/index_2150_seed_14979_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 11:22:22.105 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:22:39.998 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:22:40.007 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:22:40.010 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:23:47.419 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000042/group_13/index_2150_seed_14980_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 11:23:52.605 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:24:12.062 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:24:12.073 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:24:13.228 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:26:07.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.32s (1.91min) +2026-06-19 11:26:08.279 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:28:06.796 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:35:06.774 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 43, train_steps: 1720 +2026-06-19 11:35:07.458 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 11:35:07.622 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:35:07.705 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:35:07.833 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:36:13.427 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_13/index_3937_seed_22912_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 11:36:17.460 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:36:35.574 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:36:35.583 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:36:35.584 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:37:46.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_13/index_3937_seed_22913_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 11:37:49.537 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:38:07.812 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:38:10.299 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:38:11.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:40:05.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.38s (1.91min) +2026-06-19 11:40:06.549 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:42:03.526 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:49:04.436 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 44, train_steps: 1760 +2026-06-19 11:49:05.113 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 11:49:05.211 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:49:05.253 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:49:05.302 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:50:11.522 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000044/group_13/index_4019_seed_18659_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 11:50:14.903 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:50:34.160 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:50:34.170 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:50:34.170 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:51:41.718 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000044/group_13/index_4019_seed_18660_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 11:51:46.308 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:52:06.481 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:52:06.491 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:52:07.498 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:54:01.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.69s (1.89min) +2026-06-19 11:54:01.910 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:56:01.236 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:03:03.877 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 45, train_steps: 1800 +2026-06-19 12:03:04.557 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 12:03:04.698 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:03:04.743 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:03:04.845 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:04:10.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000045/group_13/index_9986_seed_26839_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 12:04:15.019 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:04:32.899 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:04:32.908 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:04:32.909 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:05:40.329 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000045/group_13/index_9986_seed_26840_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 12:05:43.413 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:06:01.582 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:06:05.347 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:06:06.278 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:08:01.417 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.05s (1.92min) +2026-06-19 12:08:02.101 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:10:01.383 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:17:03.597 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 46, train_steps: 1840 +2026-06-19 12:17:04.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 12:17:04.473 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:17:04.566 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:17:04.662 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:18:10.622 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_13/index_7606_seed_5572_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 12:18:15.898 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:18:34.177 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:18:34.185 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:18:34.185 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:19:40.430 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_13/index_7606_seed_5573_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 12:19:44.845 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:20:03.076 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:20:05.333 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:20:06.448 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:22:01.733 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.17s (1.92min) +2026-06-19 12:22:02.442 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:24:01.452 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:31:02.925 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 47, train_steps: 1880 +2026-06-19 12:31:03.607 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 12:31:03.725 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:31:03.767 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:31:03.836 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:32:09.890 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000047/group_13/index_6342_seed_33207_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 12:32:15.325 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:32:33.797 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:32:33.806 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:32:33.806 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:33:39.440 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000047/group_13/index_6342_seed_33208_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 12:33:42.609 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:34:00.751 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:34:03.182 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:34:04.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:35:58.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-19 12:35:58.983 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:37:55.892 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:44:57.687 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 48, train_steps: 1920 +2026-06-19 12:44:58.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 12:44:58.563 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:44:58.728 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:44:58.806 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:46:04.757 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_13/index_6392_seed_14817_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 12:46:07.873 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:46:25.887 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:46:25.896 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:46:25.898 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:47:34.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_13/index_6392_seed_14818_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 12:47:37.940 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:47:56.229 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:47:58.931 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:47:59.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:49:54.620 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-19 12:49:55.343 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:51:52.291 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:58:52.099 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 49, train_steps: 1960 +2026-06-19 12:58:52.782 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 12:58:52.878 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:58:52.924 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:58:52.971 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:59:59.122 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_13/index_777_seed_27062_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 13:00:04.429 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:00:22.670 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:00:22.679 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:00:22.680 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:01:30.078 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_13/index_777_seed_27063_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 13:01:35.031 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:01:53.438 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:01:54.369 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:01:55.361 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:03:48.446 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.04s (1.88min) +2026-06-19 13:03:49.148 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:05:45.468 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:12:45.278 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 50, train_steps: 2000 +2026-06-19 13:14:21.378 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=50 (missing=0, unexpected=0) +2026-06-19 13:14:23.203 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 13:14:23.406 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:14:23.531 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:14:23.610 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:15:31.417 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_13/index_6668_seed_24574_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 13:15:35.792 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:15:53.661 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:15:53.671 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:15:53.671 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:17:00.942 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_13/index_6668_seed_24575_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 13:17:06.074 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:17:24.847 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:17:25.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:17:26.229 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:19:20.081 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.84s (1.90min) +2026-06-19 13:19:20.807 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:21:17.526 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:28:18.476 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 51, train_steps: 2040 +2026-06-19 13:28:19.184 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 13:28:19.290 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:28:19.326 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:28:19.373 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:29:25.329 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000051/group_13/index_8139_seed_14655_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 13:29:30.141 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:29:48.435 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:29:48.444 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:29:48.445 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:30:53.808 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000051/group_13/index_8139_seed_14656_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 13:30:58.034 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:31:16.413 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:31:18.028 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:31:19.020 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:33:13.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.83s (1.90min) +2026-06-19 13:33:13.742 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:35:13.231 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:42:13.622 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 52, train_steps: 2080 +2026-06-19 13:42:14.277 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 13:42:14.475 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:42:14.582 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:42:14.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:43:21.154 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_13/index_1988_seed_27070_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 13:43:26.336 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:43:44.659 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:43:44.665 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:43:44.666 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:44:50.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_13/index_1988_seed_27071_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 13:44:55.205 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:45:13.428 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:45:15.287 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:45:16.402 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:47:10.481 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.03s (1.90min) +2026-06-19 13:47:11.197 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:49:11.267 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:56:12.603 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 53, train_steps: 2120 +2026-06-19 13:56:13.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 13:56:13.508 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:56:13.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:56:13.710 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:57:20.512 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_13/index_8143_seed_20295_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 13:57:25.421 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:57:43.707 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:57:43.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:57:43.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:58:48.623 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_13/index_8143_seed_20296_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 13:58:53.536 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:59:13.122 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:59:13.130 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:59:14.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:01:08.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.41s (1.91min) +2026-06-19 14:01:09.529 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:03:10.475 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:10:12.173 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 54, train_steps: 2160 +2026-06-19 14:10:12.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 14:10:13.017 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:10:13.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:10:13.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:11:19.774 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_13/index_3453_seed_2888_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 14:11:23.897 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:11:41.790 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:11:41.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:11:41.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:12:47.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_13/index_3453_seed_2889_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 14:12:50.870 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:13:09.246 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:13:12.866 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:13:13.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:15:08.493 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.42s (1.91min) +2026-06-19 14:15:09.218 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:17:09.308 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:24:11.919 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 55, train_steps: 2200 +2026-06-19 14:24:12.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 14:24:12.813 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:24:12.925 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:24:13.004 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:25:19.745 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_13/index_6570_seed_6179_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 14:25:25.002 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:25:43.066 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:25:43.076 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:25:43.076 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:26:48.272 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_13/index_6570_seed_6180_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 14:26:53.297 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:27:12.746 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:27:12.754 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:27:13.913 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:29:09.888 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.82s (1.93min) +2026-06-19 14:29:10.605 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:31:10.765 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:38:12.121 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 56, train_steps: 2240 +2026-06-19 14:38:12.811 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 14:38:12.937 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:38:12.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:38:13.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:39:20.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_13/index_2363_seed_5832_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 14:39:25.029 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:39:43.198 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:39:43.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:39:43.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:40:48.640 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_13/index_2363_seed_5833_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 14:40:53.128 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:41:12.427 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:41:13.092 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:41:14.034 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:43:08.781 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.66s (1.91min) +2026-06-19 14:43:09.475 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:45:09.719 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:52:11.855 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 57, train_steps: 2280 +2026-06-19 14:52:12.526 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 14:52:12.651 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:52:12.693 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:52:12.767 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:53:19.490 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_13/index_7125_seed_23364_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 14:53:24.703 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:53:43.119 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:53:43.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:53:43.129 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:54:48.248 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_13/index_7125_seed_23365_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 14:54:53.212 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:55:12.984 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:55:13.057 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:55:14.063 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:57:08.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.79s (1.91min) +2026-06-19 14:57:09.597 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:59:09.882 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:06:12.294 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 58, train_steps: 2320 +2026-06-19 15:06:12.985 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 15:06:13.105 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:06:13.141 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:06:13.207 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:07:20.148 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_13/index_966_seed_18340_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 15:07:25.135 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:07:43.468 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:07:43.478 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:07:43.478 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:08:48.654 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_13/index_966_seed_18341_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 15:08:53.470 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:09:12.702 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:09:13.062 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:09:14.051 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:11:08.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.71s (1.91min) +2026-06-19 15:11:09.498 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:13:09.810 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:20:12.058 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 59, train_steps: 2360 +2026-06-19 15:20:12.738 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 15:20:12.898 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:20:12.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:20:13.117 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:21:19.607 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_13/index_6316_seed_32727_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 15:21:24.445 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:21:42.841 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:21:42.849 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:21:42.852 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:22:47.820 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_13/index_6316_seed_32728_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 15:22:51.823 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:23:10.145 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:23:12.372 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:23:13.461 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:25:09.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.28s (1.92min) +2026-06-19 15:25:09.743 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:27:10.368 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:34:12.105 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 60, train_steps: 2400 +2026-06-19 15:34:48.550 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=60 (missing=0, unexpected=0) +2026-06-19 15:34:50.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 15:34:50.864 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:34:50.939 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:34:51.002 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:35:58.699 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_13/index_9776_seed_28286_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 15:36:03.999 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:36:22.447 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:36:22.456 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:36:22.456 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:37:28.373 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_13/index_9776_seed_28287_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 15:37:32.767 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:37:51.059 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:37:53.529 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:37:54.681 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:39:47.707 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-19 15:39:48.403 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:41:45.236 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:48:44.587 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 61, train_steps: 2440 +2026-06-19 15:48:45.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 15:48:45.469 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:48:45.535 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:48:45.651 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:49:51.631 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_13/index_7314_seed_24802_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 15:49:54.803 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:50:13.114 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:50:13.124 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:50:13.124 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:51:23.451 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_13/index_7314_seed_24803_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 15:51:28.825 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:51:47.599 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:51:47.609 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:51:48.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:53:41.681 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.04s (1.88min) +2026-06-19 15:53:42.400 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:55:39.328 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:02:39.009 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 62, train_steps: 2480 +2026-06-19 16:02:39.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 16:02:39.836 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:02:39.899 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:02:40.012 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:03:45.906 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000062/group_13/index_6137_seed_6575_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 16:03:50.355 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:04:08.258 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:04:08.266 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:04:08.267 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:05:18.943 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000062/group_13/index_6137_seed_6576_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 16:05:22.117 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:05:40.028 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:05:44.132 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:05:45.094 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:07:38.761 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.59s (1.89min) +2026-06-19 16:07:39.471 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:09:38.856 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:16:39.322 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 63, train_steps: 2520 +2026-06-19 16:16:40.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 16:16:40.217 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:16:40.309 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:16:40.412 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:17:46.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_13/index_1415_seed_23248_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 16:17:51.689 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:18:10.193 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:18:10.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:18:10.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:19:20.149 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_13/index_1415_seed_23249_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 16:19:25.109 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:19:43.853 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:19:45.197 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:19:46.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:21:39.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.36s (1.89min) +2026-06-19 16:21:40.303 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:23:38.883 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:30:40.942 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 64, train_steps: 2560 +2026-06-19 16:30:41.626 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 16:30:41.828 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:30:41.923 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:30:42.022 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:31:48.170 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_13/index_8480_seed_12503_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 16:31:52.119 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:32:10.188 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:32:10.197 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:32:10.197 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:33:21.124 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_13/index_8480_seed_12504_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 16:33:26.502 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:33:45.460 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:33:45.471 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:33:46.481 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:35:40.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.63s (1.89min) +2026-06-19 16:35:40.844 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:37:39.399 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:44:41.299 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 65, train_steps: 2600 +2026-06-19 16:44:41.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 16:44:42.127 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:44:42.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:44:42.326 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:45:48.651 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_13/index_2062_seed_32651_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 16:45:53.189 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:46:12.875 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:46:12.886 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:46:12.886 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:47:22.852 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_13/index_2062_seed_32652_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 16:47:26.073 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:47:44.406 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:47:47.482 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:47:48.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:49:42.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.72s (1.90min) +2026-06-19 16:49:42.959 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:51:40.038 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:58:41.285 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 66, train_steps: 2640 +2026-06-19 16:58:41.990 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 16:58:42.137 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:58:42.223 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:58:42.346 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:59:48.027 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_13/index_5792_seed_32994_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 16:59:51.171 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:00:09.304 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:00:09.313 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:00:09.314 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:01:19.584 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_13/index_5792_seed_32995_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 17:01:22.732 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:01:41.001 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:01:43.599 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:01:44.506 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:03:37.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.34s (1.89min) +2026-06-19 17:03:38.682 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:05:38.384 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:12:40.131 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 67, train_steps: 2680 +2026-06-19 17:12:40.908 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 17:12:41.033 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:12:41.263 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:12:41.379 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:13:47.762 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_13/index_2524_seed_1989_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 17:13:50.889 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:14:08.675 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:14:08.685 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:14:08.685 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:15:20.604 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_13/index_2524_seed_1990_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 17:15:25.841 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:15:44.272 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:15:44.529 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:15:45.541 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:17:38.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 17:17:39.679 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:19:39.113 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:26:39.548 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 68, train_steps: 2720 +2026-06-19 17:26:40.215 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 17:26:40.369 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:26:40.446 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:26:40.570 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:27:47.053 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_13/index_1214_seed_2800_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 17:27:51.279 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:28:09.487 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:28:09.495 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:28:09.495 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:29:20.228 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_13/index_1214_seed_2801_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 17:29:23.428 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:29:42.822 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:29:43.886 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:29:44.855 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:31:38.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.76s (1.90min) +2026-06-19 17:31:39.421 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:33:38.460 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:40:38.379 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 69, train_steps: 2760 +2026-06-19 17:40:39.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 17:40:39.139 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:40:39.175 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:40:39.223 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:41:45.108 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000069/group_13/index_4263_seed_15716_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 17:41:48.203 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:42:06.351 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:42:06.360 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:42:06.361 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:43:16.994 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000069/group_13/index_4263_seed_15717_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 17:43:22.239 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:43:40.747 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:43:41.069 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:43:42.170 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:45:35.854 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.55s (1.89min) +2026-06-19 17:45:36.570 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:47:37.117 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:54:38.582 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 70, train_steps: 2800 +2026-06-19 17:55:14.862 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=70 (missing=0, unexpected=0) +2026-06-19 17:55:16.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 17:55:17.143 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:55:17.231 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:55:17.330 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:56:25.297 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_13/index_989_seed_9324_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 17:56:29.425 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:56:47.424 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:56:47.433 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:56:47.433 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:57:54.549 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_13/index_989_seed_9325_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 17:57:59.750 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:58:18.003 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:58:18.228 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:58:19.470 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:00:13.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.58s (1.89min) +2026-06-19 18:00:13.838 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:02:13.854 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:09:13.899 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 71, train_steps: 2840 +2026-06-19 18:09:14.623 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 18:09:14.819 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:09:14.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:09:15.067 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:10:21.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_13/index_3425_seed_8335_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 18:10:25.743 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:10:43.715 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:10:43.725 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:10:43.725 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:11:51.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_13/index_3425_seed_8336_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 18:11:55.080 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:12:13.240 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:12:16.021 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:12:16.983 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:14:12.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.67s (1.93min) +2026-06-19 18:14:13.461 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:16:12.694 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:23:13.834 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 72, train_steps: 2880 +2026-06-19 18:23:14.506 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 18:23:14.686 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:23:14.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:23:14.899 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:24:21.196 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_13/index_4902_seed_13354_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 18:24:26.365 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:24:44.595 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:24:44.604 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:24:44.604 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:25:51.014 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_13/index_4902_seed_13355_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 18:25:56.015 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:26:14.219 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:26:14.711 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:26:15.873 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:28:10.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.45s (1.91min) +2026-06-19 18:28:11.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:30:10.990 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:37:12.862 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 73, train_steps: 2920 +2026-06-19 18:37:13.455 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 18:37:13.591 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:37:13.631 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:37:13.726 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:38:20.778 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_13/index_2501_seed_2845_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 18:38:24.543 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:38:42.814 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:38:42.823 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:38:42.826 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:39:50.038 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_13/index_2501_seed_2846_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 18:39:53.257 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:40:13.096 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:40:13.249 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:40:14.150 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:42:09.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.16s (1.92min) +2026-06-19 18:42:10.284 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:44:10.445 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:51:12.759 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 74, train_steps: 2960 +2026-06-19 18:51:13.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 18:51:13.643 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:51:13.738 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:51:13.847 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:52:20.678 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_13/index_5986_seed_15063_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 18:52:25.965 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:52:44.429 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:52:44.438 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:52:44.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:53:49.672 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_13/index_5986_seed_15064_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 18:53:53.960 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:54:13.070 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:54:13.982 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:54:14.897 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:56:10.312 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.21s (1.92min) +2026-06-19 18:56:11.031 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:58:11.445 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:05:14.090 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 75, train_steps: 3000 +2026-06-19 19:06:23.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 19:06:23.385 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:06:23.441 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:06:23.580 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:07:29.650 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000075/group_13/index_4367_seed_26415_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 19:07:32.743 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:07:50.937 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:07:50.947 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:07:50.949 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:08:59.057 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000075/group_13/index_4367_seed_26416_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 19:09:04.379 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:09:22.932 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:09:24.209 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:09:25.241 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:11:18.397 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.14s (1.89min) +2026-06-19 19:11:19.101 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:13:16.254 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:20:16.335 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 76, train_steps: 3040 +2026-06-19 19:20:16.989 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 19:20:17.176 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:20:17.273 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:20:17.371 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:21:22.946 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000076/group_13/index_5995_seed_23474_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 19:21:27.148 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:21:45.026 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:21:45.036 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:21:45.036 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:22:53.666 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000076/group_13/index_5995_seed_23475_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 19:22:58.587 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:23:17.948 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:23:18.455 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:23:19.582 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:25:13.441 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.66s (1.89min) +2026-06-19 19:25:14.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:27:12.736 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:34:13.279 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 77, train_steps: 3080 +2026-06-19 19:34:13.949 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 19:34:14.087 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:34:14.123 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:34:14.214 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:35:20.843 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_13/index_1338_seed_6930_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 19:35:25.972 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:35:43.956 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:35:43.965 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:35:43.965 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:36:49.200 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_13/index_1338_seed_6931_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 19:36:53.218 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:37:12.708 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:37:13.073 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:37:13.983 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:39:08.734 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.60s (1.91min) +2026-06-19 19:39:09.441 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:41:09.771 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:48:12.659 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 78, train_steps: 3120 +2026-06-19 19:48:13.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 19:48:13.522 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:48:13.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:48:13.720 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:49:20.766 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_13/index_2049_seed_10812_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 19:49:25.979 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:49:44.220 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:49:44.228 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:49:44.229 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:50:49.784 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_13/index_2049_seed_10813_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 19:50:54.044 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:51:13.167 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:51:13.323 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:51:14.244 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:53:09.267 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-19 19:53:09.977 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:55:09.987 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:02:13.178 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 79, train_steps: 3160 +2026-06-19 20:02:13.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 20:02:13.979 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:02:14.021 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:02:14.050 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:03:20.701 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000079/group_13/index_147_seed_5199_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 20:03:23.718 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:03:41.540 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:03:41.549 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:03:41.550 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:04:49.849 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000079/group_13/index_147_seed_5200_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 20:04:53.044 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:05:11.806 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:05:13.093 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:05:13.983 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:07:09.660 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.53s (1.93min) +2026-06-19 20:07:10.386 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:09:11.035 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:16:13.946 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 80, train_steps: 3200 +2026-06-19 20:16:50.657 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=80 (missing=0, unexpected=0) +2026-06-19 20:16:53.093 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 20:16:53.289 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:16:53.417 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:16:53.503 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:18:01.952 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_13/index_7507_seed_24421_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 20:18:05.077 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:18:22.094 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:18:22.105 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:18:22.105 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:19:30.869 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_13/index_7507_seed_24422_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 20:19:33.893 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:19:51.213 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:19:54.752 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:19:55.669 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:21:49.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-19 20:21:50.363 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:23:47.272 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:30:47.437 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 81, train_steps: 3240 +2026-06-19 20:30:48.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 20:30:48.268 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:30:48.307 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:30:48.383 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:31:54.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_13/index_3919_seed_701_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 20:31:57.813 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:32:14.384 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:32:14.394 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:32:14.394 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:33:25.002 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_13/index_3919_seed_702_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 20:33:27.987 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:33:44.909 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:33:49.273 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:33:50.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:35:44.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.70s (1.89min) +2026-06-19 20:35:44.752 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:37:41.679 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:44:39.998 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 82, train_steps: 3280 +2026-06-19 20:44:40.696 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 20:44:40.813 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:44:40.854 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:44:40.926 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:45:46.932 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_13/index_2289_seed_6623_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 20:45:49.941 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:46:07.138 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:46:07.149 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:46:07.149 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:47:19.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_13/index_2289_seed_6624_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 20:47:23.741 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:47:41.175 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:47:44.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:47:45.082 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:49:38.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.41s (1.89min) +2026-06-19 20:49:39.275 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:51:38.425 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:58:38.996 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 83, train_steps: 3320 +2026-06-19 20:58:39.739 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 20:58:39.896 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:58:39.989 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:58:40.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:59:46.126 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_13/index_8216_seed_1916_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 20:59:49.136 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:00:06.154 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:00:06.163 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:00:06.164 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:01:19.720 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_13/index_8216_seed_1917_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 21:01:22.752 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:01:39.772 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:01:44.304 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:01:45.269 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:03:38.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.28s (1.89min) +2026-06-19 21:03:39.340 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:05:38.820 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:12:38.940 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 84, train_steps: 3360 +2026-06-19 21:12:39.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 21:12:39.788 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:12:39.830 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:12:39.948 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:13:46.159 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_13/index_6012_seed_113_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 21:13:49.207 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:14:06.027 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:14:06.036 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:14:06.036 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:15:19.438 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_13/index_6012_seed_114_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 21:15:22.470 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:15:39.605 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:15:43.791 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:15:44.700 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:17:38.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.31s (1.89min) +2026-06-19 21:17:38.864 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:19:38.524 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:26:39.504 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 85, train_steps: 3400 +2026-06-19 21:26:40.170 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 21:26:40.394 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:26:40.526 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:26:40.588 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:27:46.810 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_13/index_8791_seed_30121_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 21:27:51.838 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:28:09.069 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:28:09.077 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:28:09.078 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:29:20.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_13/index_8791_seed_30122_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 21:29:24.807 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:29:42.378 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:29:43.697 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:29:44.717 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:31:38.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-19 21:31:38.988 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:33:38.757 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:40:38.654 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 86, train_steps: 3440 +2026-06-19 21:40:39.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 21:40:39.458 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:40:39.507 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:40:39.559 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:41:46.102 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_13/index_9862_seed_994_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 21:41:50.791 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:42:07.935 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:42:07.944 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:42:07.945 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:43:18.571 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_13/index_9862_seed_995_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 21:43:23.410 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:43:41.002 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:43:42.867 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:43:43.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:45:37.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 21:45:37.988 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:47:37.710 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:54:39.052 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 87, train_steps: 3480 +2026-06-19 21:54:39.743 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 21:54:39.870 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:54:39.926 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:54:40.018 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:55:45.840 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000087/group_13/index_2508_seed_18552_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 21:55:50.559 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:56:08.039 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:56:08.049 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:56:08.049 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:57:19.189 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000087/group_13/index_2508_seed_18553_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 21:57:24.001 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:57:41.821 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:57:43.169 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:57:44.172 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:59:37.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.39s (1.89min) +2026-06-19 21:59:38.318 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:01:38.557 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:08:39.996 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 88, train_steps: 3520 +2026-06-19 22:08:40.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 22:08:40.835 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:08:40.903 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:08:41.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:09:46.782 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_13/index_4851_seed_14535_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 22:09:50.743 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:10:08.167 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:10:08.177 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:10:08.177 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:11:19.655 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_13/index_4851_seed_14536_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 22:11:23.389 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:11:40.556 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:11:43.755 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:11:44.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:13:38.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.57s (1.89min) +2026-06-19 22:13:39.103 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:15:38.305 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:22:38.709 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 89, train_steps: 3560 +2026-06-19 22:22:39.424 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 22:22:39.618 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:22:39.761 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:22:39.854 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:23:45.511 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_13/index_3953_seed_11254_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 22:23:48.658 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:24:05.491 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:24:05.501 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:24:05.501 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:25:19.695 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_13/index_3953_seed_11255_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 22:25:24.783 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:25:42.193 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:25:43.683 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:25:44.783 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:27:38.162 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 22:27:38.887 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:29:38.496 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:36:39.430 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 90, train_steps: 3600 +2026-06-19 22:37:15.150 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=90 (missing=0, unexpected=0) +2026-06-19 22:37:17.307 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 22:37:17.492 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:37:17.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:37:17.682 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:38:24.480 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_13/index_9668_seed_22378_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 22:38:28.186 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:38:44.359 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:38:44.368 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:38:44.368 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:39:55.121 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_13/index_9668_seed_22379_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 22:39:58.286 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:40:14.544 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:40:19.241 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:40:20.235 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:42:14.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.70s (1.89min) +2026-06-19 22:42:14.863 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:44:13.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:51:14.856 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 91, train_steps: 3640 +2026-06-19 22:51:15.699 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 22:51:15.817 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:51:15.847 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:51:15.896 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:52:22.030 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000091/group_13/index_5555_seed_4522_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 22:52:25.125 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:52:41.483 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:52:41.493 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:52:41.493 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:53:51.749 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000091/group_13/index_5555_seed_4523_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 22:53:56.553 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:54:13.381 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:54:15.937 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:54:16.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:56:12.672 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.67s (1.93min) +2026-06-19 22:56:13.389 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:58:13.112 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:05:14.271 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 92, train_steps: 3680 +2026-06-19 23:05:15.118 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 23:05:15.330 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:05:15.455 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:05:15.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:06:21.448 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000092/group_13/index_9001_seed_7189_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 23:06:25.061 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:06:41.570 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:06:41.580 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:06:41.580 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:07:51.299 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000092/group_13/index_9001_seed_7190_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 23:07:56.199 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:08:13.201 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:08:16.058 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:08:17.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:10:11.751 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.62s (1.91min) +2026-06-19 23:10:12.471 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:12:11.284 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:19:13.485 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 93, train_steps: 3720 +2026-06-19 23:19:14.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 23:19:14.271 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:19:14.305 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:19:14.352 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:20:20.970 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_13/index_6810_seed_9096_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 23:20:25.664 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:20:42.758 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:20:42.768 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:20:42.768 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:21:50.154 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_13/index_6810_seed_9097_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 23:21:54.127 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:22:11.190 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:22:14.354 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:22:15.497 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:24:10.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.33s (1.92min) +2026-06-19 23:24:11.708 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:26:11.286 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:33:13.942 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 94, train_steps: 3760 +2026-06-19 23:33:14.641 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 23:33:14.848 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:33:14.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:33:15.057 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:34:21.275 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_13/index_9350_seed_14301_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 23:34:24.553 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:34:41.389 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:34:41.399 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:34:41.400 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:35:51.525 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_13/index_9350_seed_14302_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 23:35:54.578 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:36:11.791 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:36:16.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:36:17.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:38:12.062 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.89s (1.91min) +2026-06-19 23:38:12.784 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:40:11.321 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:47:11.885 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 95, train_steps: 3800 +2026-06-19 23:47:12.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-19 23:47:12.775 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:47:12.866 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:47:12.977 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:48:19.521 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_13/index_8849_seed_24974_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-19 23:48:22.781 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:48:39.539 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:48:39.548 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:48:39.549 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:49:48.123 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_13/index_8849_seed_24975_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-19 23:49:53.105 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:50:09.966 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:50:12.004 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:50:13.254 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:52:08.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.82s (1.91min) +2026-06-19 23:52:08.791 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:54:07.678 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:01:09.783 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 96, train_steps: 3840 +2026-06-20 00:01:10.475 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 00:01:10.680 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:01:10.897 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:01:11.013 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:02:17.377 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_13/index_1544_seed_3864_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 00:02:21.112 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:02:38.100 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:02:38.110 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:02:38.110 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:03:47.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_13/index_1544_seed_3865_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 00:03:52.592 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:04:09.635 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:04:11.783 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:04:13.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:06:07.927 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 00:06:08.645 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:08:07.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:15:08.512 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 97, train_steps: 3880 +2026-06-20 00:15:09.220 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 00:15:09.418 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:15:09.529 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:15:09.628 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:16:15.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_13/index_3151_seed_13375_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 00:16:19.218 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:16:36.102 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:16:36.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:16:36.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:17:47.264 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_13/index_3151_seed_13376_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 00:17:52.251 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:18:09.553 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:18:11.238 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:18:12.451 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:20:07.671 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.20s (1.92min) +2026-06-20 00:20:08.401 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:22:07.127 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:29:07.015 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 98, train_steps: 3920 +2026-06-20 00:29:07.691 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 00:29:07.785 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:29:07.836 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:29:07.883 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:30:13.903 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000098/group_13/index_2185_seed_28618_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 00:30:18.639 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:30:35.986 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:30:35.996 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:30:35.996 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:31:46.022 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000098/group_13/index_2185_seed_28619_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 00:31:49.992 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:32:06.624 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:32:10.419 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:32:11.506 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:34:06.389 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 00:34:07.117 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:36:03.852 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:43:05.349 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 99, train_steps: 3960 +2026-06-20 00:43:06.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 00:43:06.233 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:43:06.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:43:06.437 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:44:12.333 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_13/index_5759_seed_5854_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 00:44:17.288 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:44:34.421 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:44:34.431 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:44:34.431 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:45:46.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_13/index_5759_seed_5855_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 00:45:50.178 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:46:06.516 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:46:09.899 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:46:10.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:48:05.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.75s (1.91min) +2026-06-20 00:48:06.440 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:50:03.399 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:57:04.807 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 100, train_steps: 4000 +2026-06-20 00:58:47.733 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=100 (missing=0, unexpected=0) +2026-06-20 00:58:49.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 00:58:49.739 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:58:49.819 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:58:49.939 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:59:57.276 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_13/index_6556_seed_20228_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 01:00:02.199 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:00:19.006 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:00:19.013 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:00:19.013 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:01:26.662 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_13/index_6556_seed_20229_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 01:01:31.421 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:01:48.568 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:01:50.303 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:01:51.316 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:03:44.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-20 01:03:45.434 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:05:42.220 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:12:39.869 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 101, train_steps: 4040 +2026-06-20 01:12:40.567 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 01:12:40.727 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:12:40.794 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:12:40.888 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:13:47.129 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_13/index_1275_seed_11048_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 01:13:50.900 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:14:07.874 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:14:07.881 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:14:07.881 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:15:20.148 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_13/index_1275_seed_11049_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 01:15:23.331 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:15:40.082 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:15:43.499 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:15:44.427 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:17:38.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.62s (1.89min) +2026-06-20 01:17:38.931 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:19:38.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:26:37.827 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 102, train_steps: 4080 +2026-06-20 01:26:38.503 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 01:26:38.659 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:26:38.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:26:38.832 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:27:44.521 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_13/index_3251_seed_30214_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 01:27:47.651 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:28:04.464 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:28:04.474 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:28:04.474 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:29:17.177 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_13/index_3251_seed_30215_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 01:29:20.370 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:29:36.921 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:29:41.183 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:29:42.401 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:31:36.383 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-20 01:31:37.098 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:33:37.226 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:40:38.546 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 103, train_steps: 4120 +2026-06-20 01:40:39.244 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 01:40:39.357 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:40:39.393 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:40:39.442 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:41:45.619 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000103/group_13/index_6184_seed_13847_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 01:41:48.605 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:42:05.301 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:42:05.310 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:42:05.310 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:43:19.204 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000103/group_13/index_6184_seed_13848_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 01:43:24.147 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:43:40.784 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:43:43.334 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:43:44.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:45:37.767 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-20 01:45:38.488 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:47:38.266 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:54:37.797 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 104, train_steps: 4160 +2026-06-20 01:54:38.702 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 01:54:38.877 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:54:38.951 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:54:39.075 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:55:45.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_13/index_9683_seed_27122_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 01:55:49.859 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:56:06.937 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:56:06.947 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:56:06.947 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:57:18.651 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_13/index_9683_seed_27123_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 01:57:22.586 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:57:39.224 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:57:42.889 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:57:43.863 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:59:37.314 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-20 01:59:38.031 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:01:38.021 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:08:38.461 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 105, train_steps: 4200 +2026-06-20 02:08:39.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 02:08:39.301 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:08:39.350 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:08:39.458 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:09:45.658 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000105/group_13/index_2006_seed_13145_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 02:09:48.663 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:10:05.552 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:10:05.562 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:10:05.562 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:11:19.328 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000105/group_13/index_2006_seed_13146_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 02:11:24.236 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:11:41.185 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:11:43.945 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:11:44.974 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:13:38.217 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 02:13:38.926 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:15:38.270 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:22:37.864 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 106, train_steps: 4240 +2026-06-20 02:22:38.564 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 02:22:38.723 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:22:38.790 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:22:38.899 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:23:44.881 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000106/group_13/index_2035_seed_3302_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 02:23:48.825 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:24:05.749 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:24:05.758 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:24:05.758 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:25:16.774 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000106/group_13/index_2035_seed_3303_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 02:25:19.864 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:25:36.725 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:25:40.847 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:25:41.784 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:27:35.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.90s (1.90min) +2026-06-20 02:27:36.583 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:29:36.814 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:36:37.099 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 107, train_steps: 4280 +2026-06-20 02:36:37.801 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 02:36:37.980 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:36:38.094 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:36:38.202 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:37:43.998 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_13/index_3856_seed_26868_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 02:37:46.940 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:38:03.756 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:38:03.765 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:38:03.766 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:39:15.142 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_13/index_3856_seed_26869_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 02:39:18.271 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:39:34.633 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:39:39.229 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:39:40.323 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:41:35.347 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 02:41:36.072 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:43:35.148 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:50:35.008 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 108, train_steps: 4320 +2026-06-20 02:50:35.727 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 02:50:35.903 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:50:35.996 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:50:36.077 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:51:41.983 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_13/index_5348_seed_17441_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 02:51:47.002 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:52:03.718 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:52:03.728 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:52:03.728 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:53:13.092 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_13/index_5348_seed_17442_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 02:53:17.835 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:53:34.776 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:53:37.573 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:53:38.593 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:55:33.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.45s (1.91min) +2026-06-20 02:55:33.973 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:57:30.650 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:04:30.704 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 109, train_steps: 4360 +2026-06-20 03:04:31.407 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 03:04:31.597 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:04:31.702 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:04:31.811 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:05:37.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000109/group_13/index_1364_seed_23009_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 03:05:41.452 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:05:57.781 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:05:57.790 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:05:57.790 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:07:09.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000109/group_13/index_1364_seed_23010_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 03:07:13.170 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:07:29.549 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:07:33.407 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:07:34.369 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:09:29.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-20 03:09:30.014 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:11:26.577 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:18:25.347 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 110, train_steps: 4400 +2026-06-20 03:19:01.546 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=110 (missing=0, unexpected=0) +2026-06-20 03:19:03.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 03:19:03.499 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:19:03.587 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:19:03.682 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:20:11.672 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_13/index_7512_seed_14828_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 03:20:14.912 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:20:31.598 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:20:31.609 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:20:31.609 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:21:43.337 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_13/index_7512_seed_14829_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 03:21:48.237 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:22:04.506 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:22:07.524 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:22:08.537 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:24:03.966 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.41s (1.92min) +2026-06-20 03:24:04.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:26:02.474 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:33:05.400 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 111, train_steps: 4440 +2026-06-20 03:33:06.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 03:33:06.179 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:33:06.221 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:33:06.269 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:34:12.298 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_13/index_635_seed_32242_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 03:34:16.807 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:34:32.910 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:34:32.920 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:34:32.920 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:35:46.341 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_13/index_635_seed_32243_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 03:35:49.894 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:36:05.961 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:36:10.098 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:36:11.183 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:38:06.021 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 03:38:06.732 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:40:04.483 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:47:04.320 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 112, train_steps: 4480 +2026-06-20 03:47:05.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 03:47:05.245 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:47:05.357 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:47:05.433 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:48:11.273 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_13/index_9147_seed_8078_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 03:48:16.147 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:48:32.438 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:48:32.448 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:48:32.450 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:49:40.866 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_13/index_9147_seed_8079_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 03:49:45.499 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:50:01.748 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:50:04.983 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:50:06.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:52:00.943 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-20 03:52:01.658 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:54:00.965 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:01:04.011 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 113, train_steps: 4520 +2026-06-20 04:01:04.875 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 04:01:05.002 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:01:05.046 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:01:05.131 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:02:10.744 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000113/group_13/index_9284_seed_30050_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 04:02:14.369 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:02:30.010 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:02:30.020 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:02:30.022 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:03:42.065 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000113/group_13/index_9284_seed_30051_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 04:03:46.957 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:04:03.272 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:04:06.611 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:04:07.652 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:06:02.814 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.15s (1.92min) +2026-06-20 04:06:03.532 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:08:01.584 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:15:03.784 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 114, train_steps: 4560 +2026-06-20 04:15:04.474 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 04:15:04.650 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:15:04.747 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:15:04.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:16:10.681 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000114/group_13/index_6270_seed_2827_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 04:16:14.459 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:16:30.080 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:16:30.086 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:16:30.086 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:17:40.330 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000114/group_13/index_6270_seed_2828_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 04:17:45.183 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:18:01.449 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:18:04.152 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:18:05.373 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:20:00.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.72s (1.91min) +2026-06-20 04:20:00.895 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:22:00.101 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:29:01.474 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 115, train_steps: 4600 +2026-06-20 04:29:02.193 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 04:29:02.386 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:29:02.487 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:29:02.586 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:30:08.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_13/index_5278_seed_32106_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 04:30:11.638 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:30:28.581 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:30:28.590 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:30:28.591 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:31:38.978 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_13/index_5278_seed_32107_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 04:31:42.029 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:31:57.921 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:32:02.494 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:32:03.399 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:33:58.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.33s (1.91min) +2026-06-20 04:33:58.776 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:35:55.627 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:42:56.481 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 116, train_steps: 4640 +2026-06-20 04:42:57.160 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 04:42:57.352 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:42:57.454 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:42:57.537 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:44:03.734 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_13/index_2676_seed_21921_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 04:44:06.567 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:44:22.498 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:44:22.508 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:44:22.510 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:45:34.079 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_13/index_2676_seed_21922_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 04:45:38.850 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:45:55.390 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:45:58.380 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:45:59.408 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:47:53.698 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.25s (1.90min) +2026-06-20 04:47:54.407 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:49:50.890 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:56:50.233 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 117, train_steps: 4680 +2026-06-20 04:56:50.955 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 04:56:51.148 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:56:51.248 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:56:51.339 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:57:57.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_13/index_7079_seed_5683_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 04:58:01.750 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:58:17.965 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:58:17.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:58:17.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:59:28.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_13/index_7079_seed_5684_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 04:59:32.579 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:59:48.943 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:59:51.798 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:59:52.808 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:01:46.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.49s (1.89min) +2026-06-20 05:01:47.078 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:03:44.145 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:10:42.591 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 118, train_steps: 4720 +2026-06-20 05:10:43.295 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 05:10:43.424 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:10:43.461 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:10:43.530 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:11:49.825 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000118/group_13/index_3406_seed_3984_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 05:11:53.524 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:12:09.123 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:12:09.133 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:12:09.133 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:13:22.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000118/group_13/index_3406_seed_3985_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 05:13:27.806 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:13:44.064 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:13:46.957 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:13:47.996 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:15:41.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.26s (1.89min) +2026-06-20 05:15:42.001 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:17:38.996 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:24:37.910 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 119, train_steps: 4760 +2026-06-20 05:24:38.622 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 05:24:38.763 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:24:38.803 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:24:38.893 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:25:44.812 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_13/index_4248_seed_24472_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 05:25:49.459 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:26:05.429 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:26:05.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:26:05.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:27:14.939 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_13/index_4248_seed_24473_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 05:27:18.896 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:27:34.562 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:27:40.119 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:27:41.236 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:29:34.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-20 05:29:35.560 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:31:33.107 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:38:32.153 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 120, train_steps: 4800 +2026-06-20 05:39:09.282 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=120 (missing=0, unexpected=0) +2026-06-20 05:39:10.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 05:39:11.052 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:39:11.132 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:39:11.235 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:40:19.822 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000120/group_13/index_7807_seed_8862_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 05:40:22.703 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:40:38.906 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:40:38.916 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:40:38.917 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:41:48.448 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000120/group_13/index_7807_seed_8863_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 05:41:53.336 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:42:09.877 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:42:12.566 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:42:13.594 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:44:08.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.96s (1.92min) +2026-06-20 05:44:09.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:46:10.088 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:53:12.066 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 121, train_steps: 4840 +2026-06-20 05:53:12.775 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 05:53:12.909 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:53:12.944 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:53:13.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:54:20.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_13/index_9596_seed_22236_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 05:54:24.679 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:54:40.937 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:54:40.946 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:54:40.947 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:55:48.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_13/index_9596_seed_22237_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 05:55:52.131 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:56:08.329 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:56:13.156 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:56:14.131 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:58:09.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.97s (1.92min) +2026-06-20 05:58:10.076 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:00:10.576 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:07:11.957 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 122, train_steps: 4880 +2026-06-20 06:07:12.649 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 06:07:12.858 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:07:12.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:07:13.062 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:08:19.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_13/index_3897_seed_26347_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 06:08:22.790 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:08:38.662 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:08:38.672 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:08:38.672 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:09:48.757 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_13/index_3897_seed_26348_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 06:09:51.800 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:10:07.847 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:10:13.195 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:10:14.119 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:12:08.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.25s (1.90min) +2026-06-20 06:12:09.327 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:14:10.582 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:21:11.886 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 123, train_steps: 4920 +2026-06-20 06:21:12.565 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 06:21:12.769 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:21:12.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:21:12.946 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:22:19.259 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_13/index_9788_seed_17707_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 06:22:24.124 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:22:40.380 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:22:40.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:22:40.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:23:47.960 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_13/index_9788_seed_17708_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 06:23:52.647 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:24:09.172 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:24:12.071 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:24:13.294 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:26:08.788 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.47s (1.92min) +2026-06-20 06:26:09.511 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:28:09.835 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:35:11.860 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 124, train_steps: 4960 +2026-06-20 06:35:12.538 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 06:35:12.727 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:35:12.796 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:35:12.898 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:36:19.294 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_13/index_7815_seed_11192_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 06:36:23.981 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:36:40.277 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:36:40.288 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:36:40.288 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:37:47.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_13/index_7815_seed_11193_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 06:37:51.529 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:38:07.694 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:38:12.129 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:38:13.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:40:08.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.44s (1.92min) +2026-06-20 06:40:09.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:42:09.775 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:49:11.660 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 125, train_steps: 5000 +2026-06-20 06:50:15.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 06:50:15.392 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:50:15.486 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:50:15.619 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:51:21.871 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_13/index_9467_seed_12199_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 06:51:24.758 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:51:40.658 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:51:40.668 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:51:40.668 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:52:51.282 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_13/index_9467_seed_12200_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 06:52:54.328 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:53:10.287 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:53:15.926 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:53:16.850 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:55:11.654 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.68s (1.91min) +2026-06-20 06:55:12.372 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:57:11.765 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:04:12.000 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 126, train_steps: 5040 +2026-06-20 07:04:12.632 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 07:04:12.853 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:04:12.933 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:04:13.027 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:05:19.923 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_13/index_365_seed_18377_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 07:05:22.947 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:05:38.481 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:05:38.490 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:05:38.491 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:06:49.350 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_13/index_365_seed_18378_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 07:06:54.231 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:07:10.495 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:07:13.988 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:07:15.235 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:09:10.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.08s (1.92min) +2026-06-20 07:09:11.147 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:11:11.167 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:18:11.941 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 127, train_steps: 5080 +2026-06-20 07:18:12.629 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 07:18:12.790 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:18:12.850 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:18:12.971 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:19:19.419 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_13/index_7377_seed_27572_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 07:19:23.224 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:19:38.837 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:19:38.846 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:19:38.847 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:20:48.816 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_13/index_7377_seed_27573_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 07:20:52.178 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:21:08.347 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:21:13.050 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:21:14.010 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:23:09.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.50s (1.93min) +2026-06-20 07:23:10.327 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:25:10.159 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:32:11.191 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 128, train_steps: 5120 +2026-06-20 07:32:11.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 07:32:12.075 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:32:12.167 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:32:12.247 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:33:18.722 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_13/index_9519_seed_25439_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 07:33:21.771 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:33:38.685 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:33:38.694 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:33:38.695 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:34:48.781 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_13/index_9519_seed_25440_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 07:34:52.428 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:35:09.105 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:35:13.016 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:35:14.040 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:37:08.641 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.56s (1.91min) +2026-06-20 07:37:09.373 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:39:09.932 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:46:11.993 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 129, train_steps: 5160 +2026-06-20 07:46:12.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 07:46:12.909 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:46:13.021 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:46:13.119 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:47:20.113 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_13/index_6141_seed_28742_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 07:47:24.750 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:47:40.609 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:47:40.618 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:47:40.619 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:48:48.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_13/index_6141_seed_28743_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 07:48:52.614 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:49:08.535 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:49:13.070 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:49:14.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:51:09.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-20 07:51:09.758 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:53:10.306 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:00:11.795 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 130, train_steps: 5200 +2026-06-20 08:00:49.332 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=130 (missing=0, unexpected=0) +2026-06-20 08:00:51.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 08:00:51.440 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:00:51.559 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:00:51.638 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:02:00.310 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_13/index_3889_seed_24953_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 08:02:05.105 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:02:21.730 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:02:21.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:02:21.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:03:28.802 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_13/index_3889_seed_24954_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 08:03:32.384 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:03:48.458 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:03:52.509 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:03:53.465 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:05:47.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.72s (1.90min) +2026-06-20 08:05:48.018 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:07:44.846 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:14:45.324 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 131, train_steps: 5240 +2026-06-20 08:14:46.026 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 08:14:46.240 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:14:46.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:14:46.433 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:15:52.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_13/index_1204_seed_29056_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 08:15:55.309 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:16:12.396 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:16:12.407 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:16:12.410 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:17:23.309 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_13/index_1204_seed_29057_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 08:17:26.933 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:17:43.250 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:17:46.990 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:17:47.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:19:41.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.38s (1.89min) +2026-06-20 08:19:42.139 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:21:39.469 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:28:37.843 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 132, train_steps: 5280 +2026-06-20 08:28:38.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 08:28:38.654 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:28:38.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:28:38.774 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:29:44.954 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000132/group_13/index_87_seed_29687_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 08:29:49.359 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:30:05.596 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:30:05.606 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:30:05.606 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:31:15.655 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000132/group_13/index_87_seed_29688_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 08:31:19.541 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:31:35.503 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:31:40.849 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:31:41.832 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:33:35.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.44s (1.89min) +2026-06-20 08:33:36.119 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:35:35.264 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:42:34.340 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 133, train_steps: 5320 +2026-06-20 08:42:35.020 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 08:42:35.232 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:42:35.343 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:42:35.409 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:43:41.621 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_13/index_2570_seed_15362_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 08:43:44.483 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:44:01.639 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:44:01.649 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:44:01.649 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:45:11.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_13/index_2570_seed_15363_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 08:45:15.680 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:45:33.572 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:45:35.888 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:45:36.927 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:47:32.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.77s (1.93min) +2026-06-20 08:47:33.450 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:49:30.693 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:56:29.435 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 134, train_steps: 5360 +2026-06-20 08:56:30.116 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 08:56:30.309 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:56:30.411 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:56:30.491 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:57:36.425 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000134/group_13/index_769_seed_196_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 08:57:41.133 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:57:57.458 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:57:57.469 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:57:57.469 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:59:07.672 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000134/group_13/index_769_seed_197_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 08:59:12.616 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:59:29.480 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:59:32.040 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:59:33.068 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:01:26.319 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 09:01:27.048 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:03:23.806 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:10:22.990 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 135, train_steps: 5400 +2026-06-20 09:10:23.676 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 09:10:23.875 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:10:23.980 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:10:24.060 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:11:30.335 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000135/group_13/index_619_seed_26367_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 09:11:35.005 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:11:51.139 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:11:51.149 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:11:51.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:13:00.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000135/group_13/index_619_seed_26368_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 09:13:04.297 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:13:20.280 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:13:23.932 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:13:24.901 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:15:18.588 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.61s (1.89min) +2026-06-20 09:15:19.310 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:17:17.101 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:24:16.803 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 136, train_steps: 5440 +2026-06-20 09:24:17.533 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 09:24:17.736 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:24:17.834 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:24:17.929 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:25:23.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_13/index_5624_seed_29387_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 09:25:28.721 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:25:44.665 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:25:44.674 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:25:44.674 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:26:54.216 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_13/index_5624_seed_29388_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 09:26:57.931 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:27:14.029 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:27:18.656 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:27:19.634 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:29:12.958 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.15s (1.89min) +2026-06-20 09:29:13.686 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:31:13.179 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:38:13.223 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 137, train_steps: 5480 +2026-06-20 09:38:13.939 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 09:38:14.140 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:38:14.251 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:38:14.330 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:39:20.845 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000137/group_13/index_4489_seed_5134_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 09:39:24.004 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:39:40.051 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:39:40.062 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:39:40.064 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:40:50.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000137/group_13/index_4489_seed_5135_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 09:40:53.324 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:41:09.531 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:41:14.371 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:41:15.513 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:43:09.522 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-20 09:43:10.251 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:45:10.505 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:52:12.158 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 138, train_steps: 5520 +2026-06-20 09:52:12.867 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 09:52:13.050 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:52:13.132 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:52:13.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:53:20.257 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_13/index_4678_seed_12859_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 09:53:25.053 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:53:41.086 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:53:41.096 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:53:41.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:54:49.311 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_13/index_4678_seed_12860_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 09:54:54.011 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:55:10.462 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:55:13.383 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:55:14.604 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:57:10.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.46s (1.92min) +2026-06-20 09:57:10.921 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:59:11.153 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:06:12.677 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 139, train_steps: 5560 +2026-06-20 10:06:13.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 10:06:13.568 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:06:13.659 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:06:13.737 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:07:20.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_13/index_8876_seed_11909_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 10:07:23.677 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:07:39.415 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:07:39.425 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:07:39.425 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:08:50.190 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_13/index_8876_seed_11910_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 10:08:53.196 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:09:09.212 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:09:13.264 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:09:14.169 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:11:08.675 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.20s (1.90min) +2026-06-20 10:11:09.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:13:09.783 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:20:12.699 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 140, train_steps: 5600 +2026-06-20 10:20:49.837 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=140 (missing=0, unexpected=0) +2026-06-20 10:20:51.813 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 10:20:51.989 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:20:52.078 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:20:52.190 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:21:59.840 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000140/group_13/index_4226_seed_1121_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 10:22:04.790 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:22:21.527 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:22:21.537 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:22:21.537 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:23:29.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000140/group_13/index_4226_seed_1122_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 10:23:33.281 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:23:49.485 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:23:53.040 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:23:53.995 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:25:47.600 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-20 10:25:48.313 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:27:45.365 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:34:44.009 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 141, train_steps: 5640 +2026-06-20 10:34:44.692 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 10:34:44.911 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:34:45.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:34:45.094 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:35:50.908 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_13/index_6245_seed_7335_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 10:35:53.977 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:36:10.214 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:36:10.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:36:10.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:37:22.875 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_13/index_6245_seed_7336_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 10:37:25.916 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:37:42.203 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:37:46.846 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:37:47.784 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:39:41.415 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.51s (1.89min) +2026-06-20 10:39:42.148 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:41:39.464 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:48:38.465 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 142, train_steps: 5680 +2026-06-20 10:48:39.159 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 10:48:39.374 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:48:39.478 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:48:39.586 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:49:45.871 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000142/group_13/index_1590_seed_31296_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 10:49:49.007 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:50:04.948 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:50:04.958 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:50:04.959 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:51:19.255 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000142/group_13/index_1590_seed_31297_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 10:51:22.256 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:51:38.243 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:51:43.153 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:51:44.077 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:53:37.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-20 10:53:37.851 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:55:37.995 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:02:37.948 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 143, train_steps: 5720 +2026-06-20 11:02:38.631 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 11:02:38.794 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:02:38.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:02:38.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:03:45.232 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000143/group_13/index_4827_seed_706_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 11:03:50.087 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:04:06.415 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:04:06.425 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:04:06.426 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:05:18.286 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000143/group_13/index_4827_seed_707_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 11:05:23.139 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:05:39.729 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:05:42.169 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:05:43.339 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:07:36.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 11:07:37.313 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:09:36.845 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:16:37.106 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 144, train_steps: 5760 +2026-06-20 11:16:37.800 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 11:16:37.997 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:16:38.097 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:16:38.178 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:17:43.873 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_13/index_589_seed_17212_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 11:17:48.487 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:18:04.388 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:18:04.398 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:18:04.400 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:19:13.075 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_13/index_589_seed_17213_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 11:19:17.736 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:19:34.166 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:19:38.707 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:19:39.833 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:21:33.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.47s (1.89min) +2026-06-20 11:21:34.087 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:23:30.891 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:30:30.394 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 145, train_steps: 5800 +2026-06-20 11:30:31.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 11:30:31.319 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:30:31.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:30:31.502 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:31:37.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000145/group_13/index_1666_seed_17972_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 11:31:41.728 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:31:58.074 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:31:58.085 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:31:58.085 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:33:08.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000145/group_13/index_1666_seed_17973_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 11:33:13.886 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:33:30.623 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:33:33.207 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:33:34.240 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:35:28.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.89s (1.90min) +2026-06-20 11:35:28.900 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:37:25.682 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:44:23.925 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 146, train_steps: 5840 +2026-06-20 11:44:24.629 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 11:44:24.805 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:44:24.908 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:44:25.002 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:45:30.998 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_13/index_9152_seed_13464_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 11:45:35.770 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:45:51.945 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:45:51.954 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:45:51.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:46:59.147 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_13/index_9152_seed_13465_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 11:47:02.780 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:47:19.074 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:47:23.247 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:47:24.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:49:19.133 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-20 11:49:19.856 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:51:17.041 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:58:15.925 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 147, train_steps: 5880 +2026-06-20 11:58:16.619 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 11:58:16.831 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:58:16.955 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:58:17.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:59:23.229 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_13/index_8628_seed_27218_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 11:59:26.216 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:59:42.057 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:59:42.068 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:59:42.068 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:00:52.744 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_13/index_8628_seed_27219_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 12:00:55.764 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:01:12.401 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:01:16.232 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:01:17.200 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:03:12.586 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.29s (1.92min) +2026-06-20 12:03:13.294 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:05:12.906 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:12:12.709 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 148, train_steps: 5920 +2026-06-20 12:12:13.405 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 12:12:13.574 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:12:13.657 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:12:13.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:13:20.802 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000148/group_13/index_8319_seed_8832_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 12:13:23.829 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:13:39.960 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:13:39.970 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:13:39.971 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:14:49.693 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000148/group_13/index_8319_seed_8833_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 12:14:52.812 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:15:09.350 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:15:13.144 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:15:14.067 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:17:09.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.69s (1.91min) +2026-06-20 12:17:09.789 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:19:10.590 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:26:11.396 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 149, train_steps: 5960 +2026-06-20 12:26:12.133 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 12:26:12.254 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:26:12.290 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:26:12.355 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:27:19.177 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000149/group_13/index_3388_seed_12319_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 12:27:22.117 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:27:38.604 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:27:38.614 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:27:38.615 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:28:48.592 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000149/group_13/index_3388_seed_12320_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 12:28:53.321 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:29:09.901 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:29:13.029 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:29:14.060 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:31:08.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.72s (1.91min) +2026-06-20 12:31:09.532 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:33:09.829 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:40:10.992 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 150, train_steps: 6000 +2026-06-20 12:41:53.768 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=150 (missing=0, unexpected=0) +2026-06-20 12:41:56.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 12:41:56.195 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:41:56.262 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:41:56.397 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:43:04.124 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_13/index_8369_seed_19154_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 12:43:08.897 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:43:24.973 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:43:24.983 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:43:24.983 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:44:32.774 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_13/index_8369_seed_19155_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 12:44:36.730 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:44:53.037 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:44:59.744 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:45:00.692 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:46:55.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.70s (1.91min) +2026-06-20 12:46:56.236 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:48:52.990 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:55:51.610 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 151, train_steps: 6040 +2026-06-20 12:55:52.306 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 12:55:52.473 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:55:52.549 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:55:52.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:56:58.551 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_13/index_415_seed_16169_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 12:57:01.537 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:57:17.745 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:57:17.755 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:57:17.755 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:58:27.108 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_13/index_415_seed_16170_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 12:58:30.244 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:58:47.165 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:58:50.721 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:58:51.695 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:00:46.189 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.41s (1.91min) +2026-06-20 13:00:46.908 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:02:43.654 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:09:41.910 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 152, train_steps: 6080 +2026-06-20 13:09:42.596 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 13:09:42.724 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:09:42.759 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:09:42.829 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:10:49.097 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_13/index_3755_seed_5371_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 13:10:53.996 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:11:10.264 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:11:10.274 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:11:10.274 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:12:20.497 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_13/index_3755_seed_5372_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 13:12:25.418 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:12:41.977 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:12:43.910 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:12:44.931 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:14:38.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-20 13:14:39.492 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:16:38.552 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:23:37.222 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 153, train_steps: 6120 +2026-06-20 13:23:38.095 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 13:23:38.310 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:23:38.441 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:23:38.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:24:45.250 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_13/index_6229_seed_18112_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 13:24:49.102 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:25:04.784 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:25:04.794 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:25:04.794 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:26:15.089 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_13/index_6229_seed_18113_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 13:26:18.240 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:26:34.425 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:26:40.127 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:26:41.258 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:28:34.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.39s (1.89min) +2026-06-20 13:28:35.473 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:30:32.909 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:37:31.703 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 154, train_steps: 6160 +2026-06-20 13:37:32.394 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 13:37:32.580 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:37:32.657 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:37:32.754 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:38:38.704 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_13/index_4191_seed_19795_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 13:38:43.619 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:38:59.708 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:38:59.718 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:38:59.719 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:40:09.291 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_13/index_4191_seed_19796_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 13:40:13.950 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:40:30.494 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:40:33.788 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:40:34.817 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:42:29.329 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.47s (1.91min) +2026-06-20 13:42:30.062 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:44:27.160 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:51:24.983 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 155, train_steps: 6200 +2026-06-20 13:51:25.699 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 13:51:25.806 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:51:25.850 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:51:25.895 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:52:32.012 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_13/index_420_seed_20087_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 13:52:36.688 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:52:53.155 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:52:53.164 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:52:53.165 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:54:01.078 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_13/index_420_seed_20088_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 13:54:04.869 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:54:21.246 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:54:24.447 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:54:25.487 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:56:19.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.63s (1.89min) +2026-06-20 13:56:19.859 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:58:16.751 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:05:15.346 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 156, train_steps: 6240 +2026-06-20 14:05:16.037 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 14:05:16.263 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:05:16.393 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:05:16.475 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:06:22.588 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_13/index_3371_seed_8824_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 14:06:27.394 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:06:43.654 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:06:43.664 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:06:43.667 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:07:51.710 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_13/index_3371_seed_8825_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 14:07:55.580 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:08:13.024 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:08:15.931 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:08:16.903 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:10:11.722 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.74s (1.91min) +2026-06-20 14:10:12.457 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:12:11.107 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:19:12.248 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 157, train_steps: 6280 +2026-06-20 14:19:12.924 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 14:19:13.068 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:19:13.105 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:19:13.197 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:20:19.959 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000157/group_13/index_4294_seed_25660_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 14:20:24.857 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:20:41.029 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:20:41.039 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:20:41.039 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:21:47.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000157/group_13/index_4294_seed_25661_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 14:21:51.431 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:22:07.819 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:22:12.398 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:22:13.449 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:24:08.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.22s (1.92min) +2026-06-20 14:24:09.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:26:09.906 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:33:10.832 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 158, train_steps: 6320 +2026-06-20 14:33:11.520 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 14:33:11.721 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:33:11.817 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:33:11.913 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:34:18.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_13/index_5135_seed_8191_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 14:34:21.651 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:34:37.319 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:34:37.329 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:34:37.329 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:35:47.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_13/index_5135_seed_8192_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 14:35:52.026 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:36:08.325 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:36:10.935 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:36:12.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:38:07.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.88s (1.91min) +2026-06-20 14:38:07.784 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:40:05.259 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:47:04.569 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 159, train_steps: 6360 +2026-06-20 14:47:05.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 14:47:05.444 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:47:05.495 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:47:05.607 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:48:11.402 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_13/index_2210_seed_29662_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 14:48:14.992 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:48:31.145 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:48:31.154 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:48:31.155 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:49:40.116 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_13/index_2210_seed_29663_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 14:49:45.148 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:50:01.837 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:50:04.048 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:50:05.262 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:51:59.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.86s (1.90min) +2026-06-20 14:51:59.929 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:53:57.323 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:00:59.183 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 160, train_steps: 6400 +2026-06-20 15:01:37.149 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=160 (missing=0, unexpected=0) +2026-06-20 15:01:39.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 15:01:39.446 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:01:39.511 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:01:39.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:02:46.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_13/index_3882_seed_20183_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 15:02:50.155 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:03:05.928 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:03:05.939 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:03:05.939 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:04:19.333 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_13/index_3882_seed_20184_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 15:04:22.601 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:04:38.869 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:04:42.879 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:04:43.841 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:06:37.409 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.48s (1.89min) +2026-06-20 15:06:38.137 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:08:38.302 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:15:37.530 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 161, train_steps: 6440 +2026-06-20 15:15:38.205 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 15:15:38.296 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:15:38.345 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:15:38.393 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:16:44.541 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_13/index_5232_seed_17896_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 15:16:47.519 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:17:03.227 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:17:03.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:17:03.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:18:16.575 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_13/index_5232_seed_17897_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 15:18:21.457 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:18:37.604 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:18:40.375 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:18:41.590 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:20:35.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-20 15:20:36.332 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:22:36.151 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:29:36.901 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 162, train_steps: 6480 +2026-06-20 15:29:37.570 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 15:29:37.769 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:29:37.786 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:29:37.904 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:30:43.909 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_13/index_9568_seed_2887_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 15:30:47.642 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:31:03.131 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:31:03.141 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:31:03.142 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:32:11.872 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_13/index_9568_seed_2888_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 15:32:14.874 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:32:30.663 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:32:37.444 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:32:38.419 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:34:33.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.48s (1.91min) +2026-06-20 15:34:33.827 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:36:31.320 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:43:31.491 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 163, train_steps: 6520 +2026-06-20 15:43:32.171 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 15:43:32.305 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:43:32.343 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:43:32.435 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:44:38.646 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000163/group_13/index_3871_seed_27025_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 15:44:41.649 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:44:57.419 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:44:57.430 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:44:57.430 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:46:09.453 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000163/group_13/index_3871_seed_27026_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 15:46:13.148 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:46:29.215 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:46:33.676 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:46:34.609 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:48:29.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.47s (1.91min) +2026-06-20 15:48:29.935 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:50:26.540 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:57:25.529 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 164, train_steps: 6560 +2026-06-20 15:57:26.228 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 15:57:26.397 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:57:26.472 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:57:26.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:58:32.323 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_13/index_373_seed_8582_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 15:58:37.155 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:58:53.268 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:58:53.279 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:58:53.281 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:00:02.482 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_13/index_373_seed_8583_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 16:00:07.146 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:00:23.583 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:00:26.401 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:00:27.412 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:02:21.064 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.61s (1.89min) +2026-06-20 16:02:21.793 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:04:18.345 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:11:18.378 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 165, train_steps: 6600 +2026-06-20 16:11:19.089 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 16:11:19.304 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:11:19.424 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:11:19.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:12:25.827 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_13/index_527_seed_21889_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 16:12:30.432 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:12:46.440 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:12:46.449 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:12:46.450 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:13:54.683 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_13/index_527_seed_21890_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 16:13:59.277 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:14:15.704 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:14:19.105 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:14:20.276 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:16:14.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.75s (1.90min) +2026-06-20 16:16:14.870 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:18:14.127 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:25:13.934 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 166, train_steps: 6640 +2026-06-20 16:25:14.621 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 16:25:14.822 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:25:14.946 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:25:15.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:26:21.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_13/index_2916_seed_4401_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 16:26:26.116 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:26:42.006 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:26:42.017 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:26:42.017 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:27:50.085 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_13/index_2916_seed_4402_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 16:27:53.627 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:28:09.341 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:28:13.831 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:28:14.905 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:30:08.844 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.78s (1.90min) +2026-06-20 16:30:09.574 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:32:10.308 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:39:12.162 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 167, train_steps: 6680 +2026-06-20 16:39:12.849 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 16:39:12.994 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:39:13.043 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:39:13.157 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:40:20.263 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_13/index_7874_seed_7002_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 16:40:25.054 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:40:40.598 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:40:40.609 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:40:40.609 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:41:48.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_13/index_7874_seed_7003_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 16:41:51.570 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:42:07.861 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:42:12.095 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:42:13.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:44:08.745 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.36s (1.92min) +2026-06-20 16:44:09.477 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:46:10.022 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:53:11.386 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 168, train_steps: 6720 +2026-06-20 16:53:12.088 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 16:53:12.204 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:53:12.251 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:53:12.320 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:54:19.617 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_13/index_8294_seed_4232_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 16:54:23.161 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:54:39.102 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:54:39.113 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:54:39.113 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:55:47.473 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_13/index_8294_seed_4233_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 16:55:50.482 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:56:07.478 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:56:11.767 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:56:12.770 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:58:07.736 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-20 16:58:08.464 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:00:07.581 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:07:09.184 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 169, train_steps: 6760 +2026-06-20 17:07:09.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 17:07:10.069 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 17:07:10.166 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:07:10.301 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:08:16.649 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_13/index_8808_seed_5849_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 17:08:19.766 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 17:08:35.490 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:08:35.501 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:08:35.501 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:09:47.602 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_13/index_8808_seed_5850_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 17:09:52.538 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 17:10:08.960 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:10:11.904 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 17:10:13.042 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:12:08.254 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.20s (1.92min) +2026-06-20 17:12:08.969 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:14:09.323 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:21:11.059 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 170, train_steps: 6800 +2026-06-20 17:21:47.551 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=170 (missing=0, unexpected=0) +2026-06-20 17:21:49.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 17:21:49.430 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 17:21:49.548 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:21:49.649 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:22:56.412 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_13/index_8639_seed_30511_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 17:23:00.047 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 17:23:16.487 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:23:16.498 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:23:16.498 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:24:25.549 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_13/index_8639_seed_30512_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 17:24:30.144 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 17:24:46.417 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:24:49.925 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 17:24:50.940 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:26:44.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.05s (1.88min) +2026-06-20 17:26:44.743 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:28:41.427 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:35:39.738 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 54, sp_rank 0: update step: 171, train_steps: 6840 +2026-06-20 17:35:40.433 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 54, dp_rank 27: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=13, rank_in_group=1 +2026-06-20 17:35:40.665 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 17:35:40.899 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:35:41.020 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:36:47.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000171/group_13/index_8490_seed_30297_dp_27_rank_54_sample_2_slot_0.mp4 +2026-06-20 17:36:51.616 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 17:37:07.677 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:37:07.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 54 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:37:07.688 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 54 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:38:19.204 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 54: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000171/group_13/index_8490_seed_30298_dp_27_rank_54_sample_3_slot_1.mp4 +2026-06-20 17:38:23.825 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 17:38:40.100 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:38:42.537 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 54: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 17:38:43.742 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 54: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:40:36.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 54: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.89s (1.88min) +2026-06-20 17:40:37.378 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 54, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-29 01:30:50.580 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 01:30:51.255 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 01:30:51.256 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 01:53:18.794 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 01:53:19.453 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 01:53:19.455 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 02:22:42.610 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 02:22:43.281 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 02:22:43.283 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 02:48:32.585 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 02:48:33.235 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 02:48:33.237 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:09:17.895 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 03:09:18.555 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:09:18.556 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:34:32.240 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 03:34:32.909 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:34:32.909 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:00:48.930 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 04:00:49.588 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:00:49.591 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:28:44.611 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 04:28:45.292 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:28:45.293 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:54:00.638 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 04:54:01.309 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:54:01.309 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 05:17:21.819 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 05:17:22.504 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 05:17:22.505 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 05:40:48.034 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 05:40:48.718 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 05:40:48.718 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:04:33.987 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 06:04:34.666 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:04:34.666 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:27:29.339 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 06:27:30.001 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:27:30.001 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:51:08.331 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 06:51:08.992 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:51:08.992 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 07:14:09.520 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 07:14:10.183 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 07:14:10.186 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 07:38:36.916 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 07:38:37.573 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 07:38:37.574 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:09:35.403 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 08:09:36.075 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:09:36.076 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:32:25.959 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 08:32:26.621 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:32:26.622 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:01:07.973 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 09:01:08.644 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:01:08.645 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:26:49.179 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 09:26:49.874 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:26:49.874 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:49:41.123 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 09:49:41.802 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:49:41.803 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 10:12:07.330 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 10:12:07.999 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 10:12:08.000 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 10:38:21.747 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 10:38:22.409 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 10:38:22.412 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:05:35.753 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 11:05:36.433 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:05:36.433 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:29:13.708 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 11:29:14.381 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:29:14.382 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:57:16.931 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 11:57:17.586 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:57:17.587 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 12:22:52.432 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 12:22:53.092 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 12:22:53.092 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 12:50:48.012 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 12:50:48.686 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 12:50:48.687 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 13:14:45.450 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 13:14:46.125 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 13:14:46.126 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 13:38:09.098 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 13:38:09.754 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 13:38:09.755 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:07:57.110 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 14:07:57.787 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:07:57.787 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:28:52.231 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 14:28:52.894 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:28:52.896 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:51:31.736 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 14:51:32.406 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:51:32.407 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 15:22:59.516 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 15:23:00.170 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 15:23:00.171 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 15:49:42.069 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 15:49:42.728 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 15:49:42.729 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 16:17:45.331 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 16:17:45.994 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 16:17:45.996 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 16:42:21.796 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 16:42:22.481 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 16:42:22.481 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:11:39.727 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 17:11:40.396 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:11:40.398 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:34:41.950 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 17:34:42.630 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:34:42.631 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:56:06.017 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 17:56:06.670 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:56:06.671 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 18:18:49.468 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 18:18:50.141 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 18:18:50.142 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 18:41:54.695 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 18:41:55.344 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 18:41:55.345 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_6.log b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_6.log new file mode 100644 index 0000000000000000000000000000000000000000..a76e3aabc935ce216aa4dade725c229d17d5f746 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_6.log @@ -0,0 +1,3241 @@ +2026-06-19 01:30:11.770 | INFO | __main__:__init__:933 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-19 01:30:12.440 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-19 01:30:12.441 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-19 01:30:17.294 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 01:30:31.243 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 01:30:31.247 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:30:31.247 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:31:52.745 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_1/index_8214_seed_1874_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 01:31:57.723 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 01:32:15.351 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:32:15.360 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:32:15.360 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:33:22.371 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_1/index_8214_seed_1875_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 01:33:26.068 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 01:33:43.584 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:33:45.514 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 01:34:00.959 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 01:35:56.996 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.94s (1.93min) +2026-06-19 01:35:57.689 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 01:37:54.182 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 01:37:57.219 | INFO | hyvideo.utils.grpo_utils:train_one_step:3127 - Rank 6, sp_rank 0: init adaptive KL controller (init=1.000e-07, min=1.000e-07, max=1.000e-05) +2026-06-19 01:45:11.970 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 1, train_steps: 40 +2026-06-19 01:45:12.575 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 01:45:12.717 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 01:45:12.754 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:45:12.808 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:46:18.656 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000001/group_1/index_3217_seed_28162_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 01:46:23.651 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 01:46:41.081 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:46:41.090 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:46:41.090 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:47:47.189 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000001/group_1/index_3217_seed_28163_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 01:47:51.910 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 01:48:09.124 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:48:10.944 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 01:48:12.041 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 01:50:07.820 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.70s (1.93min) +2026-06-19 01:50:08.465 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 01:52:06.953 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 01:59:06.764 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 2, train_steps: 80 +2026-06-19 01:59:07.382 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 01:59:07.507 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 01:59:07.571 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:59:07.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:00:13.392 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000002/group_1/index_8238_seed_31700_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 02:00:17.101 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:00:35.007 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:00:35.016 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:00:35.016 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:01:46.305 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000002/group_1/index_8238_seed_31701_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 02:01:49.420 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:02:07.114 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:02:10.528 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:02:11.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:04:06.723 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.88s (1.91min) +2026-06-19 02:04:07.374 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:06:03.754 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:13:04.860 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 3, train_steps: 120 +2026-06-19 02:13:05.465 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 02:13:05.550 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:13:05.551 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:13:05.551 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:14:11.142 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_1/index_7920_seed_8834_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 02:14:14.221 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:14:31.176 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:14:31.183 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:14:31.183 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:15:40.544 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_1/index_7920_seed_8835_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 02:15:45.549 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:16:03.415 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:16:04.655 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:16:05.875 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:17:59.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.77s (1.90min) +2026-06-19 02:18:00.448 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:19:59.350 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:26:59.882 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 4, train_steps: 160 +2026-06-19 02:27:00.546 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 02:27:00.629 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:27:00.655 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:27:00.756 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:28:06.546 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000004/group_1/index_5336_seed_29245_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 02:28:11.361 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:28:28.967 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:28:28.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:28:28.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:29:35.110 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000004/group_1/index_5336_seed_29246_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 02:29:39.895 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:29:57.668 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:30:00.315 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:30:01.191 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:31:55.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.62s (1.91min) +2026-06-19 02:31:56.651 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:33:53.223 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:40:53.219 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 5, train_steps: 200 +2026-06-19 02:40:53.901 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 02:40:54.055 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:40:54.189 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:40:54.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:41:59.996 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_1/index_383_seed_6484_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 02:42:04.774 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:42:22.287 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:42:22.295 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:42:22.296 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:43:28.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_1/index_383_seed_6485_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 02:43:32.540 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:43:50.794 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:43:51.811 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:43:52.682 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:45:46.316 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.50s (1.89min) +2026-06-19 02:45:46.993 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:47:43.352 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:54:40.663 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 6, train_steps: 240 +2026-06-19 02:54:41.361 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 02:54:41.466 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:54:41.610 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:54:41.933 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:55:47.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_1/index_7097_seed_16271_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 02:55:50.810 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:56:07.938 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:56:07.947 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:56:07.947 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:57:19.388 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_1/index_7097_seed_16272_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 02:57:24.430 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 02:57:42.055 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:57:42.949 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:57:43.964 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:59:37.353 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.38s (1.89min) +2026-06-19 02:59:38.043 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:01:37.731 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:08:37.476 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 7, train_steps: 280 +2026-06-19 03:08:38.148 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 03:08:38.257 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:08:38.392 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:08:38.523 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:09:44.244 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_1/index_186_seed_15946_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 03:09:48.156 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:10:05.485 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:10:05.495 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:10:05.495 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:11:14.766 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_1/index_186_seed_15947_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 03:11:18.151 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:11:37.241 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:11:38.337 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:11:39.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:13:34.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.68s (1.91min) +2026-06-19 03:13:34.977 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:15:32.043 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:22:31.149 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 8, train_steps: 320 +2026-06-19 03:22:31.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 03:22:31.938 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:22:32.055 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:22:32.160 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:23:38.739 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000008/group_1/index_4634_seed_25611_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 03:23:41.895 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:23:59.284 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:23:59.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:23:59.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:25:09.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000008/group_1/index_4634_seed_25612_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 03:25:14.008 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:25:31.781 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:25:33.100 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:25:33.963 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:27:28.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-19 03:27:29.431 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:29:26.904 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:36:25.649 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 9, train_steps: 360 +2026-06-19 03:36:26.346 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 03:36:26.470 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:36:26.542 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:36:26.680 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:37:32.668 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_1/index_2247_seed_4961_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 03:37:35.772 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:37:53.232 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:37:53.242 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:37:53.242 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:39:02.091 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_1/index_2247_seed_4962_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 03:39:07.295 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:39:25.227 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:39:26.010 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:39:27.037 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:41:20.399 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.36s (1.89min) +2026-06-19 03:41:21.105 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:43:17.563 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:50:16.931 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 10, train_steps: 400 +2026-06-19 03:50:55.049 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=10 (missing=0, unexpected=0) +2026-06-19 03:50:56.181 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 03:50:56.279 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:50:56.410 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:50:56.540 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:52:03.771 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_1/index_9023_seed_384_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 03:52:07.771 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:52:25.164 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:52:25.174 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:52:25.174 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:53:32.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_1/index_9023_seed_385_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 03:53:35.588 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 03:53:53.010 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:53:57.827 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:53:58.697 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:55:52.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.65s (1.89min) +2026-06-19 03:55:53.208 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:57:49.763 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:04:48.389 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 11, train_steps: 440 +2026-06-19 04:04:49.074 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 04:04:49.250 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:04:49.354 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:04:49.468 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:05:55.117 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_1/index_7238_seed_9509_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 04:05:58.281 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:06:15.587 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:06:15.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:06:15.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:07:24.615 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_1/index_7238_seed_9510_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 04:07:27.827 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:07:46.100 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:07:48.852 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:07:49.875 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:09:43.124 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.25s (1.89min) +2026-06-19 04:09:43.797 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:11:40.270 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:18:37.304 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 12, train_steps: 480 +2026-06-19 04:18:37.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 04:18:38.065 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:18:38.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:18:38.303 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:19:44.321 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_1/index_2225_seed_25762_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 04:19:49.340 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:20:07.071 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:20:07.080 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:20:07.080 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:21:12.833 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_1/index_2225_seed_25763_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 04:21:17.052 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:21:36.535 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:21:37.180 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:21:38.109 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:23:33.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.19s (1.92min) +2026-06-19 04:23:34.070 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:25:31.244 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:32:29.819 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 13, train_steps: 520 +2026-06-19 04:32:30.512 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 04:32:30.598 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:32:30.640 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:32:30.763 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:33:36.560 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_1/index_8116_seed_21087_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 04:33:39.699 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:33:56.883 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:33:56.893 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:33:56.894 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:35:06.014 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_1/index_8116_seed_21088_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 04:35:09.214 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:35:26.791 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:35:30.312 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:35:31.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:37:24.102 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.77s (1.88min) +2026-06-19 04:37:24.791 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:39:21.077 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:46:18.658 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 14, train_steps: 560 +2026-06-19 04:46:19.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 04:46:19.346 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:46:19.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:46:19.559 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:47:26.559 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_1/index_4676_seed_29981_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 04:47:31.742 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:47:49.336 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:47:49.341 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:47:49.342 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:48:54.291 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_1/index_4676_seed_29982_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 04:48:58.298 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 04:49:16.236 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:49:17.445 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:49:18.470 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:51:12.963 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.49s (1.91min) +2026-06-19 04:51:13.650 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:53:12.996 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:00:11.829 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 15, train_steps: 600 +2026-06-19 05:00:12.526 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 05:00:12.701 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:00:12.828 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:00:12.907 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:01:19.203 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000015/group_1/index_3178_seed_10174_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 05:01:23.103 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:01:40.377 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:01:40.386 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:01:40.387 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:02:47.586 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000015/group_1/index_3178_seed_10175_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 05:02:50.768 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:03:08.213 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:03:11.349 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:03:12.557 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:05:07.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.64s (1.91min) +2026-06-19 05:05:08.060 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:07:05.978 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:14:04.896 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 16, train_steps: 640 +2026-06-19 05:14:05.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 05:14:05.700 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:14:05.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:14:05.907 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:15:11.495 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_1/index_5838_seed_31267_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 05:15:14.666 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:15:31.779 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:15:31.788 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:15:31.789 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:16:43.166 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_1/index_5838_seed_31268_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 05:16:46.476 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:17:04.174 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:17:06.988 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:17:07.837 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:19:03.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.02s (1.92min) +2026-06-19 05:19:03.741 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:21:01.960 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:28:03.321 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 17, train_steps: 680 +2026-06-19 05:28:04.007 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 05:28:04.185 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:28:04.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:28:04.398 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:29:10.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_1/index_5510_seed_29845_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 05:29:15.604 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:29:33.244 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:29:33.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:29:33.258 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:30:39.587 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_1/index_5510_seed_29846_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 05:30:43.089 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:31:00.975 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:31:03.926 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:31:04.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:32:58.688 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.48s (1.89min) +2026-06-19 05:32:59.386 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:34:56.265 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:41:58.098 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 18, train_steps: 720 +2026-06-19 05:41:58.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 05:41:58.880 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:41:58.926 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:41:59.018 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:43:05.262 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_1/index_2695_seed_29087_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 05:43:10.503 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:43:28.230 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:43:28.239 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:43:28.240 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:44:35.542 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_1/index_2695_seed_29088_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 05:44:40.444 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:44:58.383 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:45:00.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:45:01.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:46:55.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.88s (1.90min) +2026-06-19 05:46:55.797 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:48:52.331 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:55:50.498 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 19, train_steps: 760 +2026-06-19 05:55:51.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 05:55:51.333 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:55:51.430 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:55:51.541 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:56:57.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_1/index_5421_seed_15590_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 05:57:02.058 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:57:20.342 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:57:20.350 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:57:20.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:58:25.477 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_1/index_5421_seed_15591_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 05:58:30.405 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 05:58:48.468 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:58:49.068 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:58:50.076 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:00:43.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.10s (1.89min) +2026-06-19 06:00:43.876 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:02:40.342 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:09:39.760 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 20, train_steps: 800 +2026-06-19 06:10:16.340 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=20 (missing=0, unexpected=0) +2026-06-19 06:10:18.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 06:10:18.956 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:10:19.065 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:10:19.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:11:27.078 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_1/index_148_seed_22590_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 06:11:31.959 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:11:49.632 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:11:49.641 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:11:49.641 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:12:54.662 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_1/index_148_seed_22591_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 06:12:59.574 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:13:17.457 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:13:19.050 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:13:20.312 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:15:13.224 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.90s (1.88min) +2026-06-19 06:15:13.910 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:17:12.689 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:24:12.558 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 21, train_steps: 840 +2026-06-19 06:24:13.271 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 06:24:13.418 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:24:13.552 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:24:13.643 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:25:19.465 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_1/index_5218_seed_9006_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 06:25:24.383 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:25:42.220 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:25:42.228 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:25:42.229 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:26:47.544 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_1/index_5218_seed_9007_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 06:26:51.744 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:27:09.278 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:27:11.260 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:27:12.260 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:29:06.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.51s (1.91min) +2026-06-19 06:29:07.667 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:31:04.585 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:38:04.283 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 22, train_steps: 880 +2026-06-19 06:38:04.947 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 06:38:05.037 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:38:05.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:38:05.159 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:39:10.972 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_1/index_1588_seed_28764_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 06:39:14.110 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:39:31.638 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:39:31.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:39:31.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:40:39.826 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_1/index_1588_seed_28765_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 06:40:43.082 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:41:01.156 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:41:03.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:41:04.200 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:42:58.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.99s (1.90min) +2026-06-19 06:42:59.165 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:44:56.504 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:51:58.173 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 23, train_steps: 920 +2026-06-19 06:51:58.842 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 06:51:58.984 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:51:59.124 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:51:59.240 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:53:04.826 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000023/group_1/index_4441_seed_10167_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 06:53:08.032 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:53:25.599 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:53:25.608 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:53:25.608 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:54:35.053 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000023/group_1/index_4441_seed_10168_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 06:54:38.261 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 06:54:55.998 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:54:59.789 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:55:00.680 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:56:54.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.11s (1.90min) +2026-06-19 06:56:55.622 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:58:52.319 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:05:51.319 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 24, train_steps: 960 +2026-06-19 07:05:52.005 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 07:05:52.154 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:05:52.290 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:05:52.408 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:06:58.334 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_1/index_7804_seed_9786_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 07:07:01.605 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:07:20.483 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:07:20.491 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:07:20.491 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:08:27.619 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_1/index_7804_seed_9787_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 07:08:32.820 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:08:52.207 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:08:52.217 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:08:53.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:10:46.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-19 07:10:46.871 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:12:44.154 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:19:41.972 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 25, train_steps: 1000 +2026-06-19 07:20:52.105 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 07:20:52.230 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:20:52.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:20:52.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:21:58.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_1/index_6214_seed_6185_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 07:22:02.856 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:22:21.174 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:22:21.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:22:21.183 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:23:28.688 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_1/index_6214_seed_6186_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 07:23:32.872 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:23:51.454 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:23:52.666 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:23:53.608 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:25:46.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.81s (1.88min) +2026-06-19 07:25:47.206 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:27:44.382 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:34:43.687 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 26, train_steps: 1040 +2026-06-19 07:34:44.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 07:34:44.463 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:34:44.530 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:34:44.672 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:35:50.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_1/index_4563_seed_21866_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 07:35:53.588 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:36:11.665 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:36:11.674 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:36:11.675 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:37:21.035 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_1/index_4563_seed_21867_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 07:37:26.315 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:37:45.074 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:37:45.443 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:37:46.471 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:39:39.638 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.14s (1.89min) +2026-06-19 07:39:40.348 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:41:39.072 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:48:39.675 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 27, train_steps: 1080 +2026-06-19 07:48:40.347 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 07:48:40.438 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:48:40.520 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:48:40.667 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:49:46.442 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_1/index_8919_seed_23385_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 07:49:50.620 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:50:08.741 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:50:08.750 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:50:08.750 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:51:19.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_1/index_8919_seed_23386_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 07:51:22.538 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 07:51:40.720 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:51:43.285 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:51:44.195 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:53:38.298 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.99s (1.90min) +2026-06-19 07:53:39.007 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:55:38.423 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:02:38.173 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 28, train_steps: 1120 +2026-06-19 08:02:38.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 08:02:38.968 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:02:39.026 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:02:39.172 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:03:45.109 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_1/index_6203_seed_17102_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 08:03:48.323 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:04:06.422 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:04:06.431 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:04:06.431 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:05:17.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_1/index_6203_seed_17103_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 08:05:22.725 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:05:41.503 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:05:41.901 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:05:43.100 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:07:37.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.24s (1.90min) +2026-06-19 08:07:38.091 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:09:37.936 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:16:37.848 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 29, train_steps: 1160 +2026-06-19 08:16:38.691 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 08:16:38.785 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:16:38.873 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:16:39.177 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:17:44.742 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_1/index_7038_seed_18279_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 08:17:49.541 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:18:07.886 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:18:07.895 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:18:07.895 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:19:17.980 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_1/index_7038_seed_18280_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 08:19:22.905 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:19:41.661 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:19:42.881 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:19:43.866 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:21:37.264 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 08:21:37.961 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:23:37.407 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:30:39.545 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 30, train_steps: 1200 +2026-06-19 08:31:15.592 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=30 (missing=0, unexpected=0) +2026-06-19 08:31:18.258 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 08:31:18.465 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:31:18.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:31:18.673 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:32:27.512 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000030/group_1/index_7076_seed_4936_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 08:32:31.422 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:32:49.691 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:32:49.701 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:32:49.701 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:33:55.883 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000030/group_1/index_7076_seed_4937_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 08:33:59.209 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:34:17.553 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:34:19.980 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:34:21.038 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:36:14.510 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-19 08:36:15.227 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:38:14.172 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:45:14.284 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 31, train_steps: 1240 +2026-06-19 08:45:14.992 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 08:45:15.098 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:45:15.197 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:45:15.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:46:21.469 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_1/index_7881_seed_11786_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 08:46:26.815 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:46:45.230 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:46:45.236 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:46:45.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:47:50.611 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_1/index_7881_seed_11787_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 08:47:55.450 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 08:48:13.942 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:48:15.288 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:48:16.463 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:50:10.322 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.84s (1.90min) +2026-06-19 08:50:11.024 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:52:11.150 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:59:13.097 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 32, train_steps: 1280 +2026-06-19 08:59:13.777 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 08:59:13.904 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:59:14.028 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:59:14.177 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:00:20.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000032/group_1/index_3351_seed_7095_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 09:00:23.970 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:00:42.054 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:00:42.063 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:00:42.063 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:01:48.763 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000032/group_1/index_3351_seed_7096_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 09:01:51.884 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:02:09.951 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:02:13.436 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:02:14.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:04:08.713 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.91s (1.90min) +2026-06-19 09:04:09.421 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:06:10.410 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:13:13.280 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 33, train_steps: 1320 +2026-06-19 09:13:13.952 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 09:13:14.037 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:13:14.067 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:13:14.210 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:14:20.092 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000033/group_1/index_3402_seed_29783_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 09:14:25.219 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:14:44.927 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:14:44.937 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:14:44.937 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:15:49.939 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000033/group_1/index_3402_seed_29784_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 09:15:54.877 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:16:14.812 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:16:14.822 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:16:15.905 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:18:11.065 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.06s (1.92min) +2026-06-19 09:18:11.786 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:20:11.628 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:27:13.731 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 34, train_steps: 1360 +2026-06-19 09:27:14.446 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 09:27:14.589 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:27:14.711 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:27:14.818 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:28:20.547 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_1/index_3118_seed_7934_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 09:28:25.331 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:28:43.715 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:28:43.725 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:28:43.725 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:29:49.965 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_1/index_3118_seed_7935_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 09:29:54.848 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:30:13.637 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:30:13.695 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:30:15.007 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:32:08.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.81s (1.90min) +2026-06-19 09:32:09.591 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:34:09.778 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:41:12.399 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 35, train_steps: 1400 +2026-06-19 09:41:13.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 09:41:13.167 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:41:13.250 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:41:13.377 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:42:19.327 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000035/group_1/index_7018_seed_25876_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 09:42:23.341 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:42:41.545 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:42:41.551 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:42:41.551 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:43:48.490 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000035/group_1/index_7018_seed_25877_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 09:43:53.710 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:44:12.579 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:44:13.079 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:44:14.124 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:46:10.686 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 116.56s (1.94min) +2026-06-19 09:46:11.399 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:48:10.897 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:55:11.960 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 36, train_steps: 1440 +2026-06-19 09:55:12.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 09:55:12.754 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:55:12.851 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:55:12.989 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:56:19.212 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_1/index_6948_seed_28197_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 09:56:23.343 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:56:41.822 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:56:41.828 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:56:41.828 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:57:47.920 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_1/index_6948_seed_28198_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 09:57:53.096 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 09:58:11.897 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:58:12.646 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:58:13.881 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:00:08.966 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-19 10:00:09.672 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:02:10.007 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:09:11.707 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 37, train_steps: 1480 +2026-06-19 10:09:12.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 10:09:12.548 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:09:12.694 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:09:12.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:10:18.465 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_1/index_7510_seed_31429_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 10:10:23.405 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:10:41.921 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:10:41.930 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:10:41.930 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:11:47.656 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_1/index_7510_seed_31430_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 10:11:51.846 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:12:10.401 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:12:13.041 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:12:13.936 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:14:09.029 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.94s (1.92min) +2026-06-19 10:14:09.738 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:16:10.367 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:23:12.346 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 38, train_steps: 1520 +2026-06-19 10:23:13.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 10:23:13.177 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:23:13.257 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:23:13.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:24:19.177 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000038/group_1/index_839_seed_25386_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 10:24:22.308 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:24:40.170 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:24:40.180 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:24:40.180 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:25:48.640 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000038/group_1/index_839_seed_25387_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 10:25:53.844 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:26:12.396 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:26:12.666 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:26:13.894 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:28:09.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.16s (1.92min) +2026-06-19 10:28:09.954 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:30:10.742 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:37:11.949 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 39, train_steps: 1560 +2026-06-19 10:37:12.646 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 10:37:12.734 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:37:12.766 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:37:12.888 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:38:19.028 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000039/group_1/index_7722_seed_22548_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 10:38:23.144 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:38:41.498 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:38:41.504 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:38:41.504 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:41:17.501 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000039/group_1/index_7722_seed_22549_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 10:41:20.955 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:41:39.386 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:41:42.526 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:41:43.592 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:43:36.471 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.70s (1.88min) +2026-06-19 10:43:37.180 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:45:37.176 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:52:37.864 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 40, train_steps: 1600 +2026-06-19 10:53:14.006 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=40 (missing=0, unexpected=0) +2026-06-19 10:53:15.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 10:53:15.851 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:53:15.866 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:53:15.973 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:54:23.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_1/index_6989_seed_24240_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 10:54:28.822 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:54:45.672 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:54:45.679 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:54:45.679 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:55:52.364 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_1/index_6989_seed_24241_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 10:55:56.078 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 10:56:13.088 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:56:16.410 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:56:17.295 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:58:11.717 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.27s (1.90min) +2026-06-19 10:58:12.416 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:00:11.329 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:07:12.593 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 41, train_steps: 1640 +2026-06-19 11:07:13.291 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 11:07:13.379 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:07:13.412 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:07:13.538 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:08:19.593 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_1/index_1154_seed_11024_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 11:08:22.523 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:08:40.661 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:08:40.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:08:40.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:09:48.481 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_1/index_1154_seed_11025_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 11:09:53.841 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:10:12.210 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:10:12.668 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:10:13.950 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:12:08.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.28s (1.90min) +2026-06-19 11:12:09.087 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:14:08.757 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:21:10.893 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 42, train_steps: 1680 +2026-06-19 11:21:11.805 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 11:21:11.945 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:21:12.019 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:21:12.158 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:22:18.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000042/group_1/index_4362_seed_14654_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 11:22:22.293 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:22:40.291 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:22:40.300 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:22:40.300 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:23:47.368 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000042/group_1/index_4362_seed_14655_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 11:23:52.634 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:24:11.299 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:24:12.072 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:24:13.251 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:26:07.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.32s (1.91min) +2026-06-19 11:26:08.279 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:28:06.798 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:35:06.773 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 43, train_steps: 1720 +2026-06-19 11:35:07.458 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 11:35:07.588 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:35:07.660 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:35:07.794 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:36:13.638 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_1/index_5954_seed_23956_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 11:36:18.526 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:36:38.219 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:36:38.228 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:36:38.228 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:37:46.183 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_1/index_5954_seed_23957_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 11:37:51.015 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:38:09.720 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:38:10.298 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:38:11.341 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:40:05.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.38s (1.91min) +2026-06-19 11:40:06.549 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:42:03.524 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:49:04.518 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 44, train_steps: 1760 +2026-06-19 11:49:05.114 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 11:49:05.259 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:49:05.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:49:05.502 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:50:11.551 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000044/group_1/index_1638_seed_30338_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 11:50:16.405 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:50:34.880 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:50:34.889 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:50:34.889 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:51:41.563 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000044/group_1/index_1638_seed_30339_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 11:51:45.893 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 11:52:04.843 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:52:06.491 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:52:07.520 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:54:01.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.69s (1.89min) +2026-06-19 11:54:01.910 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:56:01.236 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:03:03.870 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 45, train_steps: 1800 +2026-06-19 12:03:04.557 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 12:03:04.671 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:03:04.733 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:03:04.877 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:04:10.856 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000045/group_1/index_2469_seed_31820_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 12:04:14.808 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:04:33.334 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:04:33.340 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:04:33.340 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:05:40.605 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000045/group_1/index_2469_seed_31821_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 12:05:43.807 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:06:02.382 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:06:05.347 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:06:06.265 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:08:01.417 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.05s (1.92min) +2026-06-19 12:08:02.101 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:10:01.383 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:17:03.587 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 46, train_steps: 1840 +2026-06-19 12:17:04.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 12:17:04.370 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:17:04.417 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:17:04.513 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:18:10.534 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_1/index_3744_seed_22923_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 12:18:15.762 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:18:34.372 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:18:34.380 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:18:34.380 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:19:40.639 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_1/index_3744_seed_22924_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 12:19:45.536 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:20:04.105 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:20:05.333 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:20:06.532 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:22:01.733 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.17s (1.92min) +2026-06-19 12:22:02.443 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:24:01.450 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:31:02.928 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 47, train_steps: 1880 +2026-06-19 12:31:03.608 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 12:31:03.738 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:31:03.850 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:31:03.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:32:09.790 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000047/group_1/index_1448_seed_32260_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 12:32:13.991 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:32:32.375 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:32:32.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:32:32.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:33:39.400 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000047/group_1/index_1448_seed_32261_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 12:33:44.587 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:34:03.156 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:34:03.182 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:34:04.368 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:35:58.279 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-19 12:35:58.982 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:37:55.890 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:44:57.686 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 48, train_steps: 1920 +2026-06-19 12:44:58.355 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 12:44:58.443 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:44:58.483 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:44:58.637 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:46:04.632 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_1/index_5045_seed_7070_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 12:46:08.602 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:46:26.557 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:46:26.566 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:46:26.568 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:47:34.899 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_1/index_5045_seed_7071_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 12:47:40.216 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 12:47:58.651 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:47:58.930 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:47:59.970 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:49:54.620 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-19 12:49:55.343 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:51:52.290 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:58:52.089 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 49, train_steps: 1960 +2026-06-19 12:58:52.783 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 12:58:52.966 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:58:53.079 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:58:53.166 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:59:59.213 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_1/index_7729_seed_3224_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 13:00:03.263 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:00:21.178 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:00:21.187 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:00:21.189 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:01:30.354 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_1/index_7729_seed_3225_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 13:01:33.572 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:01:51.498 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:01:54.369 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:01:55.288 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:03:48.446 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.04s (1.88min) +2026-06-19 13:03:49.149 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:05:45.467 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:12:45.273 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 50, train_steps: 2000 +2026-06-19 13:14:21.276 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=50 (missing=0, unexpected=0) +2026-06-19 13:14:23.203 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 13:14:23.404 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:14:23.513 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:14:23.608 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:15:31.422 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_1/index_7573_seed_14178_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 13:15:36.723 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:15:53.466 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:15:53.475 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:15:53.476 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:17:00.939 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_1/index_7573_seed_14179_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 13:17:05.626 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:17:22.777 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:17:25.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:17:26.226 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:19:20.081 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.84s (1.90min) +2026-06-19 13:19:20.807 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:21:17.524 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:28:18.471 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 51, train_steps: 2040 +2026-06-19 13:28:19.186 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 13:28:19.357 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:28:19.494 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:28:19.578 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:29:26.318 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000051/group_1/index_1403_seed_11091_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 13:29:31.087 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:29:47.950 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:29:47.960 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:29:47.960 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:30:54.150 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000051/group_1/index_1403_seed_11092_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 13:30:58.928 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:31:16.207 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:31:18.028 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:31:19.200 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:33:13.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.83s (1.90min) +2026-06-19 13:33:13.741 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:35:13.230 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:42:13.618 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 52, train_steps: 2080 +2026-06-19 13:42:14.277 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 13:42:14.388 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:42:14.479 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:42:14.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:43:20.692 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_1/index_5822_seed_3352_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 13:43:25.436 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:43:42.527 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:43:42.536 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:43:42.536 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:44:50.241 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_1/index_5822_seed_3353_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 13:44:55.115 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:45:12.362 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:45:15.287 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:45:16.424 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:47:10.481 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.03s (1.90min) +2026-06-19 13:47:11.197 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:49:11.265 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:56:12.599 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 53, train_steps: 2120 +2026-06-19 13:56:13.303 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 13:56:13.443 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:56:13.561 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:56:13.673 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:57:19.503 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_1/index_4424_seed_17767_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 13:57:24.124 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:57:41.060 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:57:41.069 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:57:41.069 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:58:48.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_1/index_4424_seed_17768_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 13:58:53.367 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 13:59:10.635 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:59:13.130 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:59:14.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:01:08.798 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.41s (1.91min) +2026-06-19 14:01:09.529 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:03:10.473 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:10:12.162 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 54, train_steps: 2160 +2026-06-19 14:10:12.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 14:10:12.970 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:10:13.049 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:10:13.195 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:11:18.994 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_1/index_3397_seed_17918_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 14:11:22.757 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:11:41.090 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:11:41.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:11:41.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:12:47.921 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_1/index_3397_seed_17919_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 14:12:51.219 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:13:09.462 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:13:12.866 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:13:13.792 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:15:08.493 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.42s (1.91min) +2026-06-19 14:15:09.217 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:17:09.305 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:24:11.914 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 55, train_steps: 2200 +2026-06-19 14:24:12.602 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 14:24:12.741 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:24:12.818 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:24:12.923 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:25:19.021 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_1/index_3886_seed_31943_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 14:25:24.007 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:25:42.518 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:25:42.528 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:25:42.528 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:26:48.102 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_1/index_3886_seed_31944_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 14:26:52.873 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:27:11.423 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:27:12.753 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:27:13.928 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:29:09.888 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.82s (1.93min) +2026-06-19 14:29:10.608 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:31:10.764 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:38:12.112 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 56, train_steps: 2240 +2026-06-19 14:38:12.810 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 14:38:12.904 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:38:12.970 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:38:13.114 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:39:18.955 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_1/index_3297_seed_25007_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 14:39:23.016 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:39:39.653 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:39:39.663 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:39:39.663 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:40:48.922 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_1/index_3297_seed_25008_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 14:40:54.005 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:41:11.249 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:41:13.092 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:41:14.123 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:43:08.780 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.66s (1.91min) +2026-06-19 14:43:09.475 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:45:09.717 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:52:11.846 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 57, train_steps: 2280 +2026-06-19 14:52:12.527 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 14:52:12.689 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:52:12.825 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:52:12.937 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:53:18.585 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_1/index_6560_seed_22398_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 14:53:23.413 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:53:40.347 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:53:40.357 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:53:40.357 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:54:48.479 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_1/index_6560_seed_22399_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 14:54:52.496 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 14:55:09.339 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:55:13.057 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:55:13.959 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:57:08.883 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.79s (1.91min) +2026-06-19 14:57:09.597 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:59:09.879 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:06:12.287 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 58, train_steps: 2320 +2026-06-19 15:06:12.984 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 15:06:13.073 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:06:13.096 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:06:13.193 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:07:19.379 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_1/index_8819_seed_10700_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 15:07:22.713 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:07:39.437 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:07:39.447 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:07:39.447 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:08:48.512 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_1/index_8819_seed_10701_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 15:08:51.832 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:09:08.583 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:09:13.062 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:09:13.945 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:11:08.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.71s (1.91min) +2026-06-19 15:11:09.498 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:13:09.812 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:20:12.067 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 59, train_steps: 2360 +2026-06-19 15:20:12.738 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 15:20:12.828 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:20:12.891 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:20:13.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:21:19.121 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_1/index_778_seed_32999_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 15:21:24.126 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:21:41.033 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:21:41.041 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:21:41.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:22:47.784 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_1/index_778_seed_33000_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 15:22:51.132 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:23:08.253 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:23:12.372 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:23:13.410 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:25:09.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.28s (1.92min) +2026-06-19 15:25:09.742 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:27:10.366 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:34:12.073 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 60, train_steps: 2400 +2026-06-19 15:34:48.498 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=60 (missing=0, unexpected=0) +2026-06-19 15:34:50.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 15:34:50.822 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:34:50.929 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:34:51.008 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:35:58.922 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_1/index_9275_seed_6272_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 15:36:02.190 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:36:20.377 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:36:20.387 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:36:20.387 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:37:28.599 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_1/index_9275_seed_6273_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 15:37:32.040 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:37:48.998 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:37:53.529 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:37:54.613 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:39:47.707 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-19 15:39:48.404 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:41:45.234 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:48:44.584 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 61, train_steps: 2440 +2026-06-19 15:48:45.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 15:48:45.383 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:48:45.427 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:48:45.545 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:49:51.654 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_1/index_4023_seed_32938_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 15:49:56.733 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:50:13.623 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:50:13.632 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:50:13.632 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:51:23.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_1/index_4023_seed_32939_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 15:51:27.062 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 15:51:44.221 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:51:47.609 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:51:48.492 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:53:41.681 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.04s (1.88min) +2026-06-19 15:53:42.399 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:55:39.327 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:02:39.007 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 62, train_steps: 2480 +2026-06-19 16:02:39.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 16:02:39.776 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:02:39.794 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:02:39.860 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:03:45.768 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000062/group_1/index_9031_seed_1625_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 16:03:50.924 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:04:08.179 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:04:08.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:04:08.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:05:18.809 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000062/group_1/index_9031_seed_1626_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 16:05:23.479 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:05:40.799 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:05:44.132 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:05:45.172 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:07:38.761 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.59s (1.89min) +2026-06-19 16:07:39.471 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:09:38.854 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:16:39.319 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 63, train_steps: 2520 +2026-06-19 16:16:40.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 16:16:40.123 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:16:40.166 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:16:40.265 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:17:46.537 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_1/index_5147_seed_7579_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 16:17:50.220 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:18:07.318 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:18:07.327 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:18:07.327 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:19:19.864 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_1/index_5147_seed_7580_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 16:19:23.208 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:19:40.054 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:19:45.197 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:19:46.126 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:21:39.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.36s (1.89min) +2026-06-19 16:21:40.302 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:23:38.880 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:30:40.941 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 64, train_steps: 2560 +2026-06-19 16:30:41.626 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 16:30:41.799 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:30:41.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:30:41.995 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:31:48.235 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_1/index_5538_seed_1548_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 16:31:53.217 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:32:10.076 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:32:10.085 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:32:10.085 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:33:21.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_1/index_5538_seed_1549_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 16:33:25.087 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:33:41.741 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:33:45.471 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:33:46.476 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:35:40.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.63s (1.89min) +2026-06-19 16:35:40.844 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:37:39.397 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:44:41.301 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 65, train_steps: 2600 +2026-06-19 16:44:41.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 16:44:42.057 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:44:42.081 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:44:42.224 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:45:48.610 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_1/index_7647_seed_28570_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 16:45:51.885 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:46:08.671 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:46:08.680 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:46:08.680 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:47:22.894 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_1/index_7647_seed_28571_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 16:47:28.043 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 16:47:45.456 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:47:47.481 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:47:48.515 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:49:42.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.72s (1.90min) +2026-06-19 16:49:42.958 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:51:40.036 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:58:41.285 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 66, train_steps: 2640 +2026-06-19 16:58:41.997 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 16:58:42.126 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:58:42.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:58:42.382 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:59:47.844 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_1/index_9638_seed_31166_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 16:59:51.426 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:00:08.175 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:00:08.184 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:00:08.184 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:01:19.569 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_1/index_9638_seed_31167_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 17:01:24.721 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:01:42.003 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:01:43.599 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:01:44.633 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:03:37.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.34s (1.89min) +2026-06-19 17:03:38.681 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:05:38.383 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:12:40.124 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 67, train_steps: 2680 +2026-06-19 17:12:40.909 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 17:12:41.045 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:12:41.273 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:12:41.454 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:13:47.262 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_1/index_8578_seed_7528_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 17:13:51.151 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:14:09.427 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:14:09.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:14:09.438 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:15:20.546 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_1/index_8578_seed_7529_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 17:15:23.842 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:15:42.357 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:15:44.529 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:15:45.464 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:17:38.956 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 17:17:39.679 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:19:39.112 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:26:39.618 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 68, train_steps: 2720 +2026-06-19 17:26:40.214 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 17:26:40.320 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:26:40.451 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:26:40.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:27:46.733 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_1/index_4541_seed_13922_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 17:27:51.758 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:28:08.965 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:28:08.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:28:08.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:29:20.039 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_1/index_4541_seed_13923_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 17:29:23.799 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:29:41.140 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:29:43.886 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:29:44.796 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:31:38.707 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.76s (1.90min) +2026-06-19 17:31:39.420 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:33:38.458 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:40:38.372 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 69, train_steps: 2760 +2026-06-19 17:40:39.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 17:40:39.123 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:40:39.149 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:40:39.274 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:41:45.247 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000069/group_1/index_825_seed_4097_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 17:41:50.334 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:42:07.179 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:42:07.187 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:42:07.190 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:43:17.050 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000069/group_1/index_825_seed_4098_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 17:43:21.822 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:43:38.842 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:43:41.069 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:43:42.177 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:45:35.854 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.55s (1.89min) +2026-06-19 17:45:36.569 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:47:37.115 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:54:38.573 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 70, train_steps: 2800 +2026-06-19 17:55:14.794 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=70 (missing=0, unexpected=0) +2026-06-19 17:55:16.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 17:55:17.165 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:55:17.270 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:55:17.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:56:26.206 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_1/index_8100_seed_335_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 17:56:30.849 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:56:47.917 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:56:47.926 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:56:47.926 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:57:54.476 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_1/index_8100_seed_336_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 17:57:59.242 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 17:58:16.748 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:58:18.228 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:58:19.523 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:00:13.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.58s (1.89min) +2026-06-19 18:00:13.838 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:02:13.853 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:09:13.898 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 71, train_steps: 2840 +2026-06-19 18:09:14.615 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 18:09:14.718 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:09:14.782 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:09:14.942 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:10:20.620 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_1/index_5722_seed_22110_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 18:10:25.387 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:10:42.310 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:10:42.318 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:10:42.319 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:11:51.693 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_1/index_5722_seed_22111_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 18:11:55.929 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:12:13.098 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:12:16.021 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:12:17.051 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:14:12.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.67s (1.93min) +2026-06-19 18:14:13.460 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:16:12.695 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:23:13.827 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 72, train_steps: 2880 +2026-06-19 18:23:14.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 18:23:14.598 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:23:14.630 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:23:14.698 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:24:20.804 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_1/index_8497_seed_25703_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 18:24:25.690 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:24:42.911 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:24:42.920 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:24:42.922 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:25:51.117 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_1/index_8497_seed_25704_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 18:25:55.057 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:26:12.366 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:26:14.711 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:26:15.797 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:28:10.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.45s (1.91min) +2026-06-19 18:28:11.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:30:10.989 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:37:12.788 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 73, train_steps: 2920 +2026-06-19 18:37:13.455 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 18:37:13.544 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:37:13.576 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:37:13.679 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:38:19.630 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_1/index_4980_seed_26510_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 18:38:24.508 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:38:41.633 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:38:41.639 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:38:41.639 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:39:49.977 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_1/index_4980_seed_26511_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 18:39:53.792 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:40:10.935 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:40:13.249 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:40:14.141 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:42:09.573 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.16s (1.92min) +2026-06-19 18:42:10.285 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:44:10.447 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:51:12.751 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 74, train_steps: 2960 +2026-06-19 18:51:13.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 18:51:13.559 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:51:13.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:51:13.726 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:52:20.121 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_1/index_451_seed_11838_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 18:52:23.270 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:52:39.956 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:52:39.965 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:52:39.965 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:53:49.784 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_1/index_451_seed_11839_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 18:53:54.798 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 18:54:11.915 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:54:13.982 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:54:15.016 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:56:10.312 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.21s (1.92min) +2026-06-19 18:56:11.031 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:58:11.443 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:05:14.017 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 75, train_steps: 3000 +2026-06-19 19:06:23.229 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 19:06:23.318 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:06:23.367 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:06:23.521 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:07:29.470 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000075/group_1/index_7451_seed_5362_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 19:07:33.397 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:07:50.573 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:07:50.580 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:07:50.580 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:08:58.925 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000075/group_1/index_7451_seed_5363_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 19:09:03.770 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:09:21.455 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:09:24.209 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:09:25.253 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:11:18.397 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.14s (1.89min) +2026-06-19 19:11:19.100 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:13:16.254 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:20:16.311 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 76, train_steps: 3040 +2026-06-19 19:20:16.989 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 19:20:17.151 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:20:17.263 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:20:17.361 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:21:23.220 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000076/group_1/index_7191_seed_28139_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 19:21:27.937 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:21:45.007 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:21:45.016 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:21:45.016 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:22:53.555 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000076/group_1/index_7191_seed_28140_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 19:22:58.255 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:23:15.483 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:23:18.455 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:23:19.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:25:13.441 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.66s (1.89min) +2026-06-19 19:25:14.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:27:12.734 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:34:13.265 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 77, train_steps: 3080 +2026-06-19 19:34:13.951 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 19:34:14.132 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:34:14.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:34:14.353 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:35:20.093 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_1/index_3928_seed_17716_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 19:35:23.902 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:35:40.590 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:35:40.596 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:35:40.597 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:36:49.215 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_1/index_3928_seed_17717_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 19:36:52.448 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:37:09.424 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:37:13.073 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:37:14.011 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:39:08.734 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.60s (1.91min) +2026-06-19 19:39:09.440 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:41:09.769 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:48:12.648 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 78, train_steps: 3120 +2026-06-19 19:48:13.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 19:48:13.455 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:48:13.484 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:48:13.587 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:49:19.829 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_1/index_1908_seed_7362_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 19:49:22.951 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:49:39.670 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:49:39.680 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:49:39.680 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:50:49.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_1/index_1908_seed_7363_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 19:50:53.051 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 19:51:10.057 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:51:13.323 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:51:14.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:53:09.267 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-19 19:53:09.975 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:55:09.985 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:02:13.174 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 79, train_steps: 3160 +2026-06-19 20:02:13.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 20:02:13.995 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:02:14.071 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:02:14.221 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:03:20.189 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000079/group_1/index_689_seed_8183_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 20:03:25.143 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:03:42.248 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:03:42.257 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:03:42.257 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:04:49.735 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000079/group_1/index_689_seed_8184_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 20:04:54.538 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:05:11.787 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:05:13.092 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:05:14.129 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:07:09.660 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.53s (1.93min) +2026-06-19 20:07:10.385 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:09:11.033 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:16:13.942 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 80, train_steps: 3200 +2026-06-19 20:16:51.154 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=80 (missing=0, unexpected=0) +2026-06-19 20:16:53.092 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 20:16:53.292 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:16:53.419 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:16:53.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:18:01.925 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_1/index_4951_seed_6386_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 20:18:06.708 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:18:23.988 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:18:23.996 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:18:23.996 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:19:30.938 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_1/index_4951_seed_6387_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 20:19:35.838 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:19:53.370 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:19:54.752 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:19:55.787 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:21:49.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-19 20:21:50.363 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:23:47.274 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:30:47.431 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 81, train_steps: 3240 +2026-06-19 20:30:48.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 20:30:48.293 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:30:48.411 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:30:48.547 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:31:55.014 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_1/index_8381_seed_13652_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 20:31:59.857 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:32:16.717 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:32:16.726 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:32:16.726 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:33:25.210 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_1/index_8381_seed_13653_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 20:33:29.254 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:33:46.038 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:33:49.273 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:33:50.213 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:35:44.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.70s (1.89min) +2026-06-19 20:35:44.752 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:37:41.677 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:44:39.995 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 82, train_steps: 3280 +2026-06-19 20:44:40.696 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 20:44:40.785 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:44:40.810 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:44:40.896 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:45:46.980 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_1/index_2715_seed_6820_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 20:45:50.162 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:46:07.185 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:46:07.195 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:46:07.195 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:47:20.079 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_1/index_2715_seed_6821_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 20:47:25.002 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 20:47:42.584 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:47:44.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:47:45.158 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:49:38.578 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.41s (1.89min) +2026-06-19 20:49:39.275 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:51:38.423 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:58:38.991 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 83, train_steps: 3320 +2026-06-19 20:58:39.739 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 20:58:39.954 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:58:40.074 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:58:40.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:59:46.110 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_1/index_2018_seed_30899_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 20:59:50.112 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:00:07.180 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:00:07.189 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:00:07.189 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:01:19.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_1/index_2018_seed_30900_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 21:01:22.960 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:01:40.304 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:01:44.303 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:01:45.248 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:03:38.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.28s (1.89min) +2026-06-19 21:03:39.339 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:05:38.827 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:12:38.933 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 84, train_steps: 3360 +2026-06-19 21:12:39.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 21:12:39.738 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:12:39.770 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:12:39.883 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:13:45.932 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_1/index_176_seed_28409_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 21:13:49.219 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:14:06.261 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:14:06.271 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:14:06.272 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:15:19.634 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_1/index_176_seed_28410_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 21:15:24.598 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:15:42.059 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:15:43.791 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:15:44.841 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:17:38.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.31s (1.89min) +2026-06-19 21:17:38.864 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:19:38.522 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:26:39.500 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 85, train_steps: 3400 +2026-06-19 21:26:40.168 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 21:26:40.274 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:26:40.346 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:26:40.490 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:27:46.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_1/index_6754_seed_12823_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 21:27:51.666 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:28:08.840 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:28:08.850 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:28:08.851 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:29:20.054 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_1/index_6754_seed_12824_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 21:29:23.979 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:29:41.120 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:29:43.697 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:29:44.630 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:31:38.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-19 21:31:38.988 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:33:38.756 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:40:38.647 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 86, train_steps: 3440 +2026-06-19 21:40:39.367 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 21:40:39.513 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:40:39.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:40:39.761 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:41:45.573 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_1/index_6454_seed_30642_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 21:41:50.588 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:42:07.807 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:42:07.817 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:42:07.817 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:43:18.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_1/index_6454_seed_30643_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 21:43:23.516 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:43:40.883 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:43:42.867 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:43:43.904 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:45:37.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 21:45:37.987 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:47:37.709 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:54:39.047 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 87, train_steps: 3480 +2026-06-19 21:54:39.743 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 21:54:39.834 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:54:39.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:54:40.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:55:45.768 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000087/group_1/index_6140_seed_2930_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 21:55:49.682 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:56:06.465 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:56:06.474 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:56:06.474 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:57:19.000 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000087/group_1/index_6140_seed_2931_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 21:57:24.049 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 21:57:41.211 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:57:43.169 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:57:44.115 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:59:37.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.39s (1.89min) +2026-06-19 21:59:38.317 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:01:38.554 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:08:39.983 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 88, train_steps: 3520 +2026-06-19 22:08:40.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 22:08:40.759 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:08:40.782 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:08:40.837 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:09:46.696 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_1/index_9572_seed_20774_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 22:09:50.737 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:10:07.720 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:10:07.729 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:10:07.729 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:11:19.305 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_1/index_9572_seed_20775_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 22:11:22.526 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:11:39.841 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:11:43.755 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:11:44.655 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:13:38.389 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.57s (1.89min) +2026-06-19 22:13:39.103 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:15:38.303 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:22:38.698 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 89, train_steps: 3560 +2026-06-19 22:22:39.424 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 22:22:39.513 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:22:39.547 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:22:39.680 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:23:45.639 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_1/index_621_seed_30122_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 22:23:48.916 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:24:06.013 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:24:06.022 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:24:06.022 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:25:19.575 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_1/index_621_seed_30123_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 22:25:22.722 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:25:39.820 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:25:43.683 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:25:44.630 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:27:38.162 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 22:27:38.886 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:29:38.494 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:36:39.419 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 90, train_steps: 3600 +2026-06-19 22:37:15.490 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=90 (missing=0, unexpected=0) +2026-06-19 22:37:17.307 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 22:37:17.476 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:37:17.581 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:37:17.668 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:38:24.462 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_1/index_4617_seed_32332_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 22:38:29.469 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:38:45.872 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:38:45.881 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:38:45.881 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:39:55.090 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_1/index_4617_seed_32333_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 22:39:59.842 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:40:16.249 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:40:19.240 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:40:20.444 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:42:14.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.70s (1.89min) +2026-06-19 22:42:14.864 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:44:13.688 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:51:14.927 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 91, train_steps: 3640 +2026-06-19 22:51:15.700 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 22:51:15.845 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:51:15.973 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:51:16.087 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:52:22.302 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000091/group_1/index_8856_seed_19017_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 22:52:25.908 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:52:41.828 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:52:41.837 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:52:41.837 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:53:51.912 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000091/group_1/index_8856_seed_19018_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 22:53:56.717 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 22:54:13.101 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:54:15.937 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:54:16.996 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:56:12.671 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.67s (1.93min) +2026-06-19 22:56:13.388 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:58:13.112 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:05:14.303 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 92, train_steps: 3680 +2026-06-19 23:05:15.117 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 23:05:15.205 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:05:15.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:05:15.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:06:21.042 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000092/group_1/index_5293_seed_25767_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 23:06:24.535 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:06:41.300 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:06:41.306 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:06:41.306 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:07:51.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000092/group_1/index_5293_seed_25768_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 23:07:54.382 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:08:11.434 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:08:16.058 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:08:16.959 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:10:11.751 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.62s (1.91min) +2026-06-19 23:10:12.471 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:12:11.281 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:19:13.498 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 93, train_steps: 3720 +2026-06-19 23:19:14.152 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 23:19:14.263 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:19:14.341 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:19:14.486 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:20:20.681 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_1/index_2466_seed_33347_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 23:20:25.606 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:20:43.074 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:20:43.084 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:20:43.084 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:21:50.242 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_1/index_2466_seed_33348_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 23:21:54.972 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:22:12.454 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:22:14.354 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:22:15.582 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:24:10.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.33s (1.92min) +2026-06-19 23:24:11.707 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:26:11.285 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:33:13.939 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 94, train_steps: 3760 +2026-06-19 23:33:14.640 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 23:33:14.790 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:33:14.920 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:33:15.029 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:34:20.956 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_1/index_6943_seed_28345_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 23:34:24.948 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:34:42.419 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:34:42.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:34:42.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:35:51.728 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_1/index_6943_seed_28346_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 23:35:54.972 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:36:13.143 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:36:16.135 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:36:17.014 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:38:12.062 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.89s (1.91min) +2026-06-19 23:38:12.784 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:40:11.320 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:47:11.873 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 95, train_steps: 3800 +2026-06-19 23:47:12.588 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-19 23:47:12.681 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:47:12.725 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:47:12.826 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:48:18.971 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_1/index_6412_seed_4897_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-19 23:48:22.952 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:48:41.888 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:48:41.895 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:48:41.897 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:49:48.450 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_1/index_6412_seed_4898_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-19 23:49:53.472 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-19 23:50:11.994 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:50:12.004 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:50:13.235 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:52:08.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.82s (1.91min) +2026-06-19 23:52:08.791 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:54:07.677 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:01:09.774 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 96, train_steps: 3840 +2026-06-20 00:01:10.475 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 00:01:10.565 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:01:10.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:01:10.733 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:02:16.710 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_1/index_8491_seed_15682_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 00:02:21.057 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:02:38.546 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:02:38.555 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:02:38.555 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:03:47.487 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_1/index_8491_seed_15683_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 00:03:52.477 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:04:10.053 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:04:11.783 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:04:12.979 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:06:07.927 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 00:06:08.644 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:08:07.688 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:15:08.508 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 97, train_steps: 3880 +2026-06-20 00:15:09.220 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 00:15:09.352 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:15:09.420 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:15:09.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:16:15.336 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_1/index_6805_seed_26033_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 00:16:20.326 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:16:37.985 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:16:37.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:16:37.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:17:47.226 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_1/index_6805_seed_26034_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 00:17:51.104 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:18:08.683 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:18:11.237 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:18:12.379 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:20:07.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.20s (1.92min) +2026-06-20 00:20:08.400 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:22:07.126 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:29:07.006 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 98, train_steps: 3920 +2026-06-20 00:29:07.692 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 00:29:07.857 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:29:07.974 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:29:08.079 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:30:13.871 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000098/group_1/index_3759_seed_15971_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 00:30:17.619 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:30:36.765 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:30:36.775 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:30:36.775 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:31:46.373 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000098/group_1/index_3759_seed_15972_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 00:31:51.541 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:32:08.737 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:32:10.419 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:32:11.471 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:34:06.389 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 00:34:07.117 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:36:03.850 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:43:05.350 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 99, train_steps: 3960 +2026-06-20 00:43:06.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 00:43:06.152 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:43:06.240 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:43:06.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:44:12.247 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_1/index_9846_seed_5197_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 00:44:16.129 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:44:33.458 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:44:33.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:44:33.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:45:46.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_1/index_9846_seed_5198_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 00:45:49.300 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 00:46:06.137 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:46:09.898 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:46:10.811 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:48:05.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.75s (1.91min) +2026-06-20 00:48:06.440 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:50:03.397 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:57:04.809 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 100, train_steps: 4000 +2026-06-20 00:58:47.487 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=100 (missing=0, unexpected=0) +2026-06-20 00:58:49.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 00:58:49.655 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:58:49.671 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:58:49.737 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:59:57.229 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_1/index_4782_seed_4756_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 01:00:00.339 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:00:16.247 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:00:16.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:00:16.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:01:26.728 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_1/index_4782_seed_4757_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 01:01:31.573 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:01:47.869 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:01:50.303 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:01:51.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:03:44.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-20 01:03:45.434 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:05:42.220 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:12:39.870 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 101, train_steps: 4040 +2026-06-20 01:12:40.567 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 01:12:40.676 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:12:40.757 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:12:40.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:13:47.331 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_1/index_3162_seed_4006_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 01:13:50.889 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:14:07.236 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:14:07.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:14:07.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:15:20.127 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_1/index_3162_seed_4007_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 01:15:23.239 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:15:39.818 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:15:43.498 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:15:44.410 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:17:38.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.62s (1.89min) +2026-06-20 01:17:38.931 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:19:38.324 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:26:37.813 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 102, train_steps: 4080 +2026-06-20 01:26:38.504 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 01:26:38.615 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:26:38.705 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:26:38.851 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:27:44.266 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_1/index_996_seed_23525_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 01:27:47.170 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:28:03.469 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:28:03.479 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:28:03.479 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:29:17.104 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_1/index_996_seed_23526_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 01:29:20.575 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:29:36.705 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:29:41.182 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:29:42.379 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:31:36.383 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-20 01:31:37.098 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:33:37.227 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:40:38.543 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 103, train_steps: 4120 +2026-06-20 01:40:39.244 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 01:40:39.334 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:40:39.362 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:40:39.429 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:41:45.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000103/group_1/index_4506_seed_19524_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 01:41:50.329 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:42:06.931 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:42:06.940 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:42:06.940 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:43:19.152 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000103/group_1/index_4506_seed_19525_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 01:43:23.728 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:43:39.786 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:43:43.334 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:43:44.383 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:45:37.767 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-20 01:45:38.487 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:47:38.265 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:54:37.793 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 104, train_steps: 4160 +2026-06-20 01:54:38.707 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 01:54:38.837 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:54:38.981 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:54:39.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:55:45.142 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_1/index_4619_seed_21852_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 01:55:48.555 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:56:04.962 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:56:04.971 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:56:04.971 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:57:18.783 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_1/index_4619_seed_21853_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 01:57:21.831 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 01:57:37.961 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:57:42.889 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:57:43.796 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:59:37.314 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-20 01:59:38.031 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:01:38.019 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:08:38.453 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 105, train_steps: 4200 +2026-06-20 02:08:39.147 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 02:08:39.337 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:08:39.465 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:08:39.544 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:09:45.269 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000105/group_1/index_7533_seed_8342_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 02:09:50.026 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:10:06.769 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:10:06.778 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:10:06.779 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:11:18.928 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000105/group_1/index_7533_seed_8343_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 02:11:22.528 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:11:38.871 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:11:43.945 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:11:44.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:13:38.216 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 02:13:38.925 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:15:38.268 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:22:37.848 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 106, train_steps: 4240 +2026-06-20 02:22:38.566 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 02:22:38.713 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:22:38.831 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:22:38.934 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:23:44.804 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000106/group_1/index_7205_seed_17704_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 02:23:47.809 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:24:04.429 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:24:04.438 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:24:04.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:25:16.750 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000106/group_1/index_7205_seed_17705_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 02:25:20.494 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:25:37.106 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:25:40.846 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:25:41.780 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:27:35.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.90s (1.90min) +2026-06-20 02:27:36.583 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:29:36.811 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:36:37.098 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 107, train_steps: 4280 +2026-06-20 02:36:37.804 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 02:36:37.927 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:36:38.060 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:36:38.206 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:37:43.914 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_1/index_9475_seed_5499_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 02:37:47.589 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:38:03.931 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:38:03.940 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:38:03.940 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:39:15.149 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_1/index_9475_seed_5500_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 02:39:18.474 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:39:36.112 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:39:39.229 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:39:40.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:41:35.347 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 02:41:36.072 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:43:35.145 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:50:35.005 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 108, train_steps: 4320 +2026-06-20 02:50:35.728 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 02:50:35.856 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:50:35.966 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:50:36.083 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:51:41.998 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_1/index_5958_seed_1822_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 02:51:44.985 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:52:01.492 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:52:01.502 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:52:01.502 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:53:13.086 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_1/index_5958_seed_1823_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 02:53:16.094 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 02:53:32.432 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:53:37.573 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:53:38.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:55:33.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.45s (1.91min) +2026-06-20 02:55:33.972 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:57:30.647 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:04:30.701 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 109, train_steps: 4360 +2026-06-20 03:04:31.406 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 03:04:31.534 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:04:31.644 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:04:31.778 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:05:37.355 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000109/group_1/index_8421_seed_2761_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 03:05:40.414 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:05:56.312 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:05:56.322 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:05:56.322 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:07:09.705 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000109/group_1/index_8421_seed_2762_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 03:07:12.745 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:07:29.163 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:07:33.406 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:07:34.308 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:09:29.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-20 03:09:30.014 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:11:26.574 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:18:25.333 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 110, train_steps: 4400 +2026-06-20 03:19:01.062 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=110 (missing=0, unexpected=0) +2026-06-20 03:19:03.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 03:19:03.424 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:19:03.488 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:19:03.606 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:20:11.702 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_1/index_3424_seed_1010_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 03:20:14.866 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:20:30.614 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:20:30.622 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:20:30.622 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:21:43.264 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_1/index_3424_seed_1011_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 03:21:48.060 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:22:03.625 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:22:07.524 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:22:08.554 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:24:03.966 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.41s (1.92min) +2026-06-20 03:24:04.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:26:02.473 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:33:05.397 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 111, train_steps: 4440 +2026-06-20 03:33:06.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 03:33:06.240 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:33:06.371 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:33:06.453 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:34:12.646 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_1/index_1171_seed_15417_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 03:34:17.168 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:34:33.096 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:34:33.106 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:34:33.106 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:35:46.534 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_1/index_1171_seed_15418_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 03:35:49.753 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:36:05.365 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:36:10.098 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:36:11.169 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:38:06.021 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 03:38:06.732 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:40:04.482 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:47:04.319 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 112, train_steps: 4480 +2026-06-20 03:47:05.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 03:47:05.176 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:47:05.285 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:47:05.386 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:48:10.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_1/index_7032_seed_33172_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 03:48:15.609 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:48:30.744 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:48:30.752 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:48:30.752 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:49:40.731 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_1/index_7032_seed_33173_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 03:49:45.265 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 03:50:00.640 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:50:04.983 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:50:06.316 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:52:00.943 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-20 03:52:01.658 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:54:00.964 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:01:04.010 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 113, train_steps: 4520 +2026-06-20 04:01:04.875 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 04:01:05.040 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:01:05.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:01:05.253 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:02:11.070 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000113/group_1/index_6461_seed_13859_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 04:02:14.352 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:02:29.535 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:02:29.544 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:02:29.544 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:03:42.089 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000113/group_1/index_6461_seed_13860_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 04:03:46.790 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:04:02.642 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:04:06.611 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:04:07.660 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:06:02.814 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.15s (1.92min) +2026-06-20 04:06:03.532 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:08:01.582 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:15:03.783 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 114, train_steps: 4560 +2026-06-20 04:15:04.474 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 04:15:04.626 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:15:04.777 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:15:04.894 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:16:10.744 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000114/group_1/index_5489_seed_5822_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 04:16:15.287 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:16:30.895 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:16:30.905 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:16:30.905 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:17:40.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000114/group_1/index_5489_seed_5823_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 04:17:44.917 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:18:00.293 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:18:04.152 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:18:05.367 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:20:00.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.72s (1.91min) +2026-06-20 04:20:00.895 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:22:00.100 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:29:01.469 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 115, train_steps: 4600 +2026-06-20 04:29:02.193 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 04:29:02.403 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:29:02.531 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:29:02.610 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:30:08.163 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_1/index_7123_seed_21059_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 04:30:11.655 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:30:26.652 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:30:26.658 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:30:26.658 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:31:38.928 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_1/index_7123_seed_21060_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 04:31:41.883 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:31:57.537 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:32:02.494 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:32:03.402 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:33:58.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.33s (1.91min) +2026-06-20 04:33:58.776 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:35:55.624 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:42:56.472 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 116, train_steps: 4640 +2026-06-20 04:42:57.158 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 04:42:57.251 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:42:57.343 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:42:57.494 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:44:03.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_1/index_3099_seed_26804_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 04:44:06.580 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:44:21.739 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:44:21.746 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:44:21.746 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:45:33.890 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_1/index_3099_seed_26805_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 04:45:36.900 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:45:52.302 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:45:58.380 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:45:59.287 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:47:53.698 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.25s (1.90min) +2026-06-20 04:47:54.407 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:49:50.889 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:56:50.232 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 117, train_steps: 4680 +2026-06-20 04:56:50.953 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 04:56:51.041 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:56:51.067 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:56:51.142 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:57:57.095 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_1/index_3254_seed_14818_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 04:58:00.045 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:58:15.341 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:58:15.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:58:15.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:59:28.464 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_1/index_3254_seed_14819_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 04:59:33.242 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 04:59:48.911 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:59:51.798 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:59:52.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:01:46.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.49s (1.89min) +2026-06-20 05:01:47.078 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:03:44.148 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:10:42.593 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 118, train_steps: 4720 +2026-06-20 05:10:43.295 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 05:10:43.388 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:10:43.454 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:10:43.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:11:49.725 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000118/group_1/index_458_seed_23056_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 05:11:54.233 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:12:09.700 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:12:09.709 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:12:09.709 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:13:22.742 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000118/group_1/index_458_seed_23057_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 05:13:26.194 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:13:43.368 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:13:46.957 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:13:47.940 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:15:41.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.26s (1.89min) +2026-06-20 05:15:42.001 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:17:38.994 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:24:37.976 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 119, train_steps: 4760 +2026-06-20 05:24:38.623 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 05:24:38.749 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:24:38.810 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:24:38.936 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:25:44.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_1/index_1184_seed_20628_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 05:25:49.513 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:26:05.178 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:26:05.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:26:05.189 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:27:15.017 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_1/index_1184_seed_20629_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 05:27:18.486 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:27:35.440 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:27:40.119 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:27:41.333 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:29:34.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-20 05:29:35.559 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:31:33.105 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:38:32.150 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 120, train_steps: 4800 +2026-06-20 05:39:09.123 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=120 (missing=0, unexpected=0) +2026-06-20 05:39:10.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 05:39:11.060 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:39:11.161 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:39:11.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:40:19.082 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000120/group_1/index_2238_seed_27645_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 05:40:22.580 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:40:38.650 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:40:38.660 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:40:38.660 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:41:48.540 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000120/group_1/index_2238_seed_27646_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 05:41:51.427 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:42:07.022 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:42:12.566 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:42:13.460 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:44:08.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.96s (1.92min) +2026-06-20 05:44:09.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:46:10.086 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:53:12.062 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 121, train_steps: 4840 +2026-06-20 05:53:12.777 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 05:53:12.952 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:53:13.080 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:53:13.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:54:19.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_1/index_2191_seed_20093_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 05:54:22.004 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:54:37.850 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:54:37.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:54:37.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:55:48.452 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_1/index_2191_seed_20094_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 05:55:51.407 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 05:56:07.208 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:56:13.155 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:56:14.108 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:58:09.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.97s (1.92min) +2026-06-20 05:58:10.076 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:00:10.575 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:07:11.959 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 122, train_steps: 4880 +2026-06-20 06:07:12.650 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 06:07:12.777 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:07:12.865 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:07:12.969 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:08:18.554 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_1/index_2102_seed_9822_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 06:08:23.317 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:08:38.928 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:08:38.937 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:08:38.939 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:09:48.722 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_1/index_2102_seed_9823_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 06:09:53.302 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:10:09.097 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:10:13.195 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:10:14.236 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:12:08.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.25s (1.90min) +2026-06-20 06:12:09.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:14:10.580 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:21:11.867 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 123, train_steps: 4920 +2026-06-20 06:21:12.561 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 06:21:12.652 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:21:12.711 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:21:12.829 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:22:18.802 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_1/index_807_seed_31584_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 06:22:22.210 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:22:37.511 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:22:37.520 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:22:37.520 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:23:48.008 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_1/index_807_seed_31585_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 06:23:50.872 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:24:06.356 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:24:12.071 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:24:13.180 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:26:08.788 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.47s (1.92min) +2026-06-20 06:26:09.511 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:28:09.834 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:35:11.849 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 124, train_steps: 4960 +2026-06-20 06:35:12.538 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 06:35:12.629 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:35:12.655 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:35:12.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:36:18.461 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_1/index_4725_seed_18324_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 06:36:23.143 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:36:38.908 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:36:38.915 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:36:38.915 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:37:48.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_1/index_4725_seed_18325_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 06:37:52.693 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:38:08.753 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:38:12.129 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:38:13.253 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:40:08.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.44s (1.92min) +2026-06-20 06:40:09.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:42:09.772 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:49:11.721 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 125, train_steps: 5000 +2026-06-20 06:50:15.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 06:50:15.322 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:50:15.443 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:50:15.584 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:51:21.544 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_1/index_911_seed_5351_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 06:51:26.207 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:51:42.189 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:51:42.198 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:51:42.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:52:51.079 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_1/index_911_seed_5352_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 06:52:54.469 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 06:53:10.351 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:53:15.926 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:53:16.838 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:55:11.653 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.68s (1.91min) +2026-06-20 06:55:12.372 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:57:11.763 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:04:11.924 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 126, train_steps: 5040 +2026-06-20 07:04:12.630 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 07:04:12.737 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:04:12.790 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:04:12.904 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:05:19.152 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_1/index_6424_seed_29911_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 07:05:22.158 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:05:37.559 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:05:37.568 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:05:37.568 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:06:49.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_1/index_6424_seed_29912_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 07:06:54.156 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:07:10.182 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:07:13.988 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:07:15.267 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:09:10.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.08s (1.92min) +2026-06-20 07:09:11.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:11:11.165 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:18:11.933 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 127, train_steps: 5080 +2026-06-20 07:18:12.634 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 07:18:12.781 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:18:12.912 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:18:13.009 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:19:18.874 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_1/index_4986_seed_21729_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 07:19:23.626 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:19:39.447 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:19:39.456 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:19:39.456 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:20:48.745 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_1/index_4986_seed_21730_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 07:20:53.402 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:21:09.488 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:21:13.050 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:21:14.074 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:23:09.600 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.50s (1.93min) +2026-06-20 07:23:10.327 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:25:10.158 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:32:11.171 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 128, train_steps: 5120 +2026-06-20 07:32:11.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 07:32:12.009 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:32:12.065 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:32:12.185 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:33:17.960 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_1/index_4968_seed_20413_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 07:33:22.731 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:33:38.470 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:33:38.479 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:33:38.479 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:34:48.939 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_1/index_4968_seed_20414_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 07:34:53.643 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:35:09.625 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:35:13.015 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:35:14.047 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:37:08.641 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.56s (1.91min) +2026-06-20 07:37:09.373 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:39:09.930 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:46:11.999 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 129, train_steps: 5160 +2026-06-20 07:46:12.711 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 07:46:12.802 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:46:12.826 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:46:12.929 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:47:19.614 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_1/index_1776_seed_1892_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 07:47:24.268 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:47:40.051 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:47:40.063 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:47:40.064 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:48:48.977 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_1/index_1776_seed_1893_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 07:48:53.766 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 07:49:09.827 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:49:13.070 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:49:14.105 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:51:09.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-20 07:51:09.758 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:53:10.304 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:00:11.782 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 130, train_steps: 5200 +2026-06-20 08:00:49.335 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=130 (missing=0, unexpected=0) +2026-06-20 08:00:51.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 08:00:51.407 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:00:51.504 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:00:51.602 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:02:00.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_1/index_7500_seed_4085_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 08:02:03.758 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:02:20.383 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:02:20.392 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:02:20.393 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:03:28.838 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_1/index_7500_seed_4086_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 08:03:31.827 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:03:48.015 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:03:52.509 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:03:53.405 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:05:47.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.72s (1.90min) +2026-06-20 08:05:48.018 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:07:44.844 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:14:45.324 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 131, train_steps: 5240 +2026-06-20 08:14:46.025 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 08:14:46.133 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:14:46.212 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:14:46.354 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:15:52.337 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_1/index_4117_seed_11490_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 08:15:57.137 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:16:13.037 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:16:13.046 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:16:13.047 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:17:23.381 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_1/index_4117_seed_11491_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 08:17:28.099 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:17:44.348 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:17:46.989 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:17:48.026 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:19:41.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.38s (1.89min) +2026-06-20 08:19:42.138 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:21:39.469 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:28:37.831 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 132, train_steps: 5280 +2026-06-20 08:28:38.522 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 08:28:38.700 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:28:38.837 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:28:38.898 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:29:45.164 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000132/group_1/index_226_seed_26181_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 08:29:49.969 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:30:05.790 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:30:05.801 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:30:05.803 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:31:15.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000132/group_1/index_226_seed_26182_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 08:31:20.389 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:31:36.723 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:31:40.849 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:31:41.908 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:33:35.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.44s (1.89min) +2026-06-20 08:33:36.119 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:35:35.263 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:42:34.332 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 133, train_steps: 5320 +2026-06-20 08:42:35.019 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 08:42:35.182 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:42:35.289 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:42:35.369 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:43:41.511 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_1/index_9702_seed_5269_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 08:43:45.052 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:44:01.015 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:44:01.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:44:01.025 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:45:11.843 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_1/index_9702_seed_5270_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 08:45:16.645 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:45:33.282 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:45:35.887 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:45:36.925 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:47:32.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.77s (1.93min) +2026-06-20 08:47:33.450 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:49:30.692 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:56:29.434 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 134, train_steps: 5360 +2026-06-20 08:56:30.114 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 08:56:30.280 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:56:30.377 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:56:30.477 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:57:36.222 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000134/group_1/index_1546_seed_4752_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 08:57:39.991 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:57:56.026 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:57:56.035 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:57:56.035 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:59:07.541 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000134/group_1/index_1546_seed_4753_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 08:59:10.513 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 08:59:26.615 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:59:32.040 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:59:32.940 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:01:26.319 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 09:01:27.048 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:03:23.805 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:10:22.990 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 135, train_steps: 5400 +2026-06-20 09:10:23.677 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 09:10:23.831 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:10:23.961 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:10:24.042 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:11:29.868 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000135/group_1/index_9276_seed_4794_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 09:11:32.850 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:11:48.418 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:11:48.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:11:48.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:13:00.395 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000135/group_1/index_9276_seed_4795_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 09:13:03.493 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:13:20.352 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:13:23.932 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:13:24.816 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:15:18.588 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.61s (1.89min) +2026-06-20 09:15:19.310 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:17:17.101 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:24:16.804 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 136, train_steps: 5440 +2026-06-20 09:24:17.532 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 09:24:17.696 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:24:17.792 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:24:17.890 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:25:23.759 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_1/index_7682_seed_20505_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 09:25:28.490 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:25:44.642 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:25:44.649 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:25:44.649 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:26:54.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_1/index_7682_seed_20506_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 09:26:57.666 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:27:13.912 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:27:18.656 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:27:19.787 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:29:12.958 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.15s (1.89min) +2026-06-20 09:29:13.686 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:31:13.177 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:38:13.221 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 137, train_steps: 5480 +2026-06-20 09:38:13.937 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 09:38:14.100 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:38:14.191 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:38:14.290 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:39:20.426 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000137/group_1/index_1231_seed_6030_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 09:39:23.629 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:39:39.770 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:39:39.780 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:39:39.780 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:40:50.302 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000137/group_1/index_1231_seed_6031_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 09:40:53.333 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:41:09.343 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:41:14.371 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:41:15.435 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:43:09.522 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-20 09:43:10.250 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:45:10.504 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:52:12.151 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 138, train_steps: 5520 +2026-06-20 09:52:12.866 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 09:52:12.991 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:52:13.027 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:52:13.112 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:53:19.014 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_1/index_7260_seed_18687_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 09:53:21.976 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:53:37.548 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:53:37.557 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:53:37.558 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:54:49.483 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_1/index_7260_seed_18688_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 09:54:52.515 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 09:55:08.359 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:55:13.383 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:55:14.490 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:57:10.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.46s (1.92min) +2026-06-20 09:57:10.921 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:59:11.151 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:06:12.678 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 139, train_steps: 5560 +2026-06-20 10:06:13.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 10:06:13.561 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:06:13.682 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:06:13.761 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:07:19.775 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_1/index_1100_seed_6281_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 10:07:24.599 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:07:40.461 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:07:40.470 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:07:40.470 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:08:49.990 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_1/index_1100_seed_6282_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 10:08:54.712 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:09:10.701 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:09:13.264 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:09:14.209 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:11:08.674 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.20s (1.90min) +2026-06-20 10:11:09.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:13:09.781 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:20:12.682 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 140, train_steps: 5600 +2026-06-20 10:20:50.191 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=140 (missing=0, unexpected=0) +2026-06-20 10:20:51.812 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 10:20:51.976 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:20:52.067 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:20:52.165 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:21:59.754 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000140/group_1/index_6455_seed_21264_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 10:22:02.701 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:22:19.986 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:22:19.996 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:22:19.999 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:23:28.917 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000140/group_1/index_6455_seed_21265_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 10:23:33.841 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:23:50.417 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:23:53.040 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:23:54.078 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:25:47.600 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-20 10:25:48.314 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:27:45.364 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:34:44.004 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 141, train_steps: 5640 +2026-06-20 10:34:44.689 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 10:34:44.867 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:34:44.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:34:45.067 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:35:50.810 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_1/index_2361_seed_15755_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 10:35:55.432 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:36:11.281 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:36:11.291 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:36:11.291 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:37:22.874 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_1/index_2361_seed_15756_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 10:37:27.543 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:37:43.873 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:37:46.846 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:37:47.890 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:39:41.415 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.51s (1.89min) +2026-06-20 10:39:42.148 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:41:39.461 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:48:38.475 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 142, train_steps: 5680 +2026-06-20 10:48:39.159 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 10:48:39.329 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:48:39.438 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:48:39.518 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:49:45.456 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000142/group_1/index_6579_seed_2455_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 10:49:48.943 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:50:05.162 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:50:05.169 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:50:05.169 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:51:19.214 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000142/group_1/index_6579_seed_2456_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 10:51:22.746 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 10:51:38.892 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:51:43.153 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:51:44.105 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:53:37.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-20 10:53:37.851 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:55:37.993 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:02:37.925 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 143, train_steps: 5720 +2026-06-20 11:02:38.634 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 11:02:38.788 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:02:38.902 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:02:38.981 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:03:44.870 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000143/group_1/index_3359_seed_4892_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 11:03:47.908 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:04:03.779 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:04:03.786 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:04:03.786 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:05:18.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000143/group_1/index_3359_seed_4893_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 11:05:21.758 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:05:37.750 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:05:42.169 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:05:43.239 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:07:36.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 11:07:37.313 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:09:36.843 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:16:37.099 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 144, train_steps: 5760 +2026-06-20 11:16:37.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 11:16:37.928 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:16:38.022 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:16:38.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:17:43.699 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_1/index_4591_seed_20846_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 11:17:48.539 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:18:04.763 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:18:04.773 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:18:04.773 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:19:13.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_1/index_4591_seed_20847_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 11:19:16.791 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:19:33.020 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:19:38.707 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:19:39.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:21:33.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.47s (1.89min) +2026-06-20 11:21:34.087 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:23:30.884 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:30:30.403 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 145, train_steps: 5800 +2026-06-20 11:30:31.098 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 11:30:31.262 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:30:31.380 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:30:31.461 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:31:37.194 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000145/group_1/index_8183_seed_23112_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 11:31:40.164 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:31:56.185 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:31:56.195 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:31:56.195 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:33:08.915 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000145/group_1/index_8183_seed_23113_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 11:33:13.841 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:33:30.062 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:33:33.207 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:33:34.270 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:35:28.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.89s (1.90min) +2026-06-20 11:35:28.900 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:37:25.686 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:44:23.920 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 146, train_steps: 5840 +2026-06-20 11:44:24.631 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 11:44:24.822 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:44:24.956 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:44:25.015 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:45:30.522 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_1/index_4905_seed_469_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 11:45:35.119 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:45:51.006 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:45:51.015 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:45:51.016 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:46:59.075 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_1/index_4905_seed_470_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 11:47:03.748 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:47:20.391 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:47:23.247 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:47:24.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:49:19.133 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-20 11:49:19.856 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:51:17.040 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:58:15.993 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 147, train_steps: 5880 +2026-06-20 11:58:16.620 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 11:58:16.809 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:58:16.934 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:58:16.998 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:59:23.426 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_1/index_7887_seed_1943_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 11:59:26.827 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 11:59:42.931 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:59:42.940 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:59:42.941 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:00:52.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_1/index_7887_seed_1944_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 12:00:55.863 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:01:12.181 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:01:16.232 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:01:17.204 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:03:12.586 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.29s (1.92min) +2026-06-20 12:03:13.293 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:05:12.905 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:12:12.708 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 148, train_steps: 5920 +2026-06-20 12:12:13.404 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 12:12:13.526 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:12:13.602 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:12:13.715 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:13:19.997 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000148/group_1/index_5635_seed_25657_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 12:13:24.890 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:13:41.209 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:13:41.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:13:41.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:14:49.529 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000148/group_1/index_5635_seed_25658_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 12:14:53.034 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:15:09.228 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:15:13.144 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:15:14.040 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:17:09.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.69s (1.91min) +2026-06-20 12:17:09.789 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:19:10.588 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:26:11.388 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 149, train_steps: 5960 +2026-06-20 12:26:12.135 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 12:26:12.279 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:26:12.363 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:26:12.465 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:27:18.367 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000149/group_1/index_9665_seed_30955_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 12:27:23.282 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:27:39.143 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:27:39.154 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:27:39.154 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:28:48.552 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000149/group_1/index_9665_seed_30956_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 12:28:52.020 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:29:08.081 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:29:13.029 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:29:13.982 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:31:08.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.71s (1.91min) +2026-06-20 12:31:09.533 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:33:09.832 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:40:10.990 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 150, train_steps: 6000 +2026-06-20 12:41:54.215 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=150 (missing=0, unexpected=0) +2026-06-20 12:41:56.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 12:41:56.244 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:41:56.339 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:41:56.461 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:43:04.129 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_1/index_4542_seed_16157_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 12:43:08.951 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:43:24.468 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:43:24.478 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:43:24.478 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:44:32.706 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_1/index_4542_seed_16158_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 12:44:37.237 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:44:52.792 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:44:59.743 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:45:00.802 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:46:55.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.70s (1.91min) +2026-06-20 12:46:56.236 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:48:52.989 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:55:51.604 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 151, train_steps: 6040 +2026-06-20 12:55:52.309 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 12:55:52.519 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:55:52.645 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:55:52.717 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:56:58.341 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_1/index_2484_seed_15246_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 12:57:01.644 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:57:17.003 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:57:17.009 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:57:17.010 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:58:27.400 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_1/index_2484_seed_15247_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 12:58:32.109 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 12:58:48.001 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:58:50.720 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:58:51.767 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:00:46.189 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.41s (1.91min) +2026-06-20 13:00:46.908 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:02:43.653 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:09:41.903 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 152, train_steps: 6080 +2026-06-20 13:09:42.599 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 13:09:42.788 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:09:42.919 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:09:43.013 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:10:49.039 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_1/index_848_seed_2290_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 13:10:52.475 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:11:07.931 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:11:07.937 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:11:07.938 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:12:20.654 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_1/index_848_seed_2291_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 13:12:23.670 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:12:39.379 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:12:43.910 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:12:44.858 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:14:38.763 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-20 13:14:39.492 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:16:38.550 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:23:37.214 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 153, train_steps: 6120 +2026-06-20 13:23:38.095 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 13:23:38.272 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:23:38.394 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:23:38.483 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:24:45.209 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_1/index_1858_seed_10774_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 13:24:49.973 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:25:07.019 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:25:07.028 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:25:07.029 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:26:15.131 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_1/index_1858_seed_10775_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 13:26:18.560 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:26:36.823 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:26:40.127 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:26:41.218 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:28:34.747 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.39s (1.89min) +2026-06-20 13:28:35.473 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:30:32.907 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:37:31.698 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 154, train_steps: 6160 +2026-06-20 13:37:32.394 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 13:37:32.564 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:37:32.655 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:37:32.737 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:38:38.545 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_1/index_3760_seed_4112_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 13:38:41.502 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:38:57.075 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:38:57.084 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:38:57.084 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:40:09.503 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_1/index_3760_seed_4113_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 13:40:14.355 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:40:30.633 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:40:33.788 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:40:34.852 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:42:29.329 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.47s (1.91min) +2026-06-20 13:42:30.062 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:44:27.161 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:51:24.982 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 155, train_steps: 6200 +2026-06-20 13:51:25.702 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 13:51:25.861 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:51:25.939 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:51:26.074 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:52:31.876 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_1/index_6583_seed_15398_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 13:52:36.449 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:52:52.400 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:52:52.410 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:52:52.411 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:54:01.104 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_1/index_6583_seed_15399_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 13:54:04.550 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 13:54:20.420 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:54:24.447 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:54:25.368 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:56:19.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.63s (1.89min) +2026-06-20 13:56:19.859 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:58:16.751 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:05:15.351 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 156, train_steps: 6240 +2026-06-20 14:05:16.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 14:05:16.136 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:05:16.185 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:05:16.254 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:06:22.825 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_1/index_7733_seed_29647_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 14:06:27.565 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:06:43.444 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:06:43.455 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:06:43.455 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:07:51.801 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_1/index_7733_seed_29648_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 14:07:55.185 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:08:11.053 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:08:15.931 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:08:16.847 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:10:11.722 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.74s (1.91min) +2026-06-20 14:10:12.457 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:12:11.104 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:19:12.246 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 157, train_steps: 6280 +2026-06-20 14:19:12.925 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 14:19:13.099 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:19:13.218 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:19:13.296 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:20:19.203 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000157/group_1/index_7014_seed_18619_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 14:20:22.248 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:20:38.260 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:20:38.269 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:20:38.269 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:21:47.997 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000157/group_1/index_7014_seed_18620_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 14:21:50.990 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:22:06.729 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:22:12.398 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:22:13.421 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:24:08.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.22s (1.92min) +2026-06-20 14:24:09.689 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:26:09.904 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:33:10.828 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 158, train_steps: 6320 +2026-06-20 14:33:11.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 14:33:11.680 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:33:11.778 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:33:11.890 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:34:17.684 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_1/index_1491_seed_4145_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 14:34:20.593 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:34:36.096 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:34:36.106 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:34:36.106 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:35:47.084 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_1/index_1491_seed_4146_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 14:35:51.831 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:36:07.800 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:36:10.935 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:36:12.077 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:38:07.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.88s (1.91min) +2026-06-20 14:38:07.784 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:40:05.258 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:47:04.572 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 159, train_steps: 6360 +2026-06-20 14:47:05.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 14:47:05.465 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:47:05.578 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:47:05.680 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:48:11.578 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_1/index_6972_seed_16233_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 14:48:16.224 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:48:31.940 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:48:31.949 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:48:31.950 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:49:40.220 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_1/index_6972_seed_16234_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 14:49:43.579 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 14:49:59.169 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:50:04.048 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:50:05.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:51:59.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.86s (1.90min) +2026-06-20 14:51:59.929 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:53:57.321 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:00:59.184 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 160, train_steps: 6400 +2026-06-20 15:01:36.740 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=160 (missing=0, unexpected=0) +2026-06-20 15:01:39.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 15:01:39.387 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:01:39.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:01:39.499 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:02:46.722 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_1/index_2899_seed_6161_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 15:02:49.737 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:03:05.315 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:03:05.325 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:03:05.326 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:04:18.953 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_1/index_2899_seed_6162_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 15:04:21.973 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:04:37.868 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:04:42.879 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:04:43.780 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:06:37.408 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.48s (1.89min) +2026-06-20 15:06:38.136 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:08:38.302 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:15:37.527 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 161, train_steps: 6440 +2026-06-20 15:15:38.205 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 15:15:38.293 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:15:38.341 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:15:38.405 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:16:44.417 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_1/index_1983_seed_31323_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 15:16:47.449 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:17:02.821 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:17:02.831 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:17:02.831 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:18:16.491 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_1/index_1983_seed_31324_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 15:18:19.736 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:18:35.596 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:18:40.375 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:18:41.450 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:20:35.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-20 15:20:36.332 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:22:36.147 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:29:36.903 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 162, train_steps: 6480 +2026-06-20 15:29:37.568 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 15:29:37.669 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:29:37.710 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:29:37.783 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:30:43.599 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_1/index_7552_seed_18206_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 15:30:46.628 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:31:02.025 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:31:02.036 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:31:02.036 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:32:11.809 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_1/index_7552_seed_18207_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 15:32:14.956 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:32:30.622 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:32:37.443 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:32:38.347 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:34:33.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.48s (1.91min) +2026-06-20 15:34:33.827 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:36:31.317 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:43:31.490 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 163, train_steps: 6520 +2026-06-20 15:43:32.169 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 15:43:32.256 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:43:32.295 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:43:32.358 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:44:38.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000163/group_1/index_2681_seed_2623_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 15:44:43.232 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:44:58.565 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:44:58.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:44:58.575 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:46:09.411 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000163/group_1/index_2681_seed_2624_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 15:46:12.772 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:46:28.577 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:46:33.676 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:46:34.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:48:29.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.47s (1.91min) +2026-06-20 15:48:29.934 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:50:26.537 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:57:25.507 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 164, train_steps: 6560 +2026-06-20 15:57:26.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 15:57:26.417 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:57:26.556 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:57:26.619 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:58:32.114 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_1/index_783_seed_32680_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 15:58:35.114 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 15:58:51.012 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:58:51.019 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:58:51.019 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:00:02.501 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_1/index_783_seed_32681_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 16:00:05.566 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:00:21.423 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:00:26.401 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:00:27.307 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:02:21.064 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.61s (1.89min) +2026-06-20 16:02:21.793 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:04:18.345 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:11:18.375 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 165, train_steps: 6600 +2026-06-20 16:11:19.089 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 16:11:19.294 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:11:19.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:11:19.508 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:12:26.672 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_1/index_4202_seed_9106_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 16:12:29.804 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:12:45.225 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:12:45.235 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:12:45.235 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:13:54.643 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_1/index_4202_seed_9107_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 16:13:59.460 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:14:15.424 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:14:19.105 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:14:20.372 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:16:14.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.75s (1.90min) +2026-06-20 16:16:14.870 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:18:14.125 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:25:13.932 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 166, train_steps: 6640 +2026-06-20 16:25:14.620 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 16:25:14.746 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:25:14.783 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:25:14.876 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:26:21.155 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_1/index_9873_seed_639_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 16:26:25.757 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:26:41.301 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:26:41.311 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:26:41.311 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:27:50.074 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_1/index_9873_seed_640_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 16:27:54.728 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:28:10.462 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:28:13.831 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:28:14.977 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:30:08.844 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.78s (1.90min) +2026-06-20 16:30:09.574 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:32:10.306 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:39:12.155 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 167, train_steps: 6680 +2026-06-20 16:39:12.849 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 16:39:12.985 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:39:13.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:39:13.148 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:40:19.254 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_1/index_4132_seed_6990_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 16:40:23.792 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:40:39.734 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:40:39.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:40:39.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:41:48.066 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_1/index_4132_seed_6991_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 16:41:51.584 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:42:07.202 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:42:12.094 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:42:13.288 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:44:08.745 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.36s (1.92min) +2026-06-20 16:44:09.477 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:46:10.019 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:53:11.384 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 168, train_steps: 6720 +2026-06-20 16:53:12.091 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 16:53:12.280 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:53:12.405 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:53:12.494 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:54:18.560 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_1/index_8012_seed_18047_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 16:54:21.920 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:54:37.597 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:54:37.608 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:54:37.608 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:55:47.308 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_1/index_8012_seed_18048_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 16:55:50.379 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 16:56:06.031 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:56:11.767 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:56:12.678 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:58:07.736 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-20 16:58:08.464 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:00:07.578 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:07:09.182 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 169, train_steps: 6760 +2026-06-20 17:07:09.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 17:07:10.069 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 17:07:10.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:07:10.336 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:08:16.380 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_1/index_3620_seed_1149_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 17:08:21.920 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 17:08:37.791 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:08:37.801 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:08:37.801 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:09:47.768 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_1/index_3620_seed_1150_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 17:09:51.027 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 17:10:06.869 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:10:11.903 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 17:10:12.818 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:12:08.254 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.20s (1.92min) +2026-06-20 17:12:08.969 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:14:09.318 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:21:11.057 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 170, train_steps: 6800 +2026-06-20 17:21:47.174 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=170 (missing=0, unexpected=0) +2026-06-20 17:21:49.218 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 17:21:49.344 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 17:21:49.362 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:21:49.426 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:22:56.782 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_1/index_8105_seed_7477_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 17:22:59.853 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 17:23:14.612 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:23:14.623 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:23:14.623 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:24:25.561 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_1/index_8105_seed_7478_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 17:24:28.462 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 17:24:43.715 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:24:49.925 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 17:24:50.829 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:26:44.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.05s (1.88min) +2026-06-20 17:26:44.742 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:28:41.425 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:35:39.741 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 6, sp_rank 0: update step: 171, train_steps: 6840 +2026-06-20 17:35:40.432 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 6, dp_rank 3: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=1, rank_in_group=1 +2026-06-20 17:35:40.626 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 17:35:40.893 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:35:41.014 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:36:46.993 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000171/group_1/index_3966_seed_10524_dp_3_rank_6_sample_2_slot_0.mp4 +2026-06-20 17:36:49.910 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 17:37:04.795 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:37:04.806 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 6 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:37:04.807 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 6 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:38:19.265 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 6: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000171/group_1/index_3966_seed_10525_dp_3_rank_6_sample_3_slot_1.mp4 +2026-06-20 17:38:23.887 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:6 for reward scoring. +2026-06-20 17:38:39.340 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:38:42.536 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 6: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 17:38:43.752 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 6: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:40:36.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 6: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.89s (1.88min) +2026-06-20 17:40:37.378 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 6, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-29 01:30:08.473 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 01:30:09.136 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 01:30:09.137 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 02:00:10.871 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 02:00:11.538 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 02:00:11.539 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 02:27:18.828 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 02:27:19.496 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 02:27:19.497 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 02:48:01.811 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 02:48:02.485 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 02:48:02.485 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:09:13.339 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 03:09:13.996 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:09:13.997 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:38:08.735 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 03:38:09.398 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:38:09.399 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:00:48.883 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 04:00:49.561 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:00:49.561 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:28:55.231 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 04:28:55.908 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:28:55.911 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:56:38.319 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 04:56:38.975 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:56:38.975 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 05:18:38.153 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 05:18:38.833 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 05:18:38.834 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 05:41:14.734 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 05:41:15.400 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 05:41:15.401 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:04:33.978 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 06:04:34.646 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:04:34.649 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:27:32.800 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 06:27:33.459 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:27:33.462 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:51:38.408 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 06:51:39.082 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:51:39.084 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 07:13:53.756 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 07:13:54.427 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 07:13:54.428 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 07:43:26.140 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 07:43:26.827 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 07:43:26.828 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:08:31.927 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 08:08:32.600 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:08:32.601 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:31:43.443 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 08:31:44.117 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:31:44.118 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:01:10.721 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 09:01:11.383 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:01:11.383 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:24:41.227 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 09:24:41.881 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:24:41.883 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:49:41.316 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 09:49:41.973 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:49:41.974 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 10:12:12.868 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 10:12:13.536 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 10:12:13.539 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 10:38:10.718 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 10:38:11.400 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 10:38:11.400 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:04:32.846 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 11:04:33.500 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:04:33.500 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:34:57.977 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 11:34:58.628 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:34:58.629 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:57:16.507 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 11:57:17.166 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:57:17.167 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 12:20:18.526 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 12:20:19.183 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 12:20:19.185 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 12:48:18.021 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 12:48:18.681 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 12:48:18.683 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 13:16:07.484 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 13:16:08.152 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 13:16:08.153 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 13:39:08.993 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 13:39:09.702 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 13:39:09.703 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:05:44.069 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 14:05:44.743 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:05:44.746 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:28:52.354 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 14:28:53.024 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:28:53.025 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:51:27.434 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 14:51:28.095 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:51:28.096 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 15:22:59.515 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 15:23:00.178 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 15:23:00.180 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 15:48:54.215 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 15:48:54.867 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 15:48:54.869 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 16:18:24.821 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 16:18:25.479 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 16:18:25.482 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 16:42:26.779 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 16:42:27.443 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 16:42:27.444 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:11:11.626 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 17:11:12.298 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:11:12.299 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:34:03.136 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 17:34:03.803 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:34:03.806 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:55:30.140 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 17:55:30.803 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:55:30.804 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 18:18:03.009 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 18:18:03.689 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 18:18:03.690 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 18:40:44.380 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 18:40:45.037 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 18:40:45.040 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_61.log b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_61.log new file mode 100644 index 0000000000000000000000000000000000000000..b94b9f3fc909453474284649d3885d97a457ad36 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_61.log @@ -0,0 +1,1994 @@ +2026-06-19 01:20:18.435 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-19 01:20:18.436 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-19 01:30:17.294 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 01:30:17.395 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:30:17.398 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:31:52.289 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:31:52.291 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:33:45.515 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 01:34:01.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 01:35:56.997 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.94s (1.93min) +2026-06-19 01:35:57.692 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 01:37:54.191 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 01:37:57.219 | INFO | hyvideo.utils.grpo_utils:train_one_step:3127 - Rank 61, sp_rank 0: init adaptive KL controller (init=1.000e-07, min=1.000e-07, max=1.000e-05) +2026-06-19 01:45:11.480 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 1, train_steps: 40 +2026-06-19 01:45:12.575 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 01:45:12.730 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:45:12.791 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:46:16.894 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:46:16.895 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:48:10.945 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 01:48:11.958 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 01:50:07.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.70s (1.93min) +2026-06-19 01:50:08.465 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 01:52:06.955 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 01:59:06.767 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 2, train_steps: 80 +2026-06-19 01:59:07.381 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 01:59:07.541 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:59:07.596 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:00:11.623 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:00:11.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:02:10.528 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 02:02:11.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:04:06.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.88s (1.91min) +2026-06-19 02:04:07.374 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:06:03.753 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:13:04.860 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 3, train_steps: 120 +2026-06-19 02:13:05.465 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 02:13:05.568 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:13:05.575 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:14:10.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:14:10.353 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:16:04.655 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 02:16:05.773 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:17:59.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.77s (1.90min) +2026-06-19 02:18:00.448 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:19:59.351 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:26:59.888 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 4, train_steps: 160 +2026-06-19 02:27:00.546 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 02:27:00.642 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:27:00.689 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:28:05.050 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:28:05.053 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:30:00.316 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 02:30:01.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:31:55.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.62s (1.91min) +2026-06-19 02:31:56.651 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:33:53.224 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:40:53.219 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 5, train_steps: 200 +2026-06-19 02:40:53.901 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 02:40:54.102 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:40:54.223 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:41:58.325 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:41:58.325 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:43:51.812 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 02:43:52.645 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:45:46.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.50s (1.89min) +2026-06-19 02:45:46.993 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:47:43.352 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:54:40.661 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 6, train_steps: 240 +2026-06-19 02:54:41.360 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 02:54:41.491 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:54:41.524 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:55:46.167 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:55:46.170 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:57:42.949 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 02:57:43.805 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:59:37.354 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.38s (1.89min) +2026-06-19 02:59:38.043 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:01:37.731 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:08:37.479 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 7, train_steps: 280 +2026-06-19 03:08:38.147 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 03:08:38.263 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:08:38.297 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:09:42.625 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:09:42.629 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:11:38.337 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 03:11:39.328 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:13:34.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.68s (1.91min) +2026-06-19 03:13:34.977 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:15:32.043 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:22:31.154 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 8, train_steps: 320 +2026-06-19 03:22:31.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 03:22:31.982 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:22:32.071 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:23:37.027 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:23:37.030 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:25:33.100 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 03:25:33.949 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:27:28.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-19 03:27:29.431 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:29:26.905 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:36:25.681 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 9, train_steps: 360 +2026-06-19 03:36:26.346 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 03:36:26.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:36:26.591 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:37:30.953 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:37:30.953 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:39:26.010 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 03:39:26.855 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:41:20.400 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.36s (1.89min) +2026-06-19 03:41:21.105 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:43:17.563 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:50:16.937 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 10, train_steps: 400 +2026-06-19 03:50:53.156 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=10 (missing=0, unexpected=0) +2026-06-19 03:50:56.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 03:50:56.310 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:50:56.346 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:52:02.021 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:52:02.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:53:57.828 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 03:53:58.673 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:55:52.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.65s (1.89min) +2026-06-19 03:55:53.208 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:57:49.764 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:04:48.381 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 11, train_steps: 440 +2026-06-19 04:04:49.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 04:04:49.212 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:04:49.244 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:05:53.605 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:05:53.606 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:07:48.853 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 04:07:49.699 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:09:43.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.25s (1.89min) +2026-06-19 04:09:43.797 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:11:40.267 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:18:37.295 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 12, train_steps: 480 +2026-06-19 04:18:37.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 04:18:38.115 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:18:38.149 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:19:42.646 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:19:42.646 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:21:37.180 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 04:21:38.084 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:23:33.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.19s (1.92min) +2026-06-19 04:23:34.070 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:25:31.243 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:32:29.819 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 13, train_steps: 520 +2026-06-19 04:32:30.514 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 04:32:30.732 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:32:30.880 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:33:35.215 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:33:35.215 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:35:30.313 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 04:35:31.200 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:37:24.102 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.77s (1.88min) +2026-06-19 04:37:24.791 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:39:21.076 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:46:18.593 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 14, train_steps: 560 +2026-06-19 04:46:19.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 04:46:19.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:46:19.419 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:47:24.345 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:47:24.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:49:17.445 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 04:49:18.355 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:51:12.963 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.49s (1.91min) +2026-06-19 04:51:13.649 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:53:12.996 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:00:11.831 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 15, train_steps: 600 +2026-06-19 05:00:12.525 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 05:00:12.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:00:12.693 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:01:17.434 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:01:17.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:03:11.349 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 05:03:12.498 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:05:07.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.64s (1.91min) +2026-06-19 05:05:08.060 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:07:05.979 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:14:04.898 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 16, train_steps: 640 +2026-06-19 05:14:05.573 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 05:14:05.751 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:14:05.832 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:15:10.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:15:10.719 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:17:06.988 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 05:17:07.826 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:19:03.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.02s (1.92min) +2026-06-19 05:19:03.741 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:21:01.962 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:28:03.323 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 17, train_steps: 680 +2026-06-19 05:28:04.007 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 05:28:04.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:28:04.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:29:09.626 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:29:09.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:31:03.926 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 05:31:04.981 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:32:58.689 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.48s (1.89min) +2026-06-19 05:32:59.386 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:34:56.266 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:41:58.115 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 18, train_steps: 720 +2026-06-19 05:41:58.777 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 05:41:58.963 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:41:59.075 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:43:03.495 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:43:03.495 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:45:00.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 05:45:01.090 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:46:55.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.88s (1.90min) +2026-06-19 05:46:55.797 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:48:52.331 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:55:50.501 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 19, train_steps: 760 +2026-06-19 05:55:51.178 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 05:55:51.377 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:55:51.476 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:56:55.673 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:56:55.674 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:58:49.068 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 05:58:49.911 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:00:43.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.10s (1.89min) +2026-06-19 06:00:43.876 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:02:40.341 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:09:39.759 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 20, train_steps: 800 +2026-06-19 06:10:15.889 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=20 (missing=0, unexpected=0) +2026-06-19 06:10:18.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 06:10:18.895 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:10:18.958 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:11:24.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:11:24.856 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:13:19.050 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 06:13:20.013 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:15:13.224 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.90s (1.88min) +2026-06-19 06:15:13.910 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:17:12.688 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:24:12.559 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 21, train_steps: 840 +2026-06-19 06:24:13.270 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 06:24:13.412 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:24:13.456 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:25:17.826 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:25:17.827 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:27:11.261 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 06:27:12.178 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:29:06.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.51s (1.91min) +2026-06-19 06:29:07.667 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:31:04.582 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:38:04.280 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 22, train_steps: 880 +2026-06-19 06:38:04.947 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 06:38:05.081 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:38:05.114 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:39:10.324 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:39:10.324 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:41:03.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 06:41:04.430 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:42:58.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.99s (1.90min) +2026-06-19 06:42:59.164 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:44:56.505 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:51:58.173 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 23, train_steps: 920 +2026-06-19 06:51:58.843 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 06:51:59.042 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:51:59.170 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:53:03.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:53:03.202 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:54:59.789 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 06:55:00.641 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:56:54.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.11s (1.90min) +2026-06-19 06:56:55.621 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:58:52.320 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:05:51.323 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 24, train_steps: 960 +2026-06-19 07:05:52.004 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 07:05:52.108 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:05:52.150 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:06:56.720 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:06:56.723 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:08:52.217 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 07:08:53.044 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:10:46.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-19 07:10:46.871 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:12:44.155 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:19:41.975 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 25, train_steps: 1000 +2026-06-19 07:20:52.105 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 07:20:52.307 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:20:52.418 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:21:56.696 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:21:56.699 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:23:52.666 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 07:23:53.536 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:25:46.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.81s (1.88min) +2026-06-19 07:25:47.205 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:27:44.382 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:34:43.697 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 26, train_steps: 1040 +2026-06-19 07:34:44.357 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 07:34:44.545 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:34:44.640 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:35:48.707 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:35:48.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:37:45.443 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 07:37:46.299 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:39:39.639 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.14s (1.89min) +2026-06-19 07:39:40.348 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:41:39.073 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:48:39.678 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 27, train_steps: 1080 +2026-06-19 07:48:40.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 07:48:40.528 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:48:40.622 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:49:44.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:49:44.881 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:51:43.285 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 07:51:44.133 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:53:38.298 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.99s (1.90min) +2026-06-19 07:53:39.007 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:55:38.421 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:02:38.176 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 28, train_steps: 1120 +2026-06-19 08:02:38.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 08:02:39.040 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:02:39.130 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:03:43.202 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:03:43.206 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:05:41.901 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 08:05:42.948 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:07:37.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.24s (1.90min) +2026-06-19 08:07:38.091 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:09:37.936 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:16:37.849 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 29, train_steps: 1160 +2026-06-19 08:16:38.691 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 08:16:38.833 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:16:38.862 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:17:43.299 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:17:43.305 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:19:42.881 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 08:19:43.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:21:37.264 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 08:21:37.960 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:23:37.407 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:30:39.546 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 30, train_steps: 1200 +2026-06-19 08:31:15.495 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=30 (missing=0, unexpected=0) +2026-06-19 08:31:18.258 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 08:31:18.426 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:31:18.518 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:32:24.989 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:32:24.989 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:34:19.980 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 08:34:20.972 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:36:14.510 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-19 08:36:15.226 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:38:14.174 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:45:14.286 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 31, train_steps: 1240 +2026-06-19 08:45:14.992 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 08:45:15.106 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:45:15.142 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:46:19.612 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:46:19.614 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:48:15.288 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 08:48:16.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:50:10.323 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.84s (1.90min) +2026-06-19 08:50:11.023 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:52:11.151 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:59:13.099 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 32, train_steps: 1280 +2026-06-19 08:59:13.775 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 08:59:13.898 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:59:13.934 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:00:18.328 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:00:18.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:02:13.436 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 09:02:14.507 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:04:08.713 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.91s (1.90min) +2026-06-19 09:04:09.421 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:06:10.410 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:13:13.282 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 33, train_steps: 1320 +2026-06-19 09:13:13.953 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 09:13:14.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:13:14.273 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:14:18.582 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:14:18.583 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:16:14.822 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 09:16:15.794 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:18:11.066 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.06s (1.92min) +2026-06-19 09:18:11.785 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:20:11.626 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:27:13.734 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 34, train_steps: 1360 +2026-06-19 09:27:14.444 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 09:27:14.550 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:27:14.585 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:28:18.937 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:28:18.938 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:30:13.695 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 09:30:14.908 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:32:08.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.81s (1.90min) +2026-06-19 09:32:09.592 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:34:09.779 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:41:12.395 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 35, train_steps: 1400 +2026-06-19 09:41:13.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 09:41:13.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:41:13.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:42:17.545 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:42:17.545 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:44:13.080 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 09:44:13.955 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:46:10.686 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 116.56s (1.94min) +2026-06-19 09:46:11.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:48:10.896 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:55:11.959 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 36, train_steps: 1440 +2026-06-19 09:55:12.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 09:55:12.791 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:55:12.863 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:56:17.640 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:56:17.641 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:58:12.646 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 09:58:13.760 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:00:08.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-19 10:00:09.671 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:02:10.007 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:09:11.706 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 37, train_steps: 1480 +2026-06-19 10:09:12.395 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 10:09:12.556 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:09:12.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:10:16.893 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:10:16.893 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:12:13.042 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 10:12:13.954 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:14:09.030 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.94s (1.92min) +2026-06-19 10:14:09.737 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:16:10.368 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:23:12.346 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 38, train_steps: 1520 +2026-06-19 10:23:13.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 10:23:13.196 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:23:13.234 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:24:17.734 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:24:17.744 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:26:12.667 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 10:26:13.797 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:28:09.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.16s (1.92min) +2026-06-19 10:28:09.954 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:30:10.742 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:37:11.949 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 39, train_steps: 1560 +2026-06-19 10:37:12.646 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 10:37:12.780 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:37:12.809 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:38:17.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:38:17.360 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:41:42.526 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 10:41:43.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:43:36.471 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.70s (1.88min) +2026-06-19 10:43:37.180 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:45:37.177 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:52:37.856 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 40, train_steps: 1600 +2026-06-19 10:53:13.819 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=40 (missing=0, unexpected=0) +2026-06-19 10:53:15.773 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 10:53:15.924 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:53:15.995 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:54:21.989 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:54:21.997 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:56:16.410 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 10:56:17.359 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:58:11.717 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.27s (1.90min) +2026-06-19 10:58:12.415 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:00:11.328 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:07:12.586 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 41, train_steps: 1640 +2026-06-19 11:07:13.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 11:07:13.501 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:07:13.642 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:08:17.972 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:08:17.973 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:10:12.668 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 11:10:13.810 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:12:08.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.28s (1.90min) +2026-06-19 11:12:09.087 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:14:08.756 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:21:10.897 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 42, train_steps: 1680 +2026-06-19 11:21:11.805 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 11:21:12.006 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:21:12.134 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:22:16.575 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:22:16.578 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:24:12.073 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 11:24:13.155 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:26:07.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.32s (1.91min) +2026-06-19 11:26:08.278 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:28:06.796 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:35:06.776 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 43, train_steps: 1720 +2026-06-19 11:35:07.458 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 11:35:07.610 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:35:07.666 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:36:11.766 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:36:11.770 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:38:10.299 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 11:38:11.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:40:05.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.38s (1.91min) +2026-06-19 11:40:06.548 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:42:03.523 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:49:04.439 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 44, train_steps: 1760 +2026-06-19 11:49:05.113 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 11:49:05.285 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:49:05.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:50:10.382 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:50:10.382 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:52:06.491 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 11:52:07.410 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:54:01.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.69s (1.89min) +2026-06-19 11:54:01.909 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:56:01.233 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:03:03.876 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 45, train_steps: 1800 +2026-06-19 12:03:04.556 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 12:03:04.663 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:03:04.710 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:04:09.824 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:04:09.824 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:06:05.347 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 12:06:06.257 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:08:01.417 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.05s (1.92min) +2026-06-19 12:08:02.101 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:10:01.382 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:17:03.591 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 46, train_steps: 1840 +2026-06-19 12:17:04.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 12:17:04.419 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:17:04.457 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:18:09.826 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:18:09.826 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:20:05.333 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 12:20:06.457 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:22:01.733 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.17s (1.92min) +2026-06-19 12:22:02.441 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:24:01.449 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:31:02.932 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 47, train_steps: 1880 +2026-06-19 12:31:03.607 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 12:31:03.787 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:31:03.886 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:32:08.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:32:08.352 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:34:03.182 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 12:34:04.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:35:58.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-19 12:35:58.982 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:37:55.891 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:44:57.677 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 48, train_steps: 1920 +2026-06-19 12:44:58.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 12:44:58.555 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:44:58.698 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:46:03.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:46:03.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:47:58.931 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 12:47:59.808 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:49:54.620 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-19 12:49:55.343 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:51:52.290 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:58:52.089 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 49, train_steps: 1960 +2026-06-19 12:58:52.782 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 12:58:52.946 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:58:53.001 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:59:57.444 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:59:57.447 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:01:54.369 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 13:01:55.235 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:03:48.446 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.04s (1.88min) +2026-06-19 13:03:49.148 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:05:45.468 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:12:45.280 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 50, train_steps: 2000 +2026-06-19 13:14:20.925 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=50 (missing=0, unexpected=0) +2026-06-19 13:14:23.203 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 13:14:23.357 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:14:23.432 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:15:29.977 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:15:29.980 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:17:25.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 13:17:26.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:19:20.081 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.84s (1.90min) +2026-06-19 13:19:20.807 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:21:17.523 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:28:18.476 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 51, train_steps: 2040 +2026-06-19 13:28:19.184 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 13:28:19.302 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:28:19.337 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:29:23.864 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:29:23.867 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:31:18.028 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 13:31:19.005 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:33:13.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.83s (1.90min) +2026-06-19 13:33:13.741 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:35:13.228 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:42:13.614 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 52, train_steps: 2080 +2026-06-19 13:42:14.278 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 13:42:14.486 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:42:14.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:43:19.139 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:43:19.143 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:45:15.287 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 13:45:16.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:47:10.481 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.03s (1.90min) +2026-06-19 13:47:11.196 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:49:11.264 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:56:12.606 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 53, train_steps: 2120 +2026-06-19 13:56:13.303 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 13:56:13.403 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:56:13.447 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:57:17.998 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:57:18.001 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:59:13.130 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 13:59:14.045 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:01:08.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.41s (1.91min) +2026-06-19 14:01:09.528 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:03:10.472 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:10:12.167 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 54, train_steps: 2160 +2026-06-19 14:10:12.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 14:10:12.980 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:10:13.020 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:11:17.303 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:11:17.306 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:13:12.866 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 14:13:13.767 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:15:08.493 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.42s (1.91min) +2026-06-19 14:15:09.217 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:17:09.305 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:24:11.917 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 55, train_steps: 2200 +2026-06-19 14:24:12.604 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 14:24:12.814 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:24:12.931 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:25:17.353 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:25:17.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:27:12.754 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 14:27:13.834 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:29:09.888 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.82s (1.93min) +2026-06-19 14:29:10.604 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:31:10.762 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:38:12.113 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 56, train_steps: 2240 +2026-06-19 14:38:12.813 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 14:38:13.022 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:38:13.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:39:17.369 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:39:17.372 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:41:13.092 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 14:41:14.006 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:43:08.781 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.66s (1.91min) +2026-06-19 14:43:09.474 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:45:09.717 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:52:11.842 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 57, train_steps: 2280 +2026-06-19 14:52:12.525 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 14:52:12.645 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:52:12.686 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:53:16.767 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:53:16.770 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:55:13.057 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 14:55:13.931 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:57:08.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.79s (1.91min) +2026-06-19 14:57:09.596 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:59:09.880 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:06:12.287 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 58, train_steps: 2320 +2026-06-19 15:06:12.985 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 15:06:13.148 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:06:13.196 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:07:17.857 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:07:17.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:09:13.062 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 15:09:13.930 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:11:08.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.71s (1.91min) +2026-06-19 15:11:09.498 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:13:09.810 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:20:12.120 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 59, train_steps: 2360 +2026-06-19 15:20:12.738 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 15:20:12.929 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:20:13.034 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:21:17.555 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:21:17.558 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:23:12.372 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 15:23:13.445 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:25:09.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.28s (1.92min) +2026-06-19 15:25:09.742 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:27:10.366 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:34:12.070 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 60, train_steps: 2400 +2026-06-19 15:34:47.964 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=60 (missing=0, unexpected=0) +2026-06-19 15:34:50.615 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 15:34:50.728 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:34:50.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:35:56.792 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:35:56.792 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:37:53.529 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 15:37:54.634 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:39:47.707 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-19 15:39:48.403 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:41:45.235 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:48:44.589 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 61, train_steps: 2440 +2026-06-19 15:48:45.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 15:48:45.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:48:45.477 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:49:49.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:49:49.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:51:47.609 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 15:51:48.546 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:53:41.681 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.04s (1.88min) +2026-06-19 15:53:42.399 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:55:39.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:02:39.010 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 62, train_steps: 2480 +2026-06-19 16:02:39.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 16:02:39.860 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:02:39.943 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:03:44.226 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:03:44.227 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:05:44.132 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 16:05:45.097 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:07:38.761 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.59s (1.89min) +2026-06-19 16:07:39.470 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:09:38.854 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:16:39.318 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 63, train_steps: 2520 +2026-06-19 16:16:40.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 16:16:40.190 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:16:40.239 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:17:44.947 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:17:44.951 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:19:45.197 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 16:19:46.068 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:21:39.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.36s (1.89min) +2026-06-19 16:21:40.302 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:23:38.881 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:30:40.945 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 64, train_steps: 2560 +2026-06-19 16:30:41.632 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 16:30:41.833 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:30:41.958 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:31:46.359 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:31:46.360 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:33:45.471 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 16:33:46.411 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:35:40.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.63s (1.89min) +2026-06-19 16:35:40.844 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:37:39.397 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:44:41.299 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 65, train_steps: 2600 +2026-06-19 16:44:41.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 16:44:42.160 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:44:42.269 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:45:46.868 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:45:46.871 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:47:47.482 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 16:47:48.400 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:49:42.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.72s (1.90min) +2026-06-19 16:49:42.958 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:51:40.035 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:58:41.283 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 66, train_steps: 2640 +2026-06-19 16:58:41.990 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 16:58:42.147 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:58:42.198 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:59:46.429 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:59:46.429 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:01:43.599 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 17:01:44.525 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:03:37.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.34s (1.89min) +2026-06-19 17:03:38.681 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:05:38.385 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:12:40.133 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 67, train_steps: 2680 +2026-06-19 17:12:40.908 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 17:12:41.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:12:41.347 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:13:45.787 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:13:45.790 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:15:44.529 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 17:15:45.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:17:38.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 17:17:39.678 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:19:39.113 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:26:39.551 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 68, train_steps: 2720 +2026-06-19 17:26:40.216 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 17:26:40.411 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:26:40.536 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:27:45.088 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:27:45.088 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:29:43.886 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 17:29:44.843 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:31:38.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.76s (1.90min) +2026-06-19 17:31:39.420 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:33:38.457 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:40:38.373 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 69, train_steps: 2760 +2026-06-19 17:40:39.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 17:40:39.217 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:40:39.324 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:41:43.658 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:41:43.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:43:41.069 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 17:43:42.086 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:45:35.854 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.55s (1.89min) +2026-06-19 17:45:36.569 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:47:37.114 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:54:38.580 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 70, train_steps: 2800 +2026-06-19 17:55:14.755 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=70 (missing=0, unexpected=0) +2026-06-19 17:55:16.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 17:55:17.132 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:55:17.216 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:56:23.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:56:23.514 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:58:18.228 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 17:58:19.362 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:00:13.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.58s (1.89min) +2026-06-19 18:00:13.837 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:02:13.853 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:09:13.899 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 71, train_steps: 2840 +2026-06-19 18:09:14.615 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 18:09:14.736 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:09:14.773 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:10:19.107 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:10:19.108 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:12:16.021 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 18:12:16.974 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:14:12.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.67s (1.93min) +2026-06-19 18:14:13.459 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:16:12.693 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:23:13.832 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 72, train_steps: 2880 +2026-06-19 18:23:14.506 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 18:23:14.663 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:23:14.725 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:24:19.378 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:24:19.379 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:26:14.711 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 18:26:15.790 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:28:10.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.45s (1.91min) +2026-06-19 18:28:11.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:30:10.989 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:37:12.789 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 73, train_steps: 2920 +2026-06-19 18:37:13.456 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 18:37:13.636 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:37:13.745 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:38:18.152 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:38:18.155 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:40:13.249 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 18:40:14.122 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:42:09.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.16s (1.92min) +2026-06-19 18:42:10.285 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:44:10.447 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:51:12.753 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 74, train_steps: 2960 +2026-06-19 18:51:13.468 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 18:51:13.668 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:51:13.773 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:52:18.449 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:52:18.452 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:54:13.982 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 18:54:14.864 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:56:10.312 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.21s (1.92min) +2026-06-19 18:56:11.030 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:58:11.443 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:05:14.029 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 75, train_steps: 3000 +2026-06-19 19:06:23.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 19:06:23.321 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:06:23.372 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:07:28.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:07:28.134 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:09:24.209 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 19:09:25.134 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:11:18.397 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.14s (1.89min) +2026-06-19 19:11:19.100 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:13:16.252 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:20:16.319 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 76, train_steps: 3040 +2026-06-19 19:20:16.990 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 19:20:17.204 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:20:17.318 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:21:21.369 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:21:21.372 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:23:18.455 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 19:23:19.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:25:13.441 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.66s (1.89min) +2026-06-19 19:25:14.145 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:27:12.733 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:34:13.272 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 77, train_steps: 3080 +2026-06-19 19:34:13.950 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 19:34:14.109 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:34:14.165 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:35:18.617 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:35:18.617 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:37:13.073 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 19:37:14.000 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:39:08.734 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.60s (1.91min) +2026-06-19 19:39:09.440 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:41:09.770 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:48:12.650 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 78, train_steps: 3120 +2026-06-19 19:48:13.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 19:48:13.483 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:48:13.510 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:49:18.170 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:49:18.171 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:51:13.323 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 19:51:14.212 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:53:09.267 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-19 19:53:09.975 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:55:09.985 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:02:13.173 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 79, train_steps: 3160 +2026-06-19 20:02:13.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 20:02:14.016 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:02:14.047 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:03:18.434 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:03:18.435 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:05:13.093 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 20:05:13.952 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:07:09.660 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.53s (1.93min) +2026-06-19 20:07:10.385 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:09:11.033 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:16:13.941 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 80, train_steps: 3200 +2026-06-19 20:16:51.150 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=80 (missing=0, unexpected=0) +2026-06-19 20:16:53.092 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 20:16:53.228 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:16:53.262 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:18:00.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:18:00.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:19:54.752 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 20:19:55.722 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:21:49.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-19 20:21:50.362 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:23:47.273 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:30:47.433 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 81, train_steps: 3240 +2026-06-19 20:30:48.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 20:30:48.309 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:30:48.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:31:53.165 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:31:53.166 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:33:49.273 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 20:33:50.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:35:44.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.70s (1.89min) +2026-06-19 20:35:44.752 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:37:41.678 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:44:39.994 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 82, train_steps: 3280 +2026-06-19 20:44:40.696 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 20:44:40.880 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:44:40.964 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:45:45.277 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:45:45.281 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:47:44.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 20:47:44.993 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:49:38.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.42s (1.89min) +2026-06-19 20:49:39.275 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:51:38.424 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:58:38.997 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 83, train_steps: 3320 +2026-06-19 20:58:39.739 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 20:58:39.849 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:58:39.978 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:59:44.380 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:59:44.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:01:44.304 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 21:01:45.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:03:38.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.28s (1.89min) +2026-06-19 21:03:39.339 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:05:38.817 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:12:38.929 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 84, train_steps: 3360 +2026-06-19 21:12:39.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 21:12:39.802 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:12:39.870 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:13:44.296 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:13:44.297 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:15:43.791 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 21:15:44.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:17:38.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.31s (1.89min) +2026-06-19 21:17:38.863 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:19:38.523 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:26:39.501 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 85, train_steps: 3400 +2026-06-19 21:26:40.168 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 21:26:40.346 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:26:40.447 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:27:44.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:27:44.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:29:43.697 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 21:29:44.635 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:31:38.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-19 21:31:38.988 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:33:38.756 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:40:38.652 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 86, train_steps: 3440 +2026-06-19 21:40:39.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 21:40:39.526 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:40:39.587 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:41:44.476 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:41:44.479 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:43:42.867 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 21:43:43.802 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:45:37.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 21:45:37.987 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:47:37.708 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:54:39.049 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 87, train_steps: 3480 +2026-06-19 21:54:39.743 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 21:54:39.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:54:39.927 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:55:44.328 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:55:44.331 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:57:43.169 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 21:57:44.097 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:59:37.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.39s (1.89min) +2026-06-19 21:59:38.317 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:01:38.553 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:08:39.988 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 88, train_steps: 3520 +2026-06-19 22:08:40.669 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 22:08:40.769 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:08:40.801 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:09:45.271 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:09:45.275 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:11:43.755 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 22:11:44.650 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:13:38.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.57s (1.89min) +2026-06-19 22:13:39.103 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:15:38.304 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:22:38.704 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 89, train_steps: 3560 +2026-06-19 22:22:39.424 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 22:22:39.556 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:22:39.593 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:23:44.196 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:23:44.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:25:43.683 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 22:25:44.629 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:27:38.162 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 22:27:38.886 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:29:38.495 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:36:39.425 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 90, train_steps: 3600 +2026-06-19 22:37:15.555 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=90 (missing=0, unexpected=0) +2026-06-19 22:37:17.307 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 22:37:17.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:37:17.487 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:38:23.235 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:38:23.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:40:19.240 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 22:40:20.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:42:14.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.70s (1.89min) +2026-06-19 22:42:14.863 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:44:13.689 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:51:14.858 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 91, train_steps: 3640 +2026-06-19 22:51:15.700 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 22:51:15.843 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:51:15.887 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:52:20.232 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:52:20.232 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:54:15.937 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 22:54:16.816 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:56:12.672 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.67s (1.93min) +2026-06-19 22:56:13.388 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:58:13.111 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:05:14.270 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 92, train_steps: 3680 +2026-06-19 23:05:15.119 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 23:05:15.326 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:05:15.447 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:06:19.774 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:06:19.777 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:08:16.058 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 23:08:16.937 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:10:11.751 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.62s (1.91min) +2026-06-19 23:10:12.471 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:12:11.282 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:19:13.483 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 93, train_steps: 3720 +2026-06-19 23:19:14.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 23:19:14.300 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:19:14.338 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:20:18.766 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:20:18.769 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:22:14.354 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 23:22:15.407 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:24:10.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.33s (1.92min) +2026-06-19 23:24:11.707 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:26:11.285 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:33:13.944 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 94, train_steps: 3760 +2026-06-19 23:33:14.640 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 23:33:14.796 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:33:14.863 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:34:19.418 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:34:19.418 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:36:16.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 23:36:17.001 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:38:12.062 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.89s (1.91min) +2026-06-19 23:38:12.783 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:40:11.320 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:47:11.876 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 95, train_steps: 3800 +2026-06-19 23:47:12.596 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-19 23:47:12.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:47:12.916 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:48:17.150 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:48:17.154 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:50:12.004 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-19 23:50:13.113 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:52:08.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.82s (1.91min) +2026-06-19 23:52:08.790 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:54:07.677 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:01:09.778 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 96, train_steps: 3840 +2026-06-20 00:01:10.476 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 00:01:10.683 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:01:10.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:02:15.335 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:02:15.335 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:04:11.783 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 00:04:12.810 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:06:07.928 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 00:06:08.644 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:08:07.689 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:15:08.505 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 97, train_steps: 3880 +2026-06-20 00:15:09.218 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 00:15:09.347 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:15:09.382 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:16:13.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:16:13.625 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:18:11.238 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 00:18:12.273 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:20:07.671 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.20s (1.92min) +2026-06-20 00:20:08.400 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:22:07.125 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:29:07.009 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 98, train_steps: 3920 +2026-06-20 00:29:07.691 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 00:29:07.813 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:29:07.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:30:12.309 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:30:12.312 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:32:10.419 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 00:32:11.344 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:34:06.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 00:34:07.117 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:36:03.851 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:43:05.360 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 99, train_steps: 3960 +2026-06-20 00:43:06.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 00:43:06.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:43:06.296 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:44:11.285 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:44:11.285 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:46:09.899 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 00:46:10.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:48:05.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.75s (1.91min) +2026-06-20 00:48:06.440 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:50:03.397 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:57:04.806 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 100, train_steps: 4000 +2026-06-20 00:58:47.472 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=100 (missing=0, unexpected=0) +2026-06-20 00:58:49.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 00:58:49.719 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:58:49.750 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:59:55.769 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:59:55.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:01:50.303 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 01:01:51.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:03:44.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-20 01:03:45.434 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:05:42.219 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:12:39.873 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 101, train_steps: 4040 +2026-06-20 01:12:40.567 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 01:12:40.691 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:12:40.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:13:45.550 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:13:45.551 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:15:43.499 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 01:15:44.440 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:17:38.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.62s (1.89min) +2026-06-20 01:17:38.931 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:19:38.324 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:26:37.807 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 102, train_steps: 4080 +2026-06-20 01:26:38.504 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 01:26:38.707 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:26:38.810 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:27:42.905 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:27:42.905 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:29:41.183 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 01:29:42.419 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:31:36.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-20 01:31:37.098 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:33:37.225 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:40:38.544 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 103, train_steps: 4120 +2026-06-20 01:40:39.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 01:40:39.447 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:40:39.557 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:41:44.140 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:41:44.143 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:43:43.334 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 01:43:44.285 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:45:37.767 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-20 01:45:38.487 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:47:38.264 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:54:37.795 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 104, train_steps: 4160 +2026-06-20 01:54:38.703 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 01:54:38.862 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:54:38.932 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:55:43.613 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:55:43.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:57:42.889 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 01:57:43.841 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:59:37.314 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-20 01:59:38.031 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:01:38.019 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:08:38.464 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 105, train_steps: 4200 +2026-06-20 02:08:39.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 02:08:39.327 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:08:39.421 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:09:43.696 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:09:43.699 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:11:43.945 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 02:11:44.889 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:13:38.217 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 02:13:38.925 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:15:38.267 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:22:37.852 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 106, train_steps: 4240 +2026-06-20 02:22:38.564 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 02:22:38.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:22:38.698 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:23:43.086 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:23:43.086 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:25:40.847 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 02:25:41.800 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:27:35.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.90s (1.90min) +2026-06-20 02:27:36.583 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:29:36.810 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:36:37.096 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 107, train_steps: 4280 +2026-06-20 02:36:37.802 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 02:36:38.008 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:36:38.140 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:37:42.437 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:37:42.440 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:39:39.229 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 02:39:40.337 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:41:35.347 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 02:41:36.071 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:43:35.145 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:50:35.009 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 108, train_steps: 4320 +2026-06-20 02:50:35.726 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 02:50:35.839 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:50:35.875 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:51:40.339 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:51:40.342 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:53:37.574 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 02:53:38.503 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:55:33.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.45s (1.91min) +2026-06-20 02:55:33.972 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:57:30.647 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:04:30.703 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 109, train_steps: 4360 +2026-06-20 03:04:31.406 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 03:04:31.543 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:04:31.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:05:36.001 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:05:36.004 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:07:33.407 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 03:07:34.315 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:09:29.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-20 03:09:30.014 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:11:26.575 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:18:25.333 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 110, train_steps: 4400 +2026-06-20 03:19:01.201 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=110 (missing=0, unexpected=0) +2026-06-20 03:19:03.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 03:19:03.394 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:19:03.440 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:20:10.628 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:20:10.628 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:22:07.524 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 03:22:08.484 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:24:03.966 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.41s (1.92min) +2026-06-20 03:24:04.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:26:02.473 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:33:05.401 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 111, train_steps: 4440 +2026-06-20 03:33:06.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 03:33:06.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:33:06.383 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:34:11.224 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:34:11.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:36:10.098 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 03:36:11.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:38:06.021 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 03:38:06.731 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:40:04.480 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:47:04.349 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 112, train_steps: 4480 +2026-06-20 03:47:05.043 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 03:47:05.172 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:47:05.204 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:48:10.316 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:48:10.319 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:50:04.983 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 03:50:06.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:52:00.943 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-20 03:52:01.657 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:54:00.963 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:01:04.012 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 113, train_steps: 4520 +2026-06-20 04:01:04.876 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 04:01:05.084 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:01:05.195 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:02:09.889 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:02:09.889 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:04:06.611 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 04:04:07.532 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:06:02.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.15s (1.92min) +2026-06-20 04:06:03.532 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:08:01.582 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:15:03.782 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 114, train_steps: 4560 +2026-06-20 04:15:04.473 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 04:15:04.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:15:04.644 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:16:10.104 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:16:10.104 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:18:04.152 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 04:18:05.357 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:20:00.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.72s (1.91min) +2026-06-20 04:20:00.894 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:22:00.100 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:29:01.473 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 115, train_steps: 4600 +2026-06-20 04:29:02.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 04:29:02.311 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:29:02.354 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:30:07.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:30:07.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:32:02.494 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 04:32:03.431 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:33:58.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.33s (1.91min) +2026-06-20 04:33:58.775 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:35:55.625 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:42:56.479 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 116, train_steps: 4640 +2026-06-20 04:42:57.159 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 04:42:57.301 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:42:57.347 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:44:01.823 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:44:01.831 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:45:58.380 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 04:45:59.262 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:47:53.698 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.25s (1.90min) +2026-06-20 04:47:54.407 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:49:50.889 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:56:50.235 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 117, train_steps: 4680 +2026-06-20 04:56:50.953 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 04:56:51.093 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:56:51.129 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:57:55.754 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:57:55.755 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:59:51.798 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 04:59:52.668 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:01:46.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.49s (1.89min) +2026-06-20 05:01:47.078 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:03:44.147 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:10:42.589 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 118, train_steps: 4720 +2026-06-20 05:10:43.295 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 05:10:43.478 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:10:43.580 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:11:47.897 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:11:47.898 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:13:46.957 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 05:13:47.943 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:15:41.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.26s (1.89min) +2026-06-20 05:15:42.000 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:17:38.994 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:24:37.911 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 119, train_steps: 4760 +2026-06-20 05:24:38.622 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 05:24:38.770 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:24:38.807 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:25:43.063 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:25:43.063 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:27:40.119 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 05:27:41.213 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:29:34.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-20 05:29:35.559 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:31:33.105 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:38:32.161 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 120, train_steps: 4800 +2026-06-20 05:39:09.114 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=120 (missing=0, unexpected=0) +2026-06-20 05:39:10.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 05:39:11.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:39:11.066 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:40:17.485 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:40:17.487 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:42:12.566 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 05:42:13.441 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:44:08.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.96s (1.92min) +2026-06-20 05:44:09.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:46:10.086 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:53:12.061 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 121, train_steps: 4840 +2026-06-20 05:53:12.775 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 05:53:12.937 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:53:12.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:54:17.811 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:54:17.812 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:56:13.156 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 05:56:14.034 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:58:09.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.97s (1.92min) +2026-06-20 05:58:10.076 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:00:10.576 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:07:11.958 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 122, train_steps: 4880 +2026-06-20 06:07:12.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 06:07:12.747 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:07:12.789 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:08:17.269 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:08:17.270 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:10:13.195 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 06:10:14.078 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:12:08.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.25s (1.90min) +2026-06-20 06:12:09.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:14:10.580 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:21:11.876 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 123, train_steps: 4920 +2026-06-20 06:21:12.561 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 06:21:12.706 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:21:12.750 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:22:17.133 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:22:17.134 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:24:12.071 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 06:24:13.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:26:08.788 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.47s (1.92min) +2026-06-20 06:26:09.511 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:28:09.834 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:35:11.873 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 124, train_steps: 4960 +2026-06-20 06:35:12.538 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 06:35:12.668 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:35:12.699 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:36:16.963 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:36:16.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:38:12.129 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 06:38:13.085 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:40:08.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.44s (1.92min) +2026-06-20 06:40:09.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:42:09.773 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:49:11.651 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 125, train_steps: 5000 +2026-06-20 06:50:15.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 06:50:15.310 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:50:15.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:51:19.731 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:51:19.734 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:53:15.926 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 06:53:16.854 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:55:11.654 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.68s (1.91min) +2026-06-20 06:55:12.371 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:57:11.762 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:04:11.924 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 126, train_steps: 5040 +2026-06-20 07:04:12.630 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 07:04:12.749 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:04:12.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:05:17.517 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:05:17.522 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:07:13.988 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 07:07:15.097 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:09:10.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.08s (1.92min) +2026-06-20 07:09:11.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:11:11.165 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:18:11.936 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 127, train_steps: 5080 +2026-06-20 07:18:12.630 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 07:18:12.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:18:12.886 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:19:17.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:19:17.195 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:21:13.050 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 07:21:13.933 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:23:09.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.50s (1.93min) +2026-06-20 07:23:10.327 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:25:10.158 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:32:11.175 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 128, train_steps: 5120 +2026-06-20 07:32:11.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 07:32:12.040 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:32:12.090 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:33:16.378 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:33:16.388 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:35:13.016 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 07:35:13.896 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:37:08.641 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.56s (1.91min) +2026-06-20 07:37:09.373 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:39:09.930 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:46:11.993 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 129, train_steps: 5160 +2026-06-20 07:46:12.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 07:46:12.835 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:46:12.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:47:17.666 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:47:17.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:49:13.070 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 07:49:14.002 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:51:09.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-20 07:51:09.758 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:53:10.302 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:00:11.789 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 130, train_steps: 5200 +2026-06-20 08:00:49.109 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=130 (missing=0, unexpected=0) +2026-06-20 08:00:51.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 08:00:51.376 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:00:51.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:01:58.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:01:58.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:03:52.509 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 08:03:53.441 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:05:47.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.72s (1.90min) +2026-06-20 08:05:48.018 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:07:44.844 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:14:45.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 131, train_steps: 5240 +2026-06-20 08:14:46.025 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 08:14:46.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:14:46.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:15:50.985 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:15:50.989 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:17:46.990 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 08:17:47.931 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:19:41.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.38s (1.89min) +2026-06-20 08:19:42.138 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:21:39.467 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:28:37.833 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 132, train_steps: 5280 +2026-06-20 08:28:38.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 08:28:38.628 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:28:38.664 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:29:43.250 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:29:43.250 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:31:40.849 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 08:31:41.794 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:33:35.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.44s (1.89min) +2026-06-20 08:33:36.119 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:35:35.261 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:42:34.347 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 133, train_steps: 5320 +2026-06-20 08:42:35.017 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 08:42:35.159 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:42:35.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:43:39.732 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:43:39.734 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:45:35.888 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 08:45:36.826 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:47:32.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.77s (1.93min) +2026-06-20 08:47:33.449 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:49:30.692 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:56:29.505 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 134, train_steps: 5360 +2026-06-20 08:56:30.114 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 08:56:30.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:56:30.307 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:57:34.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:57:34.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:59:32.041 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 08:59:32.984 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:01:26.319 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 09:01:27.048 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:03:23.804 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:10:22.992 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 135, train_steps: 5400 +2026-06-20 09:10:23.674 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 09:10:23.800 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:10:23.837 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:11:28.543 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:11:28.547 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:13:23.932 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 09:13:24.865 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:15:18.588 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.61s (1.89min) +2026-06-20 09:15:19.310 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:17:17.099 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:24:16.805 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 136, train_steps: 5440 +2026-06-20 09:24:17.531 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 09:24:17.666 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:24:17.704 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:25:22.276 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:25:22.285 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:27:18.656 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 09:27:19.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:29:12.958 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.15s (1.89min) +2026-06-20 09:29:13.686 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:31:13.176 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:38:13.223 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 137, train_steps: 5480 +2026-06-20 09:38:13.936 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 09:38:14.081 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:38:14.117 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:39:18.418 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:39:18.419 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:41:14.371 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 09:41:15.473 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:43:09.522 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-20 09:43:10.250 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:45:10.503 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:52:12.154 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 138, train_steps: 5520 +2026-06-20 09:52:12.866 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 09:52:13.021 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:52:13.069 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:53:17.582 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:53:17.590 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:55:13.383 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 09:55:14.540 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:57:10.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.46s (1.92min) +2026-06-20 09:57:10.921 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:59:11.150 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:06:12.676 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 139, train_steps: 5560 +2026-06-20 10:06:13.363 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 10:06:13.502 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:06:13.532 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:07:18.047 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:07:18.050 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:09:13.264 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 10:09:14.197 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:11:08.675 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.20s (1.90min) +2026-06-20 10:11:09.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:13:09.780 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:20:12.687 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 140, train_steps: 5600 +2026-06-20 10:20:49.901 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=140 (missing=0, unexpected=0) +2026-06-20 10:20:51.812 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 10:20:51.953 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:20:51.989 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:21:58.260 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:21:58.263 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:23:53.040 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 10:23:53.954 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:25:47.600 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-20 10:25:48.313 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:27:45.363 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:34:44.011 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 141, train_steps: 5640 +2026-06-20 10:34:44.688 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 10:34:44.871 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:34:44.948 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:35:49.242 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:35:49.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:37:46.846 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 10:37:47.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:39:41.415 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.51s (1.89min) +2026-06-20 10:39:42.148 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:41:39.460 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:48:38.471 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 142, train_steps: 5680 +2026-06-20 10:48:39.157 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 10:48:39.270 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:48:39.306 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:49:43.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:49:43.990 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:51:43.153 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 10:51:44.071 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:53:37.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-20 10:53:37.851 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:55:37.993 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:02:37.921 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 143, train_steps: 5720 +2026-06-20 11:02:38.631 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 11:02:38.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:02:38.805 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:03:43.358 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:03:43.359 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:05:42.169 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 11:05:43.258 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:07:36.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 11:07:37.312 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:09:36.842 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:16:37.099 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 144, train_steps: 5760 +2026-06-20 11:16:37.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 11:16:37.973 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:16:38.050 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:17:42.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:17:42.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:19:38.707 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 11:19:39.746 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:21:33.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.47s (1.89min) +2026-06-20 11:21:34.087 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:23:30.884 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:30:30.400 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 145, train_steps: 5800 +2026-06-20 11:30:31.096 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 11:30:31.254 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:30:31.309 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:31:35.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:31:35.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:33:33.207 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 11:33:34.157 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:35:28.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.89s (1.90min) +2026-06-20 11:35:28.900 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:37:25.683 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:44:23.920 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 146, train_steps: 5840 +2026-06-20 11:44:24.628 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 11:44:24.752 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:44:24.794 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:45:29.300 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:45:29.302 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:47:23.247 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 11:47:24.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:49:19.133 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-20 11:49:19.856 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:51:17.039 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:58:15.931 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 147, train_steps: 5880 +2026-06-20 11:58:16.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 11:58:16.737 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:58:16.769 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:59:21.675 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:59:21.794 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:01:16.232 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 12:01:17.177 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:03:12.586 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.29s (1.92min) +2026-06-20 12:03:13.293 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:05:12.904 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:12:12.702 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 148, train_steps: 5920 +2026-06-20 12:12:13.403 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 12:12:13.538 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:12:13.571 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:13:18.034 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:13:18.037 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:15:13.144 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 12:15:14.019 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:17:09.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.69s (1.91min) +2026-06-20 12:17:09.788 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:19:10.588 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:26:11.391 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 149, train_steps: 5960 +2026-06-20 12:26:12.133 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 12:26:12.232 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:26:12.270 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:27:16.618 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:27:16.621 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:29:13.029 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 12:29:13.914 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:31:08.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.72s (1.91min) +2026-06-20 12:31:09.533 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:33:09.830 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:40:10.992 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 150, train_steps: 6000 +2026-06-20 12:41:53.533 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=150 (missing=0, unexpected=0) +2026-06-20 12:41:56.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 12:41:56.392 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:41:56.490 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:43:02.308 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:43:02.312 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:44:59.744 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 12:45:00.655 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:46:55.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.70s (1.91min) +2026-06-20 12:46:56.236 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:48:52.989 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:55:51.603 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 151, train_steps: 6040 +2026-06-20 12:55:52.306 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 12:55:52.438 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:55:52.476 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:56:56.643 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:56:56.643 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:58:50.721 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 12:58:51.681 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:00:46.190 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.41s (1.91min) +2026-06-20 13:00:46.908 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:02:43.652 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:09:41.910 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 152, train_steps: 6080 +2026-06-20 13:09:42.596 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 13:09:42.756 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:09:42.837 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:10:47.342 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:10:47.342 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:12:43.910 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 13:12:44.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:14:38.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-20 13:14:39.491 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:16:38.549 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:23:37.217 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 153, train_steps: 6120 +2026-06-20 13:23:38.094 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 13:23:38.265 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:23:38.344 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:24:43.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:24:43.189 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:26:40.127 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 13:26:41.243 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:28:34.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.39s (1.89min) +2026-06-20 13:28:35.473 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:30:32.907 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:37:31.701 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 154, train_steps: 6160 +2026-06-20 13:37:32.392 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 13:37:32.518 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:37:32.549 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:38:36.998 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:38:36.998 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:40:33.788 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 13:40:34.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:42:29.329 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.47s (1.91min) +2026-06-20 13:42:30.062 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:44:27.162 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:51:24.983 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 155, train_steps: 6200 +2026-06-20 13:51:25.700 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 13:51:25.847 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:51:25.876 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:52:30.116 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:52:30.119 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:54:24.447 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 13:54:25.336 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:56:19.129 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.63s (1.89min) +2026-06-20 13:56:19.859 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:58:16.750 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:05:15.349 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 156, train_steps: 6240 +2026-06-20 14:05:16.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 14:05:16.195 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:05:16.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:06:21.174 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:06:21.177 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:08:15.931 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 14:08:16.817 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:10:11.722 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.74s (1.91min) +2026-06-20 14:10:12.457 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:12:11.104 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:19:12.243 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 157, train_steps: 6280 +2026-06-20 14:19:12.924 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 14:19:13.055 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:19:13.090 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:20:17.632 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:20:17.635 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:22:12.398 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 14:22:13.487 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:24:08.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.22s (1.92min) +2026-06-20 14:24:09.689 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:26:09.905 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:33:10.830 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 158, train_steps: 6320 +2026-06-20 14:33:11.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 14:33:11.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:33:11.808 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:34:16.103 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:34:16.106 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:36:10.935 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 14:36:11.906 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:38:07.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.88s (1.91min) +2026-06-20 14:38:07.784 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:40:05.258 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:47:04.571 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 159, train_steps: 6360 +2026-06-20 14:47:05.291 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 14:47:05.441 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:47:05.507 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:48:10.324 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:48:10.324 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:50:04.048 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 14:50:05.327 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:51:59.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.86s (1.90min) +2026-06-20 14:51:59.929 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:53:57.322 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:00:59.184 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 160, train_steps: 6400 +2026-06-20 15:01:36.977 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=160 (missing=0, unexpected=0) +2026-06-20 15:01:39.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 15:01:39.465 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:01:39.558 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:02:45.276 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:02:45.279 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:04:42.879 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 15:04:43.756 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:06:37.409 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.48s (1.89min) +2026-06-20 15:06:38.136 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:08:38.302 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:15:37.528 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 161, train_steps: 6440 +2026-06-20 15:15:38.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 15:15:38.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:15:38.391 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:16:42.726 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:16:42.730 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:18:40.375 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 15:18:41.441 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:20:35.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-20 15:20:36.332 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:22:36.149 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:29:36.901 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 162, train_steps: 6480 +2026-06-20 15:29:37.569 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 15:29:37.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:29:37.749 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:30:42.184 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:30:42.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:32:37.444 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 15:32:38.340 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:34:33.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.48s (1.91min) +2026-06-20 15:34:33.827 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:36:31.317 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:43:31.486 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 163, train_steps: 6520 +2026-06-20 15:43:32.171 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 15:43:32.315 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:43:32.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:44:36.631 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:44:36.634 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:46:33.676 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 15:46:34.568 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:48:29.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.47s (1.91min) +2026-06-20 15:48:29.934 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:50:26.538 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:57:25.511 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 164, train_steps: 6560 +2026-06-20 15:57:26.231 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 15:57:26.435 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:57:26.538 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:58:30.632 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:58:30.633 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:00:26.402 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 16:00:27.287 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:02:21.064 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.61s (1.89min) +2026-06-20 16:02:21.793 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:04:18.345 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:11:18.374 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 165, train_steps: 6600 +2026-06-20 16:11:19.088 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 16:11:19.298 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:11:19.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:12:24.175 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:12:24.178 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:14:19.106 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 16:14:20.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:16:14.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.75s (1.90min) +2026-06-20 16:16:14.870 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:18:14.125 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:25:13.944 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 166, train_steps: 6640 +2026-06-20 16:25:14.622 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 16:25:14.802 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:25:14.902 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:26:19.260 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:26:19.264 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:28:13.831 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 16:28:14.891 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:30:08.844 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.78s (1.90min) +2026-06-20 16:30:09.574 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:32:10.306 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:39:12.157 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 167, train_steps: 6680 +2026-06-20 16:39:12.849 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 16:39:12.999 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:39:13.047 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:40:17.569 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:40:17.573 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:42:12.095 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 16:42:13.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:44:08.745 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.36s (1.92min) +2026-06-20 16:44:09.477 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:46:10.019 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:53:11.391 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 168, train_steps: 6720 +2026-06-20 16:53:12.091 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 16:53:12.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:53:12.400 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:54:16.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:54:16.604 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:56:11.767 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 16:56:12.710 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:58:07.736 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-20 16:58:08.464 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:00:07.577 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:07:09.181 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 169, train_steps: 6760 +2026-06-20 17:07:09.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 17:07:10.093 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:07:10.223 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:08:14.848 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:08:14.851 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:10:11.904 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 17:10:12.844 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:12:08.255 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.20s (1.92min) +2026-06-20 17:12:08.968 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:14:09.319 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:21:11.056 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 170, train_steps: 6800 +2026-06-20 17:21:47.148 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=170 (missing=0, unexpected=0) +2026-06-20 17:21:49.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 17:21:49.401 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:21:49.498 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:22:54.850 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:22:54.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:24:49.925 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 17:24:50.860 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:26:44.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.05s (1.88min) +2026-06-20 17:26:44.742 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:28:41.424 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:35:39.740 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 61, sp_rank 0: update step: 171, train_steps: 6840 +2026-06-20 17:35:40.430 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 61, dp_rank 30: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=15, rank_in_group=0 +2026-06-20 17:35:40.528 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:35:40.582 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:36:45.515 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 61 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:36:45.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 61 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:38:42.537 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 61: Using local metric names without global sync: ['avg'] +2026-06-20 17:38:43.597 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 61: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:40:36.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 61: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.89s (1.88min) +2026-06-20 17:40:37.377 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 61, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-29 01:29:39.211 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 01:29:39.212 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 01:52:02.457 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 01:52:02.458 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 02:21:21.739 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 02:21:21.740 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 02:46:50.626 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 02:46:50.628 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:07:22.304 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:07:22.307 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:31:52.556 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:31:52.557 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:00:42.555 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:00:42.556 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:26:07.225 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:26:07.225 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:51:26.621 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:51:26.623 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 05:15:30.865 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 05:15:30.866 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 05:38:45.529 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 05:38:45.530 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:02:15.484 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:02:15.485 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:25:05.798 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:25:05.798 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:48:14.138 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:48:14.140 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 07:12:19.822 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 07:12:19.825 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 07:33:34.016 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 07:33:34.017 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:06:48.142 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:06:48.142 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:28:48.459 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:28:48.460 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:58:34.333 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:58:34.334 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:22:11.202 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:22:11.204 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:47:59.755 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:47:59.756 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 10:09:04.707 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 10:09:04.708 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 10:35:37.633 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 10:35:37.634 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:03:38.440 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:03:38.443 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:27:22.571 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:27:22.571 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:55:35.347 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:55:35.350 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 12:18:29.927 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 12:18:29.927 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 12:46:33.824 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 12:46:33.828 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 13:12:55.497 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 13:12:55.499 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 13:37:18.385 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 13:37:18.386 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:03:52.529 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:03:52.531 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:26:03.612 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:26:03.615 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:49:03.117 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:49:03.117 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 15:21:11.450 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 15:21:11.450 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 15:46:20.302 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 15:46:20.304 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 16:17:02.445 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 16:17:02.447 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 16:41:03.990 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 16:41:03.993 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:08:19.588 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:08:19.589 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:32:11.563 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:32:11.564 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:53:40.788 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:53:40.791 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 18:16:06.320 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 18:16:06.321 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 18:38:51.539 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 18:38:51.541 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_8.log b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_8.log new file mode 100644 index 0000000000000000000000000000000000000000..745f23f130773331980628bd3f2438e39becb50d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/logs/training_rank_8.log @@ -0,0 +1,3413 @@ +2026-06-19 01:22:18.215 | INFO | __main__:__init__:933 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-19 01:22:18.897 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-19 01:22:18.899 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-19 01:30:17.294 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 01:30:31.481 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 01:30:31.483 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:30:31.485 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:31:52.978 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_2/index_3698_seed_1481_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 01:31:55.876 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 01:32:14.917 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:32:14.926 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:32:14.926 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:33:22.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_2/index_3698_seed_1482_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 01:33:27.580 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 01:33:45.117 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:33:45.515 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 01:33:50.989 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 01:34:00.993 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 01:35:56.997 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.94s (1.93min) +2026-06-19 01:35:57.655 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 01:37:54.183 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 01:37:57.219 | INFO | hyvideo.utils.grpo_utils:train_one_step:3127 - Rank 8, sp_rank 0: init adaptive KL controller (init=1.000e-07, min=1.000e-07, max=1.000e-05) +2026-06-19 01:45:11.870 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 1, train_steps: 40 +2026-06-19 01:45:12.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 01:45:12.708 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 01:45:12.734 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:45:12.788 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:46:19.066 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000001/group_2/index_2519_seed_2818_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 01:46:23.774 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 01:46:40.522 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:46:40.531 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:46:40.531 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 01:47:47.289 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000001/group_2/index_2519_seed_2819_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 01:47:51.961 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 01:48:08.913 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 01:48:10.945 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 01:48:11.063 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 01:48:12.035 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 01:50:07.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.70s (1.93min) +2026-06-19 01:50:08.465 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 01:52:06.953 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 01:59:06.772 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 2, train_steps: 80 +2026-06-19 01:59:07.380 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 01:59:07.467 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 01:59:07.474 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 01:59:07.479 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:00:13.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000002/group_2/index_867_seed_29546_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 02:00:17.690 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 02:00:35.195 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:00:35.201 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:00:35.202 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:01:46.305 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000002/group_2/index_867_seed_29547_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 02:01:51.320 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 02:02:08.643 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:02:10.528 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:02:10.626 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 02:02:11.649 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:04:06.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.88s (1.91min) +2026-06-19 02:04:07.374 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:06:03.754 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:13:04.864 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 3, train_steps: 120 +2026-06-19 02:13:05.465 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 02:13:05.590 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:13:05.602 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:13:05.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:14:11.482 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_2/index_1370_seed_7854_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 02:14:15.669 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 02:14:33.326 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:14:33.335 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:14:33.337 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:15:40.561 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_2/index_1370_seed_7855_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 02:15:45.547 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 02:16:02.637 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:16:04.655 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:16:04.762 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 02:16:05.756 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:17:59.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.77s (1.90min) +2026-06-19 02:18:00.448 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:19:59.350 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:26:59.894 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 4, train_steps: 160 +2026-06-19 02:27:00.546 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 02:27:00.651 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:27:00.695 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:27:00.745 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:28:06.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000004/group_2/index_8889_seed_30524_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 02:28:10.471 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 02:28:28.305 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:28:28.314 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:28:28.314 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:29:35.134 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000004/group_2/index_8889_seed_30525_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 02:29:39.844 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 02:29:57.266 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:30:00.316 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:30:00.428 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 02:30:01.308 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:31:55.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.62s (1.91min) +2026-06-19 02:31:56.651 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:33:53.224 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:40:53.233 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 5, train_steps: 200 +2026-06-19 02:40:53.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 02:40:53.993 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:40:54.028 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:40:54.057 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:41:59.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_2/index_2388_seed_21607_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 02:42:04.443 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 02:42:22.643 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:42:22.651 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:42:22.652 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:43:28.524 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_2/index_2388_seed_21608_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 02:43:33.600 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 02:43:51.253 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:43:51.812 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:43:51.863 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 02:43:52.729 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:45:46.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.50s (1.89min) +2026-06-19 02:45:46.993 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 02:47:43.351 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 02:54:40.663 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 6, train_steps: 240 +2026-06-19 02:54:41.362 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 02:54:41.574 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 02:54:41.909 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:54:41.973 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:55:47.702 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_2/index_1599_seed_23460_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 02:55:52.392 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 02:56:09.763 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:56:09.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 02:56:09.772 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 02:57:19.376 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_2/index_1599_seed_23461_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 02:57:24.246 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 02:57:41.895 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 02:57:42.949 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 02:57:43.075 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 02:57:43.840 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 02:59:37.354 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.38s (1.89min) +2026-06-19 02:59:38.043 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:01:37.731 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:08:37.489 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 7, train_steps: 280 +2026-06-19 03:08:38.147 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 03:08:38.280 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:08:38.316 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:08:38.404 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:09:44.163 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_2/index_6520_seed_31807_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 03:09:48.638 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 03:10:06.599 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:10:06.608 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:10:06.608 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:11:14.822 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_2/index_6520_seed_31808_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 03:11:19.844 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 03:11:37.542 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:11:38.337 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:11:38.379 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 03:11:39.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:13:34.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.68s (1.91min) +2026-06-19 03:13:34.977 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:15:32.043 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:22:31.154 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 8, train_steps: 320 +2026-06-19 03:22:31.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 03:22:31.954 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:22:31.986 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:22:32.075 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:23:38.767 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000008/group_2/index_2357_seed_14891_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 03:23:43.633 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 03:24:01.087 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:24:01.096 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:24:01.097 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:25:09.167 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000008/group_2/index_2357_seed_14892_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 03:25:14.078 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 03:25:32.939 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:25:33.100 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:25:33.113 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 03:25:33.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:27:28.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-19 03:27:29.431 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:29:26.904 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:36:25.651 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 9, train_steps: 360 +2026-06-19 03:36:26.346 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 03:36:26.562 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:36:26.673 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:36:26.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:37:32.836 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_2/index_2658_seed_4194_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 03:37:37.640 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 03:37:55.230 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:37:55.239 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:37:55.239 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:39:02.011 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_2/index_2658_seed_4195_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 03:39:06.328 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 03:39:24.571 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:39:26.010 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:39:26.063 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 03:39:26.998 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:41:20.400 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.36s (1.89min) +2026-06-19 03:41:21.105 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:43:17.562 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 03:50:16.937 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 10, train_steps: 400 +2026-06-19 03:50:53.427 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=10 (missing=0, unexpected=0) +2026-06-19 03:50:56.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 03:50:56.362 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 03:50:56.466 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:50:56.542 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:52:03.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_2/index_7369_seed_31263_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 03:52:06.863 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 03:52:24.710 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:52:24.719 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 03:52:24.720 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 03:53:32.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_2/index_7369_seed_31264_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 03:53:35.638 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 03:53:53.575 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 03:53:57.828 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 03:53:57.971 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 03:53:58.840 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 03:55:52.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.65s (1.89min) +2026-06-19 03:55:53.208 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 03:57:49.764 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:04:48.378 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 11, train_steps: 440 +2026-06-19 04:04:49.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 04:04:49.181 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:04:49.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:04:49.266 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:05:55.488 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_2/index_3046_seed_24442_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 04:05:58.427 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 04:06:16.348 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:06:16.358 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:06:16.360 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:07:24.731 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_2/index_3046_seed_24443_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 04:07:29.827 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 04:07:47.614 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:07:48.853 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:07:48.965 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 04:07:49.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:09:43.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.25s (1.89min) +2026-06-19 04:09:43.797 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:11:40.269 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:18:37.294 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 12, train_steps: 480 +2026-06-19 04:18:37.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 04:18:38.118 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:18:38.153 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:18:38.248 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:19:44.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_2/index_3336_seed_17787_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 04:19:49.074 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 04:20:06.575 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:20:06.584 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:20:06.584 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:21:12.812 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_2/index_3336_seed_17788_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 04:21:16.774 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 04:21:35.007 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:21:37.180 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:21:37.192 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 04:21:38.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:23:33.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.19s (1.92min) +2026-06-19 04:23:34.070 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:25:31.243 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:32:29.828 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 13, train_steps: 520 +2026-06-19 04:32:30.513 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 04:32:30.677 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:32:30.760 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:32:30.888 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:33:36.891 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_2/index_9044_seed_15998_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 04:33:41.922 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 04:33:59.335 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:33:59.344 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:33:59.344 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:35:06.120 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_2/index_9044_seed_15999_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 04:35:11.156 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 04:35:28.884 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:35:30.313 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:35:30.375 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 04:35:31.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:37:24.102 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.77s (1.88min) +2026-06-19 04:37:24.791 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:39:21.076 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 04:46:18.587 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 14, train_steps: 560 +2026-06-19 04:46:19.247 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 04:46:19.457 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 04:46:19.582 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:46:19.680 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:47:25.927 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_2/index_5690_seed_6957_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 04:47:30.640 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 04:47:48.578 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:47:48.584 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 04:47:48.584 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 04:48:54.223 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_2/index_5690_seed_6958_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 04:48:57.938 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 04:49:16.219 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 04:49:17.445 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 04:49:17.519 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 04:49:18.433 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 04:51:12.963 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.49s (1.91min) +2026-06-19 04:51:13.650 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 04:53:12.995 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:00:11.835 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 15, train_steps: 600 +2026-06-19 05:00:12.526 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 05:00:12.733 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:00:12.826 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:00:12.905 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:01:19.241 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000015/group_2/index_1031_seed_7112_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 05:01:24.352 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 05:01:42.717 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:01:42.727 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:01:42.728 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:02:47.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000015/group_2/index_1031_seed_7113_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 05:02:50.565 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 05:03:09.072 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:03:11.349 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:03:11.414 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 05:03:12.642 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:05:07.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.64s (1.91min) +2026-06-19 05:05:08.060 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:07:05.978 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:14:04.905 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 16, train_steps: 640 +2026-06-19 05:14:05.573 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 05:14:05.731 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:14:05.789 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:14:05.920 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:15:11.735 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_2/index_232_seed_28034_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 05:15:14.892 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 05:15:33.540 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:15:33.549 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:15:33.550 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:16:43.170 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_2/index_232_seed_28035_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 05:16:46.104 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 05:17:04.480 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:17:06.988 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:17:07.072 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 05:17:07.954 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:19:03.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.02s (1.92min) +2026-06-19 05:19:03.741 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:21:01.960 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:28:03.331 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 17, train_steps: 680 +2026-06-19 05:28:04.008 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 05:28:04.225 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:28:04.347 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:28:04.440 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:29:10.643 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_2/index_7006_seed_22036_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 05:29:13.419 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 05:29:33.252 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:29:33.261 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:29:33.263 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:30:39.578 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_2/index_7006_seed_22037_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 05:30:42.357 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 05:31:00.564 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:31:03.926 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:31:04.005 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 05:31:05.002 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:32:58.689 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.48s (1.89min) +2026-06-19 05:32:59.386 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:34:56.265 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:41:58.109 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 18, train_steps: 720 +2026-06-19 05:41:58.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 05:41:58.871 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:41:58.918 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:41:58.964 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:43:05.193 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_2/index_9880_seed_6168_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 05:43:10.290 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 05:43:28.445 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:43:28.455 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:43:28.455 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:44:35.446 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_2/index_9880_seed_6169_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 05:44:40.468 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 05:44:58.508 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:45:00.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:45:00.323 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 05:45:01.138 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 05:46:55.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.88s (1.90min) +2026-06-19 05:46:55.797 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 05:48:52.331 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 05:55:50.501 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 19, train_steps: 760 +2026-06-19 05:55:51.178 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 05:55:51.325 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 05:55:51.375 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:55:51.483 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:56:57.487 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_2/index_1225_seed_29891_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 05:57:01.478 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 05:57:19.766 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:57:19.775 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 05:57:19.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 05:58:25.563 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_2/index_1225_seed_29892_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 05:58:28.414 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 05:58:46.726 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 05:58:49.068 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 05:58:49.166 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 05:58:50.076 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:00:43.199 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.10s (1.89min) +2026-06-19 06:00:43.876 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:02:40.341 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:09:39.765 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 20, train_steps: 800 +2026-06-19 06:10:15.937 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=20 (missing=0, unexpected=0) +2026-06-19 06:10:18.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 06:10:18.938 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:10:19.041 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:10:19.122 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:11:26.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_2/index_8949_seed_21966_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 06:11:31.862 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 06:11:49.664 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:11:49.674 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:11:49.674 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:12:54.671 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_2/index_8949_seed_21967_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 06:12:58.626 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 06:13:16.849 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:13:19.050 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:13:19.112 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 06:13:20.148 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:15:13.224 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.90s (1.88min) +2026-06-19 06:15:13.910 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:17:12.689 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:24:12.573 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 21, train_steps: 840 +2026-06-19 06:24:13.271 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 06:24:13.464 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:24:13.559 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:24:13.636 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:25:19.568 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_2/index_6645_seed_27340_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 06:25:24.670 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 06:25:42.581 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:25:42.590 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:25:42.590 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:26:47.645 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_2/index_6645_seed_27341_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 06:26:51.728 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 06:27:10.138 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:27:11.261 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:27:11.309 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 06:27:12.330 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:29:06.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.51s (1.91min) +2026-06-19 06:29:07.667 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:31:04.581 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:38:04.296 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 22, train_steps: 880 +2026-06-19 06:38:04.947 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 06:38:05.042 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:38:05.087 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:38:05.135 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:39:11.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_2/index_3771_seed_13325_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 06:39:16.805 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 06:39:34.771 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:39:34.781 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:39:34.781 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:40:39.684 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_2/index_3771_seed_13326_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 06:40:44.597 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 06:41:02.526 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:41:03.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:41:03.236 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 06:41:04.251 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:42:58.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.99s (1.90min) +2026-06-19 06:42:59.164 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:44:56.506 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 06:51:58.182 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 23, train_steps: 920 +2026-06-19 06:51:58.842 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 06:51:59.012 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 06:51:59.091 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:51:59.200 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:53:04.984 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000023/group_2/index_525_seed_14536_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 06:53:09.272 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 06:53:27.647 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:53:27.656 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 06:53:27.656 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 06:54:35.025 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000023/group_2/index_525_seed_14537_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 06:54:37.855 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 06:54:56.317 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 06:54:59.789 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 06:54:59.880 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 06:55:00.785 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 06:56:54.935 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.11s (1.90min) +2026-06-19 06:56:55.621 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 06:58:52.319 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:05:51.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 24, train_steps: 960 +2026-06-19 07:05:52.005 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 07:05:52.174 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:05:52.264 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:05:52.386 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:06:58.438 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_2/index_3612_seed_21936_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 07:07:01.385 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 07:07:19.545 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:07:19.554 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:07:19.554 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:08:27.469 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_2/index_3612_seed_21937_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 07:08:32.602 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 07:08:50.857 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:08:52.217 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:08:52.330 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 07:08:53.136 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:10:46.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-19 07:10:46.871 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:12:44.153 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:19:41.983 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 25, train_steps: 1000 +2026-06-19 07:20:52.105 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 07:20:52.280 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:20:52.374 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:20:52.491 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:21:58.306 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_2/index_4482_seed_4896_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 07:22:02.507 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 07:22:20.569 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:22:20.578 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:22:20.578 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:23:28.646 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000025/group_2/index_4482_seed_4897_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 07:23:32.204 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 07:23:50.739 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:23:52.666 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:23:52.718 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 07:23:53.694 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:25:46.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.81s (1.88min) +2026-06-19 07:25:47.205 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:27:44.380 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:34:43.706 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 26, train_steps: 1040 +2026-06-19 07:34:44.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 07:34:44.524 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:34:44.597 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:34:44.715 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:35:50.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_2/index_4783_seed_28852_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 07:35:53.453 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 07:36:11.762 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:36:11.771 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:36:11.771 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:37:21.160 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_2/index_4783_seed_28853_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 07:37:26.341 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 07:37:44.663 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:37:45.443 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:37:45.487 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 07:37:46.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:39:39.638 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.14s (1.89min) +2026-06-19 07:39:40.348 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:41:39.072 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 07:48:39.682 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 27, train_steps: 1080 +2026-06-19 07:48:40.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 07:48:40.491 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 07:48:40.529 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:48:40.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:49:46.851 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_2/index_6487_seed_8818_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 07:49:51.679 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 07:50:10.155 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:50:10.164 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 07:50:10.164 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 07:51:19.422 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_2/index_6487_seed_8819_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 07:51:23.446 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 07:51:42.537 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 07:51:43.285 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 07:51:43.399 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 07:51:44.285 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 07:53:38.298 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.99s (1.90min) +2026-06-19 07:53:39.008 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 07:55:38.421 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:02:38.175 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 28, train_steps: 1120 +2026-06-19 08:02:38.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 08:02:39.042 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:02:39.116 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:02:39.226 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:03:45.264 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_2/index_6483_seed_856_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 08:03:50.389 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 08:04:08.851 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:04:08.860 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:04:08.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:05:17.416 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000028/group_2/index_6483_seed_857_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 08:05:21.591 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 08:05:40.298 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:05:41.901 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:05:42.012 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 08:05:43.114 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:07:37.396 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.24s (1.90min) +2026-06-19 08:07:38.092 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:09:37.934 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:16:37.860 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 29, train_steps: 1160 +2026-06-19 08:16:38.691 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 08:16:38.816 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:16:38.852 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:16:38.969 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:17:44.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_2/index_1141_seed_8955_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 08:17:48.071 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 08:18:06.649 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:18:06.658 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:18:06.658 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:19:18.020 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_2/index_1141_seed_8956_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 08:19:21.122 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 08:19:39.671 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:19:42.881 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:19:42.960 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 08:19:43.864 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:21:37.264 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 08:21:37.961 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:23:37.406 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:30:39.556 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 30, train_steps: 1200 +2026-06-19 08:31:15.795 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=30 (missing=0, unexpected=0) +2026-06-19 08:31:18.258 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 08:31:18.415 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:31:18.480 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:31:18.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:32:26.615 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000030/group_2/index_2816_seed_29602_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 08:32:29.696 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 08:32:48.247 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:32:48.257 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:32:48.257 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:33:55.818 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000030/group_2/index_2816_seed_29603_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 08:34:01.023 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 08:34:19.393 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:34:19.980 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:34:20.122 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 08:34:21.066 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:36:14.510 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-19 08:36:15.227 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:38:14.172 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:45:14.282 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 31, train_steps: 1240 +2026-06-19 08:45:14.994 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 08:45:15.211 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:45:15.343 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:45:15.421 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:46:21.854 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_2/index_9580_seed_33459_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 08:46:25.663 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 08:46:44.254 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:46:44.264 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:46:44.264 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 08:47:50.844 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_2/index_9580_seed_33460_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 08:47:53.861 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 08:48:12.557 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 08:48:15.288 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 08:48:15.300 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 08:48:16.474 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 08:50:10.323 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.84s (1.90min) +2026-06-19 08:50:11.024 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 08:52:11.149 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 08:59:13.101 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 32, train_steps: 1280 +2026-06-19 08:59:13.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 08:59:13.984 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 08:59:14.103 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 08:59:14.193 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:00:20.202 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000032/group_2/index_7683_seed_18951_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 09:00:25.357 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 09:00:43.253 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:00:43.262 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:00:43.262 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:01:48.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000032/group_2/index_7683_seed_18952_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 09:01:52.899 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 09:02:11.595 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:02:13.436 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:02:13.487 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 09:02:14.790 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:04:08.713 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.91s (1.90min) +2026-06-19 09:04:09.421 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:06:10.409 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:13:13.283 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 33, train_steps: 1320 +2026-06-19 09:13:13.953 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 09:13:14.155 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:13:14.264 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:13:14.398 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:14:20.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000033/group_2/index_7896_seed_16688_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 09:14:26.031 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 09:14:44.597 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:14:44.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:14:44.604 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:15:49.988 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000033/group_2/index_7896_seed_16689_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 09:15:54.972 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 09:16:14.681 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:16:14.822 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:16:14.866 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 09:16:16.003 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:18:11.066 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.06s (1.92min) +2026-06-19 09:18:11.786 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:20:11.627 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:27:13.736 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 34, train_steps: 1360 +2026-06-19 09:27:14.444 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 09:27:14.585 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:27:14.631 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:27:14.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:28:21.997 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_2/index_6618_seed_2867_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 09:28:27.040 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 09:28:44.981 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:28:44.990 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:28:44.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:29:50.008 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_2/index_6618_seed_2868_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 09:29:54.430 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 09:30:12.829 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:30:13.695 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:30:13.760 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 09:30:15.085 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:32:08.900 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.81s (1.90min) +2026-06-19 09:32:09.592 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:34:09.779 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:41:12.413 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 35, train_steps: 1400 +2026-06-19 09:41:13.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 09:41:13.152 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:41:13.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:41:13.239 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:42:19.325 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000035/group_2/index_1072_seed_23571_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 09:42:22.112 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 09:42:40.628 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:42:40.637 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:42:40.637 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:43:48.590 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000035/group_2/index_1072_seed_23572_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 09:43:51.558 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 09:44:10.391 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:44:13.080 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:44:13.155 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 09:44:14.080 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 09:46:10.686 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 116.56s (1.94min) +2026-06-19 09:46:11.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 09:48:10.896 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 09:55:11.968 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 36, train_steps: 1440 +2026-06-19 09:55:12.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 09:55:12.745 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 09:55:12.779 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:55:12.810 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:56:19.195 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_2/index_1691_seed_14139_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 09:56:22.116 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 09:56:40.515 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:56:40.524 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 09:56:40.525 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 09:57:48.008 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_2/index_1691_seed_14140_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 09:57:50.951 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 09:58:09.787 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 09:58:12.646 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 09:58:12.710 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 09:58:14.053 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:00:08.966 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-19 10:00:09.672 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:02:10.007 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:09:11.711 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 37, train_steps: 1480 +2026-06-19 10:09:12.395 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 10:09:12.504 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:09:12.557 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:09:12.607 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:10:18.713 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_2/index_2341_seed_30959_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 10:10:23.840 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 10:10:41.731 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:10:41.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:10:41.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:11:47.718 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_2/index_2341_seed_30960_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 10:11:51.921 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 10:12:10.237 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:12:13.042 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:12:13.137 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 10:12:14.079 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:14:09.030 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.94s (1.92min) +2026-06-19 10:14:09.737 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:16:10.367 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:23:12.352 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 38, train_steps: 1520 +2026-06-19 10:23:13.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 10:23:13.257 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:23:13.364 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:23:13.457 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:24:19.273 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000038/group_2/index_8078_seed_11886_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 10:24:23.241 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 10:24:41.627 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:24:41.637 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:24:41.637 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:25:48.703 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000038/group_2/index_8078_seed_11887_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 10:25:51.565 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 10:26:10.705 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:26:12.667 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:26:12.717 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 10:26:14.064 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:28:09.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.16s (1.92min) +2026-06-19 10:28:09.955 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:30:10.743 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:37:11.955 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 39, train_steps: 1560 +2026-06-19 10:37:12.646 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 10:37:12.752 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:37:12.791 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:37:12.838 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:38:19.137 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000039/group_2/index_440_seed_8373_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 10:38:22.129 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 10:38:40.887 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:38:40.897 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:38:40.899 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:41:17.592 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000039/group_2/index_440_seed_8374_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 10:41:20.531 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 10:41:39.406 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:41:42.526 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:41:42.596 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 10:41:43.737 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:43:36.471 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.70s (1.88min) +2026-06-19 10:43:37.180 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 10:45:37.175 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 10:52:37.870 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 40, train_steps: 1600 +2026-06-19 10:53:13.912 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=40 (missing=0, unexpected=0) +2026-06-19 10:53:15.773 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 10:53:15.926 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 10:53:16.002 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:53:16.108 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:54:23.736 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_2/index_4793_seed_25720_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 10:54:26.771 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 10:54:44.661 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:54:44.671 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 10:54:44.672 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 10:55:52.539 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_2/index_4793_seed_25721_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 10:55:55.471 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 10:56:14.703 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 10:56:16.410 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 10:56:16.422 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 10:56:17.404 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 10:58:11.717 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.27s (1.90min) +2026-06-19 10:58:12.416 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:00:11.328 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:07:12.594 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 41, train_steps: 1640 +2026-06-19 11:07:13.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 11:07:13.494 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:07:13.633 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:07:13.733 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:08:19.771 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_2/index_8743_seed_15788_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 11:08:22.676 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 11:08:40.702 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:08:40.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:08:40.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:09:48.290 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_2/index_8743_seed_15789_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 11:09:51.184 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 11:10:09.402 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:10:12.668 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:10:12.782 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 11:10:14.040 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:12:08.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.28s (1.90min) +2026-06-19 11:12:09.087 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:14:08.758 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:21:10.900 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 42, train_steps: 1680 +2026-06-19 11:21:11.805 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 11:21:12.013 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:21:12.123 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:21:12.242 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:22:18.338 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000042/group_2/index_4621_seed_17935_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 11:22:21.205 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 11:22:39.549 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:22:39.558 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:22:39.558 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:23:47.502 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000042/group_2/index_4621_seed_17936_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 11:23:52.641 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 11:24:10.756 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:24:12.073 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:24:12.175 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 11:24:13.174 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:26:07.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.32s (1.91min) +2026-06-19 11:26:08.279 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:28:06.798 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:35:06.782 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 43, train_steps: 1720 +2026-06-19 11:35:07.458 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 11:35:07.617 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:35:07.674 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:35:07.807 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:36:13.717 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_2/index_1909_seed_12454_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 11:36:18.052 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 11:36:35.851 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:36:35.860 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:36:35.860 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:37:46.265 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000043/group_2/index_1909_seed_12455_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 11:37:49.005 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 11:38:07.276 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:38:10.299 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:38:10.387 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 11:38:11.318 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:40:05.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.38s (1.91min) +2026-06-19 11:40:06.548 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:42:03.524 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 11:49:04.444 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 44, train_steps: 1760 +2026-06-19 11:49:05.114 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 11:49:05.285 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 11:49:05.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:49:05.496 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:50:11.399 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000044/group_2/index_6995_seed_28986_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 11:50:14.594 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 11:50:33.700 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:50:33.711 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 11:50:33.711 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 11:51:41.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000044/group_2/index_6995_seed_28987_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 11:51:46.909 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 11:52:05.288 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 11:52:06.491 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 11:52:06.605 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 11:52:07.489 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 11:54:01.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.69s (1.89min) +2026-06-19 11:54:01.910 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 11:56:01.235 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:03:03.882 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 45, train_steps: 1800 +2026-06-19 12:03:04.557 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 12:03:04.719 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:03:04.789 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:03:04.909 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:04:11.061 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000045/group_2/index_6985_seed_17085_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 12:04:14.015 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 12:04:33.527 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:04:33.537 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:04:33.537 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:05:40.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000045/group_2/index_6985_seed_17086_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 12:05:45.557 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 12:06:03.645 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:06:05.347 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:06:05.437 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 12:06:06.287 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:08:01.417 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.05s (1.92min) +2026-06-19 12:08:02.101 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:10:01.382 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:17:03.593 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 46, train_steps: 1840 +2026-06-19 12:17:04.281 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 12:17:04.495 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:17:04.605 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:17:04.699 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:18:10.660 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_2/index_2294_seed_3872_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 12:18:15.862 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 12:18:33.637 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:18:33.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:18:33.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:19:40.685 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_2/index_2294_seed_3873_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 12:19:44.816 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 12:20:02.874 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:20:05.333 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:20:05.474 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 12:20:06.544 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:22:01.733 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.17s (1.92min) +2026-06-19 12:22:02.442 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:24:01.450 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:31:02.941 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 47, train_steps: 1880 +2026-06-19 12:31:03.607 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 12:31:03.755 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:31:03.813 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:31:03.945 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:32:09.783 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000047/group_2/index_3711_seed_5432_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 12:32:12.594 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 12:32:30.770 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:32:30.780 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:32:30.780 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:33:39.421 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000047/group_2/index_3711_seed_5433_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 12:33:42.406 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 12:34:00.623 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:34:03.182 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:34:03.289 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 12:34:04.337 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:35:58.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-19 12:35:58.982 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:37:55.891 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:44:57.685 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 48, train_steps: 1920 +2026-06-19 12:44:58.355 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 12:44:58.469 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:44:58.511 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:44:58.583 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:46:04.674 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_2/index_3448_seed_23846_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 12:46:07.526 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 12:46:25.470 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:46:25.480 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:46:25.480 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:47:34.727 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_2/index_3448_seed_23847_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 12:47:38.752 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 12:47:56.785 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 12:47:58.931 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 12:47:58.996 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 12:47:59.946 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 12:49:54.620 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-19 12:49:55.344 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 12:51:52.290 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 12:58:52.102 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 49, train_steps: 1960 +2026-06-19 12:58:52.782 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 12:58:52.877 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 12:58:52.926 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 12:58:52.972 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 12:59:58.950 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_2/index_2666_seed_31483_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 13:00:01.715 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 13:00:19.620 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:00:19.630 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:00:19.630 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:01:30.206 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000049/group_2/index_2666_seed_31484_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 13:01:33.267 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 13:01:51.450 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:01:54.369 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:01:54.479 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 13:01:55.319 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:03:48.446 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.04s (1.88min) +2026-06-19 13:03:49.149 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:05:45.467 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:12:45.277 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 50, train_steps: 2000 +2026-06-19 13:14:21.235 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=50 (missing=0, unexpected=0) +2026-06-19 13:14:23.203 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 13:14:23.420 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:14:23.523 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:14:23.604 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:15:31.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_2/index_3304_seed_27065_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 13:15:36.564 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 13:15:54.341 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:15:54.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:15:54.350 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:17:00.848 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_2/index_3304_seed_27066_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 13:17:04.574 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 13:17:22.868 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:17:25.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:17:25.320 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 13:17:26.227 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:19:20.081 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.84s (1.90min) +2026-06-19 13:19:20.807 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:21:17.524 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:28:18.473 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 51, train_steps: 2040 +2026-06-19 13:28:19.185 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 13:28:19.334 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:28:19.388 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:28:19.498 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:29:25.502 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000051/group_2/index_2709_seed_28347_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 13:29:30.686 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 13:29:48.688 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:29:48.697 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:29:48.697 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:30:54.159 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000051/group_2/index_2709_seed_28348_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 13:30:58.246 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 13:31:16.753 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:31:18.029 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:31:18.110 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 13:31:19.096 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:33:13.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.83s (1.90min) +2026-06-19 13:33:13.741 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:35:13.227 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:42:13.632 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 52, train_steps: 2080 +2026-06-19 13:42:14.276 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 13:42:14.401 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:42:14.437 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:42:14.503 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:43:21.433 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_2/index_3650_seed_12960_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 13:43:26.498 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 13:43:44.715 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:43:44.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:43:44.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:44:50.169 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000052/group_2/index_3650_seed_12961_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 13:44:55.113 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 13:45:13.127 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:45:15.287 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:45:15.406 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 13:45:16.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 13:47:10.481 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.03s (1.90min) +2026-06-19 13:47:11.196 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 13:49:11.264 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 13:56:12.606 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 53, train_steps: 2120 +2026-06-19 13:56:13.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 13:56:13.496 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 13:56:13.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:56:13.695 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:57:19.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_2/index_3556_seed_24404_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 13:57:24.471 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 13:57:42.000 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:57:42.008 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 13:57:42.008 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 13:58:48.763 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_2/index_3556_seed_24405_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 13:58:52.851 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 13:59:11.127 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 13:59:13.130 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 13:59:13.237 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 13:59:14.385 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:01:08.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.41s (1.91min) +2026-06-19 14:01:09.529 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:03:10.472 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:10:12.174 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 54, train_steps: 2160 +2026-06-19 14:10:12.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 14:10:13.019 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:10:13.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:10:13.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:11:19.129 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_2/index_6888_seed_29256_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 14:11:22.027 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 14:11:40.349 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:11:40.359 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:11:40.359 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:12:47.768 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_2/index_6888_seed_29257_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 14:12:50.710 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 14:13:09.058 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:13:12.866 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:13:12.961 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 14:13:14.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:15:08.493 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.42s (1.91min) +2026-06-19 14:15:09.218 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:17:09.304 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:24:11.918 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 55, train_steps: 2200 +2026-06-19 14:24:12.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 14:24:12.761 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:24:12.796 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:24:12.883 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:25:19.328 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_2/index_4222_seed_11610_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 14:25:22.262 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 14:25:40.525 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:25:40.535 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:25:40.535 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:26:48.263 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_2/index_4222_seed_11611_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 14:26:51.241 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 14:27:09.784 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:27:12.754 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:27:12.862 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 14:27:14.039 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:29:09.888 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.82s (1.93min) +2026-06-19 14:29:10.604 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:31:10.762 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:38:12.129 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 56, train_steps: 2240 +2026-06-19 14:38:12.810 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 14:38:12.905 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:38:12.947 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:38:12.995 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:39:19.444 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_2/index_5858_seed_11650_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 14:39:24.587 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 14:39:42.277 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:39:42.286 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:39:42.286 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:40:48.782 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_2/index_5858_seed_11651_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 14:40:53.695 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 14:41:11.913 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:41:13.092 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:41:13.199 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 14:41:14.064 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:43:08.781 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.66s (1.91min) +2026-06-19 14:43:09.475 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:45:09.716 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 14:52:11.858 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 57, train_steps: 2280 +2026-06-19 14:52:12.526 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 14:52:12.690 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 14:52:12.754 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:52:12.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:53:18.870 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_2/index_6957_seed_16210_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 14:53:23.165 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 14:53:41.107 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:53:41.116 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 14:53:41.116 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 14:54:48.454 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_2/index_6957_seed_16211_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 14:54:53.515 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 14:55:11.226 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 14:55:13.057 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 14:55:13.186 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 14:55:14.036 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 14:57:08.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.79s (1.91min) +2026-06-19 14:57:09.597 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 14:59:09.879 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:06:12.298 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 58, train_steps: 2320 +2026-06-19 15:06:12.986 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 15:06:13.189 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:06:13.295 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:06:13.391 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:07:19.925 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_2/index_5844_seed_14693_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 15:07:24.978 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 15:07:42.958 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:07:42.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:07:42.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:08:48.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_2/index_5844_seed_14694_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 15:08:53.032 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 15:09:11.395 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:09:13.062 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:09:13.146 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 15:09:14.091 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:11:08.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.71s (1.91min) +2026-06-19 15:11:09.498 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:13:09.811 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:20:12.064 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 59, train_steps: 2360 +2026-06-19 15:20:12.739 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 15:20:12.954 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:20:13.084 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:20:13.163 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:21:19.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_2/index_6912_seed_11777_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 15:21:24.401 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 15:21:42.053 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:21:42.063 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:21:42.063 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:22:47.973 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_2/index_6912_seed_11778_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 15:22:52.064 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 15:23:10.395 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:23:12.372 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:23:12.473 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 15:23:13.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:25:09.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.28s (1.92min) +2026-06-19 15:25:09.742 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:27:10.365 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:34:12.083 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 60, train_steps: 2400 +2026-06-19 15:34:48.527 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=60 (missing=0, unexpected=0) +2026-06-19 15:34:50.616 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 15:34:50.804 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:34:50.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:34:50.980 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:35:58.602 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_2/index_2967_seed_24369_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 15:36:01.482 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 15:36:19.325 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:36:19.335 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:36:19.335 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:37:28.408 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_2/index_2967_seed_24370_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 15:37:33.618 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 15:37:51.746 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:37:53.529 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:37:53.567 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 15:37:54.683 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:39:47.707 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-19 15:39:48.403 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:41:45.234 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 15:48:44.582 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 61, train_steps: 2440 +2026-06-19 15:48:45.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 15:48:45.490 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 15:48:45.593 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:48:45.692 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:49:51.581 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_2/index_4901_seed_15668_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 15:49:55.923 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 15:50:15.629 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:50:15.636 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 15:50:15.636 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 15:51:23.692 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_2/index_4901_seed_15669_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 15:51:26.584 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 15:51:44.879 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 15:51:47.609 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 15:51:47.683 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 15:51:48.614 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 15:53:41.681 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.04s (1.88min) +2026-06-19 15:53:42.400 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 15:55:39.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:02:39.029 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 62, train_steps: 2480 +2026-06-19 16:02:39.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 16:02:39.811 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:02:39.856 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:02:39.931 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:03:45.831 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000062/group_2/index_7307_seed_16798_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 16:03:50.947 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 16:04:08.988 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:04:08.998 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:04:08.998 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:05:18.832 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000062/group_2/index_7307_seed_16799_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 16:05:22.549 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 16:05:41.022 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:05:44.132 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:05:44.221 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 16:05:45.152 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:07:38.761 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.59s (1.89min) +2026-06-19 16:07:39.471 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:09:38.853 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:16:39.330 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 63, train_steps: 2520 +2026-06-19 16:16:40.032 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 16:16:40.211 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:16:40.289 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:16:40.426 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:17:46.706 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_2/index_2370_seed_24408_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 16:17:49.714 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 16:18:07.716 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:18:07.725 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:18:07.725 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:19:19.873 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_2/index_2370_seed_24409_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 16:19:24.849 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 16:19:43.415 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:19:45.197 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:19:45.298 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 16:19:46.202 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:21:39.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.36s (1.89min) +2026-06-19 16:21:40.302 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:23:38.880 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:30:40.941 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 64, train_steps: 2560 +2026-06-19 16:30:41.626 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 16:30:41.816 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:30:41.899 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:30:41.999 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:31:48.105 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_2/index_3081_seed_6313_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 16:31:51.079 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 16:32:09.136 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:32:09.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:32:09.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:33:21.362 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_2/index_3081_seed_6314_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 16:33:26.538 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 16:33:44.528 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:33:45.471 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:33:45.587 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 16:33:46.416 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:35:40.144 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.63s (1.89min) +2026-06-19 16:35:40.844 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:37:39.399 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:44:41.307 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 65, train_steps: 2600 +2026-06-19 16:44:41.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 16:44:42.071 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:44:42.105 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:44:42.137 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:45:48.528 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_2/index_4872_seed_28177_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 16:45:53.344 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 16:46:11.195 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:46:11.204 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:46:11.204 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:47:23.189 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_2/index_4872_seed_28178_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 16:47:28.087 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 16:47:46.130 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 16:47:47.482 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 16:47:47.552 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 16:47:48.434 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 16:49:42.246 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.72s (1.90min) +2026-06-19 16:49:42.958 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 16:51:40.035 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 16:58:41.285 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 66, train_steps: 2640 +2026-06-19 16:58:41.990 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 16:58:42.147 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 16:58:42.214 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 16:58:42.338 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 16:59:48.247 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_2/index_2971_seed_14970_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 16:59:53.121 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 17:00:11.086 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:00:11.095 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:00:11.096 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:01:19.521 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_2/index_2971_seed_14971_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 17:01:23.738 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 17:01:42.001 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:01:43.599 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:01:43.686 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 17:01:44.600 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:03:37.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.34s (1.89min) +2026-06-19 17:03:38.681 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:05:38.382 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:12:40.131 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 67, train_steps: 2680 +2026-06-19 17:12:40.908 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 17:12:41.288 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:12:41.447 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:12:41.571 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:13:47.639 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_2/index_7348_seed_30890_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 17:13:52.676 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 17:14:10.464 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:14:10.473 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:14:10.474 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:15:20.785 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_2/index_7348_seed_30891_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 17:15:25.188 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 17:15:43.426 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:15:44.529 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:15:44.596 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 17:15:45.582 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:17:38.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 17:17:39.679 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:19:39.113 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:26:39.553 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 68, train_steps: 2720 +2026-06-19 17:26:40.215 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 17:26:40.398 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:26:40.523 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:26:40.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:27:47.066 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_2/index_4589_seed_2372_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 17:27:52.146 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 17:28:09.906 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:28:09.916 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:28:09.916 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:29:20.123 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_2/index_4589_seed_2373_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 17:29:24.986 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 17:29:42.983 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:29:43.886 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:29:43.989 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 17:29:44.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:31:38.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.76s (1.90min) +2026-06-19 17:31:39.420 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:33:38.458 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:40:38.369 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 69, train_steps: 2760 +2026-06-19 17:40:39.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 17:40:39.215 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:40:39.316 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:40:39.426 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:41:45.101 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000069/group_2/index_4630_seed_15329_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 17:41:49.962 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 17:42:08.066 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:42:08.074 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:42:08.074 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:43:16.977 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000069/group_2/index_4630_seed_15330_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 17:43:21.038 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 17:43:39.711 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:43:41.069 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:43:41.082 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 17:43:42.165 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 17:45:35.854 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.55s (1.89min) +2026-06-19 17:45:36.569 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 17:47:37.117 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 17:54:38.586 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 70, train_steps: 2800 +2026-06-19 17:55:15.011 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=70 (missing=0, unexpected=0) +2026-06-19 17:55:16.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 17:55:17.055 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 17:55:17.071 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:55:17.132 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:56:25.244 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_2/index_1514_seed_25683_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 17:56:28.185 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 17:56:46.183 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:56:46.193 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 17:56:46.193 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 17:57:54.713 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_2/index_1514_seed_25684_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 17:58:00.039 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 17:58:18.221 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 17:58:18.228 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 17:58:18.241 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 17:58:19.407 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:00:13.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.58s (1.89min) +2026-06-19 18:00:13.838 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:02:13.851 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:09:13.905 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 71, train_steps: 2840 +2026-06-19 18:09:14.615 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 18:09:14.750 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:09:14.785 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:09:14.887 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:10:21.452 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_2/index_9759_seed_27453_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 18:10:25.778 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 18:10:43.869 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:10:43.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:10:43.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:11:51.762 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_2/index_9759_seed_27454_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 18:11:56.908 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 18:12:16.007 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:12:16.021 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:12:16.034 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 18:12:16.948 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:14:12.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.67s (1.93min) +2026-06-19 18:14:13.460 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:16:12.695 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:23:13.833 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 72, train_steps: 2880 +2026-06-19 18:23:14.506 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 18:23:14.724 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:23:14.839 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:23:14.951 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:24:21.978 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_2/index_8525_seed_27800_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 18:24:25.924 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 18:24:44.288 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:24:44.294 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:24:44.295 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:25:51.278 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_2/index_8525_seed_27801_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 18:25:54.342 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 18:26:12.700 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:26:14.711 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:26:14.766 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 18:26:15.980 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:28:10.439 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.45s (1.91min) +2026-06-19 18:28:11.145 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:30:10.988 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:37:12.798 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 73, train_steps: 2920 +2026-06-19 18:37:13.456 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 18:37:13.630 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:37:13.714 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:37:13.846 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:38:20.194 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_2/index_4113_seed_11490_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 18:38:25.391 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 18:38:43.321 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:38:43.330 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:38:43.330 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:39:50.035 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000073/group_2/index_4113_seed_11491_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 18:39:54.333 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 18:40:12.541 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:40:13.250 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:40:13.334 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 18:40:14.368 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:42:09.574 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.16s (1.92min) +2026-06-19 18:42:10.285 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:44:10.447 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 18:51:12.756 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 74, train_steps: 2960 +2026-06-19 18:51:13.469 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 18:51:13.683 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 18:51:13.811 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:51:13.908 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:52:20.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_2/index_9527_seed_23855_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 18:52:25.313 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 18:52:43.208 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:52:43.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 18:52:43.220 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 18:53:49.771 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_2/index_9527_seed_23856_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 18:53:54.951 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 18:54:13.142 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 18:54:13.982 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 18:54:14.120 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 18:54:15.065 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 18:56:10.312 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.21s (1.92min) +2026-06-19 18:56:11.030 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 18:58:11.442 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:05:14.020 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 75, train_steps: 3000 +2026-06-19 19:06:23.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 19:06:23.418 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:06:23.531 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:06:23.636 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:07:30.035 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000075/group_2/index_9613_seed_4576_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 19:07:35.145 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 19:07:52.650 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:07:52.659 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:07:52.659 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:08:59.098 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000075/group_2/index_9613_seed_4577_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 19:09:04.049 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 19:09:22.004 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:09:24.209 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:09:24.310 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 19:09:25.170 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:11:18.397 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.14s (1.89min) +2026-06-19 19:11:19.100 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:13:16.253 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:20:16.321 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 76, train_steps: 3040 +2026-06-19 19:20:16.989 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 19:20:17.096 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:20:17.139 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:20:17.187 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:21:23.730 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000076/group_2/index_8476_seed_7778_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 19:21:27.814 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 19:21:46.007 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:21:46.016 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:21:46.016 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:22:53.675 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000076/group_2/index_8476_seed_7779_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 19:22:56.613 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 19:23:15.607 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:23:18.455 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:23:18.548 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 19:23:19.632 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:25:13.441 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.66s (1.89min) +2026-06-19 19:25:14.145 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:27:12.733 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:34:13.271 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 77, train_steps: 3080 +2026-06-19 19:34:13.951 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 19:34:14.178 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:34:14.318 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:34:14.397 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:35:20.380 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_2/index_3647_seed_9883_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 19:35:23.367 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 19:35:41.109 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:35:41.118 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:35:41.118 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:36:49.228 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_2/index_3647_seed_9884_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 19:36:54.386 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 19:37:12.107 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:37:13.074 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:37:13.183 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 19:37:14.095 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:39:08.734 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.60s (1.91min) +2026-06-19 19:39:09.441 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:41:09.769 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 19:48:12.653 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 78, train_steps: 3120 +2026-06-19 19:48:13.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 19:48:13.515 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 19:48:13.559 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:48:13.656 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:49:19.883 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_2/index_5772_seed_9759_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 19:49:23.893 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 19:49:41.680 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:49:41.688 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 19:49:41.688 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 19:50:49.856 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_2/index_5772_seed_9760_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 19:50:55.053 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 19:51:12.934 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 19:51:13.323 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 19:51:13.387 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 19:51:14.398 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 19:53:09.267 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-19 19:53:09.975 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 19:55:09.985 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:02:13.179 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 79, train_steps: 3160 +2026-06-19 20:02:13.885 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 20:02:14.090 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:02:14.213 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:02:14.307 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:03:20.358 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000079/group_2/index_2984_seed_19942_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 20:03:25.182 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 20:03:42.706 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:03:42.715 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:03:42.715 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:04:49.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000079/group_2/index_2984_seed_19943_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 20:04:54.650 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 20:05:12.432 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:05:13.093 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:05:13.199 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 20:05:14.015 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:07:09.660 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.53s (1.93min) +2026-06-19 20:07:10.385 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:09:11.032 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:16:13.942 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 80, train_steps: 3200 +2026-06-19 20:16:51.256 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=80 (missing=0, unexpected=0) +2026-06-19 20:16:53.093 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 20:16:53.220 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:16:53.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:16:53.447 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:18:01.767 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_2/index_9548_seed_7064_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 20:18:06.619 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 20:18:24.298 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:18:24.307 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:18:24.307 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:19:30.788 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_2/index_9548_seed_7065_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 20:19:35.354 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 20:19:53.657 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:19:54.752 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:19:54.849 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 20:19:55.776 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:21:49.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-19 20:21:50.363 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:23:47.274 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:30:47.438 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 81, train_steps: 3240 +2026-06-19 20:30:48.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 20:30:48.310 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:30:48.372 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:30:48.534 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:31:54.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_2/index_6516_seed_25872_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 20:31:57.724 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 20:32:16.049 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:32:16.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:32:16.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:33:25.163 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_2/index_6516_seed_25873_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 20:33:28.170 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 20:33:46.965 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:33:49.273 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:33:49.377 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 20:33:50.267 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:35:44.031 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.70s (1.89min) +2026-06-19 20:35:44.752 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:37:41.676 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:44:40.010 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 82, train_steps: 3280 +2026-06-19 20:44:40.696 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 20:44:40.872 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:44:40.953 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:44:41.103 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:45:47.205 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_2/index_859_seed_26490_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 20:45:52.204 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 20:46:09.877 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:46:09.886 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:46:09.886 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:47:20.143 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_2/index_859_seed_26491_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 20:47:25.013 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 20:47:42.899 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 20:47:44.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 20:47:44.264 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 20:47:45.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 20:49:38.579 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.41s (1.89min) +2026-06-19 20:49:39.275 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 20:51:38.423 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 20:58:39.011 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 83, train_steps: 3320 +2026-06-19 20:58:39.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 20:58:40.035 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 20:58:40.137 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 20:58:40.237 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 20:59:46.036 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_2/index_7917_seed_11920_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 20:59:50.038 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 21:00:08.282 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:00:08.291 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:00:08.291 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:01:19.928 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_2/index_7917_seed_11921_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 21:01:25.187 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 21:01:42.832 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:01:44.304 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:01:44.414 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 21:01:45.227 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:03:38.627 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.28s (1.89min) +2026-06-19 21:03:39.339 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:05:38.817 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:12:38.927 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 84, train_steps: 3360 +2026-06-19 21:12:39.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 21:12:39.776 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:12:39.819 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:12:39.896 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:13:45.940 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_2/index_3272_seed_23637_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 21:13:50.893 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 21:14:08.635 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:14:08.644 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:14:08.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:15:19.620 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_2/index_3272_seed_23638_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 21:15:23.629 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 21:15:43.476 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:15:43.791 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:15:43.896 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 21:15:44.817 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:17:38.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.31s (1.89min) +2026-06-19 21:17:38.864 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:19:38.522 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:26:39.507 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 85, train_steps: 3400 +2026-06-19 21:26:40.167 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 21:26:40.267 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:26:40.309 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:26:40.358 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:27:47.054 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_2/index_8916_seed_12205_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 21:27:50.060 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 21:28:08.317 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:28:08.327 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:28:08.327 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:29:20.170 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_2/index_8916_seed_12206_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 21:29:23.235 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 21:29:41.733 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:29:43.697 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:29:43.811 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 21:29:44.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:31:38.268 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-19 21:31:38.988 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:33:38.755 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:40:38.659 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 86, train_steps: 3440 +2026-06-19 21:40:39.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 21:40:39.475 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:40:39.524 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:40:39.575 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:41:46.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_2/index_1907_seed_6791_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 21:41:49.175 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 21:42:07.550 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:42:07.560 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:42:07.561 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:43:18.762 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_2/index_1907_seed_6792_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 21:43:23.959 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 21:43:41.969 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:43:42.867 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:43:42.987 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 21:43:43.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:45:37.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 21:45:37.987 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 21:47:37.708 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 21:54:39.048 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 87, train_steps: 3480 +2026-06-19 21:54:39.746 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 21:54:39.973 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 21:54:40.111 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:54:40.176 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:55:45.908 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000087/group_2/index_6478_seed_21462_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 21:55:49.877 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 21:56:07.961 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:56:07.971 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 21:56:07.971 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 21:57:19.089 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000087/group_2/index_6478_seed_21463_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 21:57:21.958 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 21:57:40.342 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 21:57:43.169 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 21:57:43.290 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 21:57:44.172 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 21:59:37.604 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.39s (1.89min) +2026-06-19 21:59:38.317 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:01:38.553 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:08:39.998 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 88, train_steps: 3520 +2026-06-19 22:08:40.670 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 22:08:40.767 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:08:40.806 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:08:40.833 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:09:46.993 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_2/index_4397_seed_24354_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 22:09:52.201 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 22:10:09.887 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:10:09.898 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:10:09.898 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:11:19.522 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_2/index_4397_seed_24355_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 22:11:23.749 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 22:11:43.350 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:11:43.755 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:11:43.792 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 22:11:44.789 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:13:38.390 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.57s (1.89min) +2026-06-19 22:13:39.103 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:15:38.303 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:22:38.711 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 89, train_steps: 3560 +2026-06-19 22:22:39.424 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 22:22:39.562 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:22:39.605 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:22:39.709 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:23:45.563 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_2/index_2758_seed_20884_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 22:23:48.547 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 22:24:08.238 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:24:08.244 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:24:08.244 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:25:19.655 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_2/index_2758_seed_20885_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 22:25:22.628 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 22:25:42.697 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:25:43.683 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:25:43.791 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 22:25:44.762 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:27:38.162 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-19 22:27:38.886 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:29:38.493 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:36:39.432 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 90, train_steps: 3600 +2026-06-19 22:37:15.715 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=90 (missing=0, unexpected=0) +2026-06-19 22:37:17.308 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 22:37:17.521 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:37:17.663 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:37:17.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:38:25.029 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_2/index_7201_seed_22011_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 22:38:30.224 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 22:38:49.766 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:38:49.775 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:38:49.775 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:39:55.004 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_2/index_7201_seed_22012_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 22:39:59.135 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 22:40:19.226 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:40:19.240 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:40:19.253 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 22:40:20.293 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:42:14.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.70s (1.89min) +2026-06-19 22:42:14.863 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:44:13.687 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 22:51:14.864 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 91, train_steps: 3640 +2026-06-19 22:51:15.700 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 22:51:15.883 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 22:51:15.963 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:51:16.080 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:52:23.025 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000091/group_2/index_2829_seed_22675_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 22:52:25.988 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 22:52:44.019 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:52:44.029 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 22:52:44.029 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 22:53:51.959 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000091/group_2/index_2829_seed_22676_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 22:53:55.172 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 22:54:13.470 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 22:54:15.937 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 22:54:15.950 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 22:54:16.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 22:56:12.672 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.67s (1.93min) +2026-06-19 22:56:13.388 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 22:58:13.111 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:05:14.268 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 92, train_steps: 3680 +2026-06-19 23:05:15.118 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 23:05:15.337 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:05:15.463 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:05:15.550 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:06:22.210 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000092/group_2/index_5467_seed_30570_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 23:06:27.362 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 23:06:45.630 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:06:45.640 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:06:45.640 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:07:51.238 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000092/group_2/index_5467_seed_30571_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 23:07:55.231 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 23:08:14.370 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:08:16.058 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:08:16.097 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 23:08:17.102 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:10:11.751 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.62s (1.91min) +2026-06-19 23:10:12.471 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:12:11.281 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:19:13.532 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 93, train_steps: 3720 +2026-06-19 23:19:14.151 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 23:19:14.284 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:19:14.315 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:19:14.382 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:20:21.527 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_2/index_251_seed_28922_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 23:20:26.670 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 23:20:45.019 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:20:45.028 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:20:45.028 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:21:50.135 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_2/index_251_seed_28923_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 23:21:55.004 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 23:22:13.234 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:22:14.354 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:22:14.463 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 23:22:15.642 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:24:10.991 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.33s (1.92min) +2026-06-19 23:24:11.707 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:26:11.287 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:33:13.951 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 94, train_steps: 3760 +2026-06-19 23:33:14.640 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 23:33:14.812 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:33:14.893 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:33:15.010 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:34:21.806 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_2/index_1248_seed_19189_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 23:34:25.831 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 23:34:44.638 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:34:44.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:34:44.649 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:35:51.703 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_2/index_1248_seed_19190_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 23:35:56.962 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 23:36:16.052 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:36:16.136 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:36:16.147 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 23:36:17.075 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:38:12.062 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.89s (1.91min) +2026-06-19 23:38:12.784 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:40:11.319 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-19 23:47:11.876 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 95, train_steps: 3800 +2026-06-19 23:47:12.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-19 23:47:12.794 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-19 23:47:12.887 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:47:12.981 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:48:18.824 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_2/index_8782_seed_7214_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-19 23:48:23.567 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 23:48:41.669 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:48:41.678 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-19 23:48:41.679 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-19 23:49:48.200 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_2/index_8782_seed_7215_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-19 23:49:52.459 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-19 23:50:10.822 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-19 23:50:12.004 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-19 23:50:12.095 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-19 23:50:13.247 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-19 23:52:08.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.82s (1.91min) +2026-06-19 23:52:08.791 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-19 23:54:07.676 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:01:09.782 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 96, train_steps: 3840 +2026-06-20 00:01:10.475 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 00:01:10.599 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:01:10.643 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:01:10.850 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:02:17.084 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_2/index_8559_seed_15971_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 00:02:20.073 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 00:02:38.725 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:02:38.734 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:02:38.734 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:03:47.569 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_2/index_8559_seed_15972_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 00:03:52.646 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 00:04:10.815 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:04:11.783 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:04:11.905 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 00:04:12.897 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:06:07.927 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 00:06:08.644 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:08:07.688 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:15:08.509 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 97, train_steps: 3880 +2026-06-20 00:15:09.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 00:15:09.436 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:15:09.583 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:15:09.660 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:16:15.325 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_2/index_107_seed_24925_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 00:16:20.334 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 00:16:38.163 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:16:38.172 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:16:38.173 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:17:47.316 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_2/index_107_seed_24926_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 00:17:51.273 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 00:18:09.715 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:18:11.238 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:18:11.273 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 00:18:12.354 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:20:07.671 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.20s (1.92min) +2026-06-20 00:20:08.401 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:22:07.125 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:29:07.013 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 98, train_steps: 3920 +2026-06-20 00:29:07.692 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 00:29:07.851 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:29:07.909 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:29:08.024 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:30:14.155 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000098/group_2/index_1056_seed_31265_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 00:30:17.376 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 00:30:35.952 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:30:35.961 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:30:35.962 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:31:46.288 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000098/group_2/index_1056_seed_31266_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 00:31:51.452 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 00:32:09.213 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:32:10.419 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:32:10.469 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 00:32:11.403 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:34:06.389 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 00:34:07.118 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:36:03.850 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:43:05.347 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 99, train_steps: 3960 +2026-06-20 00:43:06.025 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 00:43:06.237 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:43:06.377 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:43:06.458 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:44:12.405 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_2/index_8280_seed_25367_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 00:44:16.606 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 00:44:35.311 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:44:35.319 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:44:35.319 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:45:46.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_2/index_8280_seed_25368_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 00:45:49.635 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 00:46:07.948 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 00:46:09.899 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 00:46:09.911 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 00:46:10.929 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 00:48:05.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.75s (1.91min) +2026-06-20 00:48:06.440 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 00:50:03.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 00:57:04.811 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 100, train_steps: 4000 +2026-06-20 00:58:47.832 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=100 (missing=0, unexpected=0) +2026-06-20 00:58:49.577 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 00:58:49.762 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 00:58:49.860 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 00:58:49.958 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 00:59:57.486 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_2/index_7011_seed_3388_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 01:00:02.769 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 01:00:20.485 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:00:20.494 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:00:20.494 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:01:26.666 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_2/index_7011_seed_3389_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 01:01:31.537 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 01:01:49.231 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:01:50.303 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:01:50.390 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 01:01:51.231 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:03:44.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-20 01:03:45.434 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:05:42.220 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:12:39.873 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 101, train_steps: 4040 +2026-06-20 01:12:40.568 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 01:12:40.769 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:12:40.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:12:40.936 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:13:47.448 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_2/index_6541_seed_30522_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 01:13:51.413 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 01:14:10.049 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:14:10.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:14:10.061 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:15:20.388 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_2/index_6541_seed_30523_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 01:15:25.688 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 01:15:43.488 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:15:43.499 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:15:43.530 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 01:15:44.468 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:17:38.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.62s (1.89min) +2026-06-20 01:17:38.931 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:19:38.324 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:26:37.814 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 102, train_steps: 4080 +2026-06-20 01:26:38.505 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 01:26:38.694 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:26:38.794 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:26:38.891 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:27:44.860 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_2/index_5400_seed_24428_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 01:27:49.667 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 01:28:07.739 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:28:07.747 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:28:07.749 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:29:17.147 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_2/index_5400_seed_24429_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 01:29:21.978 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 01:29:39.986 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:29:41.183 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:29:41.230 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 01:29:42.404 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:31:36.383 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-20 01:31:37.098 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:33:37.225 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:40:38.545 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 103, train_steps: 4120 +2026-06-20 01:40:39.247 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 01:40:39.459 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:40:39.582 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:40:39.662 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:41:45.749 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000103/group_2/index_7370_seed_29006_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 01:41:50.643 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 01:42:08.561 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:42:08.567 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:42:08.567 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:43:19.230 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000103/group_2/index_7370_seed_29007_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 01:43:23.750 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 01:43:41.784 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:43:43.334 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:43:43.403 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 01:43:44.382 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:45:37.767 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.37s (1.89min) +2026-06-20 01:45:38.487 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 01:47:38.264 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 01:54:37.798 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 104, train_steps: 4160 +2026-06-20 01:54:38.703 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 01:54:38.883 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 01:54:38.985 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:54:39.087 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:55:45.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_2/index_9075_seed_13442_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 01:55:50.335 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 01:56:08.038 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:56:08.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 01:56:08.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 01:57:18.785 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_2/index_9075_seed_13443_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 01:57:23.808 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 01:57:41.458 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 01:57:42.889 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 01:57:43.042 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 01:57:43.804 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 01:59:37.314 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.35s (1.89min) +2026-06-20 01:59:38.031 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:01:38.019 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:08:38.451 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 105, train_steps: 4200 +2026-06-20 02:08:39.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 02:08:39.306 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:08:39.357 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:08:39.463 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:09:45.426 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000105/group_2/index_9746_seed_21674_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 02:09:50.113 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 02:10:08.252 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:10:08.261 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:10:08.261 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:11:18.986 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000105/group_2/index_9746_seed_21675_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 02:11:23.899 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 02:11:41.698 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:11:43.945 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:11:44.038 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 02:11:44.862 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:13:38.217 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 02:13:38.925 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:15:38.267 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:22:37.857 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 106, train_steps: 4240 +2026-06-20 02:22:38.564 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 02:22:38.736 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:22:38.819 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:22:38.924 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:23:44.839 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000106/group_2/index_3732_seed_33259_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 02:23:49.849 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 02:24:07.407 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:24:07.416 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:24:07.416 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:25:16.843 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000106/group_2/index_3732_seed_33260_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 02:25:21.723 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 02:25:39.707 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:25:40.847 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:25:40.907 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 02:25:41.779 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:27:35.879 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.90s (1.90min) +2026-06-20 02:27:36.583 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:29:36.811 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:36:37.104 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 107, train_steps: 4280 +2026-06-20 02:36:37.801 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 02:36:37.922 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:36:37.956 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:36:38.025 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:37:44.017 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_2/index_4181_seed_23196_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 02:37:47.947 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 02:38:06.238 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:38:06.247 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:38:06.247 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:39:15.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_2/index_4181_seed_23197_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 02:39:18.470 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 02:39:37.330 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:39:39.229 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:39:39.279 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 02:39:40.420 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:41:35.348 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 02:41:36.072 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:43:35.144 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 02:50:35.012 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 108, train_steps: 4320 +2026-06-20 02:50:35.729 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 02:50:35.924 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 02:50:36.022 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:50:36.118 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:51:42.066 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_2/index_4481_seed_16987_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 02:51:47.282 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 02:52:04.933 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:52:04.941 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 02:52:04.942 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 02:53:13.053 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_2/index_4481_seed_16988_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 02:53:17.581 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 02:53:35.875 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 02:53:37.574 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 02:53:37.640 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 02:53:38.596 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 02:55:33.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.45s (1.91min) +2026-06-20 02:55:33.972 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 02:57:30.648 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:04:30.700 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 109, train_steps: 4360 +2026-06-20 03:04:31.405 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 03:04:31.512 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:04:31.549 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:04:31.599 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:05:37.614 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000109/group_2/index_7909_seed_26364_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 03:05:42.805 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 03:06:00.600 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:06:00.606 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:06:00.607 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:07:09.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000109/group_2/index_7909_seed_26365_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 03:07:14.720 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 03:07:33.277 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:07:33.407 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:07:33.420 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 03:07:34.322 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:09:29.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-20 03:09:30.014 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:11:26.575 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:18:25.332 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 110, train_steps: 4400 +2026-06-20 03:19:01.425 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=110 (missing=0, unexpected=0) +2026-06-20 03:19:03.284 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 03:19:03.451 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:19:03.517 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:19:03.599 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:20:11.591 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_2/index_579_seed_31341_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 03:20:17.120 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 03:20:35.212 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:20:35.221 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:20:35.222 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:21:43.285 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_2/index_579_seed_31342_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 03:21:47.477 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 03:22:05.821 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:22:07.524 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:22:07.640 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 03:22:08.526 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:24:03.966 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.41s (1.92min) +2026-06-20 03:24:04.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:26:02.472 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:33:05.401 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 111, train_steps: 4440 +2026-06-20 03:33:06.073 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 03:33:06.227 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:33:06.295 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:33:06.404 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:34:12.438 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_2/index_1600_seed_31904_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 03:34:17.619 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 03:34:34.933 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:34:34.942 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:34:34.942 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:35:46.558 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_2/index_1600_seed_31905_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 03:35:50.835 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 03:36:08.880 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:36:10.098 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:36:10.147 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 03:36:11.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:38:06.021 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.84s (1.91min) +2026-06-20 03:38:06.732 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:40:04.482 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 03:47:04.323 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 112, train_steps: 4480 +2026-06-20 03:47:05.042 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 03:47:05.141 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 03:47:05.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:47:05.227 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:48:11.201 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_2/index_4817_seed_22843_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 03:48:14.285 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 03:48:33.919 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:48:33.929 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 03:48:33.929 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 03:49:40.937 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_2/index_4817_seed_22844_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 03:49:46.096 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 03:50:04.045 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 03:50:04.983 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 03:50:05.120 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 03:50:06.231 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 03:52:00.943 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.63s (1.91min) +2026-06-20 03:52:01.657 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 03:54:00.963 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:01:04.010 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 113, train_steps: 4520 +2026-06-20 04:01:04.877 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 04:01:05.084 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:01:05.176 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:01:05.278 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:02:10.893 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000113/group_2/index_4697_seed_10377_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 04:02:16.192 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 04:02:35.109 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:02:35.119 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:02:35.119 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:03:42.090 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000113/group_2/index_4697_seed_10378_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 04:03:46.966 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 04:04:06.599 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:04:06.611 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:04:06.642 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 04:04:07.637 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:06:02.815 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.15s (1.92min) +2026-06-20 04:06:03.532 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:08:01.584 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:15:03.800 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 114, train_steps: 4560 +2026-06-20 04:15:04.473 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 04:15:04.581 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:15:04.632 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:15:04.688 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:16:10.626 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000114/group_2/index_4235_seed_26033_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 04:16:16.250 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 04:16:33.902 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:16:33.911 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:16:33.912 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:17:40.386 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000114/group_2/index_4235_seed_26034_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 04:17:44.422 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 04:18:03.299 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:18:04.152 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:18:04.195 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 04:18:05.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:20:00.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.72s (1.91min) +2026-06-20 04:20:00.894 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:22:00.099 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:29:01.479 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 115, train_steps: 4600 +2026-06-20 04:29:02.191 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 04:29:02.282 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:29:02.341 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:29:02.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:30:08.473 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_2/index_5087_seed_27712_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 04:30:11.463 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 04:30:29.581 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:30:29.591 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:30:29.591 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:31:38.931 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_2/index_5087_seed_27713_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 04:31:41.901 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 04:32:01.685 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:32:02.494 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:32:02.572 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 04:32:03.497 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:33:58.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.33s (1.91min) +2026-06-20 04:33:58.775 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:35:55.624 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:42:56.481 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 116, train_steps: 4640 +2026-06-20 04:42:57.160 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 04:42:57.348 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:42:57.474 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:42:57.571 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:44:03.529 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_2/index_7778_seed_16418_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 04:44:08.755 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 04:44:26.987 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:44:26.996 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:44:26.996 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:45:33.953 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_2/index_7778_seed_16419_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 04:45:37.940 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 04:45:56.532 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:45:58.380 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:45:58.431 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 04:45:59.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 04:47:53.698 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.25s (1.90min) +2026-06-20 04:47:54.407 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 04:49:50.887 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 04:56:50.236 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 117, train_steps: 4680 +2026-06-20 04:56:50.953 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 04:56:51.076 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 04:56:51.103 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:56:51.150 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:57:57.327 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_2/index_688_seed_28838_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 04:58:00.334 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 04:58:18.119 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:58:18.129 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 04:58:18.129 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 04:59:28.312 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_2/index_688_seed_28839_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 04:59:33.552 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 04:59:51.548 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 04:59:51.798 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 04:59:51.861 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 04:59:52.767 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:01:46.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.49s (1.89min) +2026-06-20 05:01:47.078 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:03:44.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:10:42.601 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 118, train_steps: 4720 +2026-06-20 05:10:43.296 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 05:10:43.474 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:10:43.576 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:10:43.679 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:11:49.494 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000118/group_2/index_4089_seed_20982_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 05:11:53.660 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 05:12:11.942 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:12:11.952 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:12:11.952 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:13:23.064 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000118/group_2/index_4089_seed_20983_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 05:13:28.279 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 05:13:46.799 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:13:46.957 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:13:47.070 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 05:13:47.942 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:15:41.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.26s (1.89min) +2026-06-20 05:15:42.001 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:17:38.994 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:24:37.902 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 119, train_steps: 4760 +2026-06-20 05:24:38.623 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 05:24:38.784 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:24:38.841 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:24:38.949 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:25:45.029 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_2/index_8708_seed_14119_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 05:25:49.961 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 05:26:07.887 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:26:07.896 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:26:07.896 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:27:14.938 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_2/index_8708_seed_14120_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 05:27:18.911 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 05:27:37.196 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:27:40.119 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:27:40.218 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 05:27:41.297 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:29:34.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-20 05:29:35.560 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:31:33.105 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:38:32.160 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 120, train_steps: 4800 +2026-06-20 05:39:08.950 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=120 (missing=0, unexpected=0) +2026-06-20 05:39:10.882 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 05:39:11.060 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:39:11.165 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:39:11.250 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:40:18.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000120/group_2/index_787_seed_1871_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 05:40:21.927 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 05:40:40.682 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:40:40.692 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:40:40.693 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:41:48.584 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000120/group_2/index_787_seed_1872_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 05:41:51.594 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 05:42:10.069 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:42:12.566 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:42:12.651 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 05:42:13.738 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:44:08.716 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.96s (1.92min) +2026-06-20 05:44:09.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 05:46:10.086 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 05:53:12.072 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 121, train_steps: 4840 +2026-06-20 05:53:12.775 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 05:53:12.895 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 05:53:12.931 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:53:13.000 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:54:19.568 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_2/index_2891_seed_27869_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 05:54:22.603 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 05:54:40.560 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:54:40.570 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 05:54:40.570 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 05:55:48.351 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_2/index_2891_seed_27870_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 05:55:51.380 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 05:56:09.586 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 05:56:13.156 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 05:56:13.264 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 05:56:14.356 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 05:58:09.349 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.97s (1.92min) +2026-06-20 05:58:10.076 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:00:10.575 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:07:11.968 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 122, train_steps: 4880 +2026-06-20 06:07:12.648 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 06:07:12.812 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:07:12.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:07:12.987 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:08:18.981 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_2/index_3378_seed_3212_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 06:08:21.967 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 06:08:40.121 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:08:40.131 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:08:40.131 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:09:48.841 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_2/index_3378_seed_3213_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 06:09:54.216 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 06:10:12.623 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:10:13.195 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:10:13.313 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 06:10:14.343 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:12:08.624 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.25s (1.90min) +2026-06-20 06:12:09.326 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:14:10.579 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:21:11.874 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 123, train_steps: 4920 +2026-06-20 06:21:12.562 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 06:21:12.746 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:21:12.840 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:21:12.933 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:22:18.822 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_2/index_2340_seed_11952_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 06:22:22.818 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 06:22:41.260 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:22:41.269 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:22:41.269 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:23:48.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_2/index_2340_seed_11953_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 06:23:51.244 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 06:24:09.744 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:24:12.071 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:24:12.170 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 06:24:13.286 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:26:08.788 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.47s (1.92min) +2026-06-20 06:26:09.511 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:28:09.833 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:35:11.853 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 124, train_steps: 4960 +2026-06-20 06:35:12.538 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 06:35:12.712 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:35:12.778 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:35:12.896 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:36:18.865 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_2/index_2859_seed_24062_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 06:36:24.030 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 06:36:42.191 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:36:42.200 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:36:42.201 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:37:48.120 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_2/index_2859_seed_24063_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 06:37:53.098 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 06:38:11.899 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:38:12.129 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:38:12.262 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 06:38:13.256 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:40:08.721 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.44s (1.92min) +2026-06-20 06:40:09.426 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:42:09.772 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 06:49:11.658 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 125, train_steps: 5000 +2026-06-20 06:50:15.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 06:50:15.370 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 06:50:15.476 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:50:15.581 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:51:22.682 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_2/index_6853_seed_10085_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 06:51:27.893 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 06:51:46.014 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:51:46.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 06:51:46.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 06:52:51.138 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_2/index_6853_seed_10086_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 06:52:55.275 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 06:53:15.596 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 06:53:15.926 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 06:53:15.939 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 06:53:16.938 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 06:55:11.654 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.68s (1.91min) +2026-06-20 06:55:12.372 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 06:57:11.763 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:04:11.933 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 126, train_steps: 5040 +2026-06-20 07:04:12.630 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 07:04:12.783 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:04:12.844 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:04:12.939 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:05:19.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_2/index_6534_seed_8715_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 07:05:24.944 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 07:05:43.081 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:05:43.091 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:05:43.091 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:06:49.498 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_2/index_6534_seed_8716_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 07:06:54.096 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 07:07:12.548 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:07:13.988 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:07:14.054 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 07:07:15.347 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:09:10.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.08s (1.92min) +2026-06-20 07:09:11.146 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:11:11.164 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:18:11.942 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 127, train_steps: 5080 +2026-06-20 07:18:12.630 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 07:18:12.836 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:18:12.934 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:18:13.027 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:19:18.821 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_2/index_7506_seed_7505_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 07:19:23.920 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 07:19:42.099 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:19:42.106 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:19:42.106 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:20:48.848 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_2/index_7506_seed_7506_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 07:20:53.573 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 07:21:11.721 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:21:13.050 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:21:13.129 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 07:21:14.054 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:23:09.601 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.50s (1.93min) +2026-06-20 07:23:10.327 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:25:10.159 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:32:11.174 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 128, train_steps: 5120 +2026-06-20 07:32:11.885 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 07:32:12.083 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:32:12.175 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:32:12.287 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:33:18.007 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_2/index_2245_seed_17318_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 07:33:22.850 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 07:33:40.369 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:33:40.378 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:33:40.381 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:34:48.827 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_2/index_2245_seed_17319_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 07:34:53.643 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 07:35:11.394 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:35:13.016 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:35:13.111 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 07:35:14.052 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:37:08.641 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.56s (1.91min) +2026-06-20 07:37:09.373 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:39:09.929 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 07:46:11.997 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 129, train_steps: 5160 +2026-06-20 07:46:12.715 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 07:46:12.940 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 07:46:13.068 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:46:13.147 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:47:19.379 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_2/index_4531_seed_32812_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 07:47:24.196 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 07:47:42.073 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:47:42.082 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 07:47:42.083 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 07:48:48.966 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_2/index_4531_seed_32813_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 07:48:52.903 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 07:49:11.280 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 07:49:13.070 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 07:49:13.128 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 07:49:14.093 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 07:51:09.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-20 07:51:09.758 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 07:53:10.303 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:00:11.803 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 130, train_steps: 5200 +2026-06-20 08:00:49.246 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=130 (missing=0, unexpected=0) +2026-06-20 08:00:51.245 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 08:00:51.412 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:00:51.512 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:00:51.594 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:02:00.281 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_2/index_274_seed_26665_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 08:02:05.205 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 08:02:23.084 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:02:23.093 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:02:23.093 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:03:28.889 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_2/index_274_seed_26666_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 08:03:33.727 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 08:03:51.202 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:03:52.509 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:03:52.573 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 08:03:53.419 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:05:47.283 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.72s (1.90min) +2026-06-20 08:05:48.018 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:07:44.845 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:14:45.401 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 131, train_steps: 5240 +2026-06-20 08:14:46.026 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 08:14:46.243 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:14:46.362 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:14:46.464 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:15:52.766 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_2/index_6431_seed_4149_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 08:15:57.523 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 08:16:15.727 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:16:15.737 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:16:15.737 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:17:23.380 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_2/index_6431_seed_4150_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 08:17:28.268 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 08:17:46.096 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:17:46.990 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:17:47.095 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 08:17:47.915 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:19:41.428 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.38s (1.89min) +2026-06-20 08:19:42.140 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:21:39.469 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:28:37.835 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 132, train_steps: 5280 +2026-06-20 08:28:38.522 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 08:28:38.742 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:28:38.867 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:28:38.927 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:29:45.116 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000132/group_2/index_5326_seed_4128_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 08:29:49.871 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 08:30:07.153 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:30:07.163 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:30:07.163 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:31:15.476 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000132/group_2/index_5326_seed_4129_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 08:31:19.749 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 08:31:37.843 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:31:40.849 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:31:40.970 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 08:31:41.901 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:33:35.413 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.44s (1.89min) +2026-06-20 08:33:36.119 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:35:35.262 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:42:34.334 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 133, train_steps: 5320 +2026-06-20 08:42:35.019 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 08:42:35.223 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:42:35.335 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:42:35.431 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:43:41.485 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_2/index_6484_seed_31645_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 08:43:46.508 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 08:44:04.308 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:44:04.318 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:44:04.318 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:45:11.898 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000133/group_2/index_6484_seed_31646_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 08:45:16.114 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 08:45:34.532 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:45:35.888 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:45:35.928 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 08:45:36.912 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 08:47:32.712 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.77s (1.93min) +2026-06-20 08:47:33.449 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 08:49:30.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 08:56:29.442 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 134, train_steps: 5360 +2026-06-20 08:56:30.114 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 08:56:30.249 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 08:56:30.287 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:56:30.379 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:57:36.602 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000134/group_2/index_5086_seed_21436_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 08:57:39.424 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 08:57:57.292 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:57:57.302 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 08:57:57.302 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 08:59:07.719 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000134/group_2/index_5086_seed_21437_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 08:59:12.640 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 08:59:32.031 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 08:59:32.040 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 08:59:32.052 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 08:59:32.967 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:01:26.319 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 09:01:27.048 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:03:23.809 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:10:22.991 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 135, train_steps: 5400 +2026-06-20 09:10:23.676 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 09:10:23.885 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:10:23.987 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:10:24.082 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:11:30.003 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000135/group_2/index_4512_seed_9424_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 09:11:33.745 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 09:11:51.969 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:11:51.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:11:51.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:13:00.322 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000135/group_2/index_4512_seed_9425_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 09:13:03.055 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 09:13:21.065 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:13:23.932 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:13:24.030 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 09:13:24.927 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:15:18.588 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.61s (1.89min) +2026-06-20 09:15:19.310 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:17:17.100 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:24:16.817 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 136, train_steps: 5440 +2026-06-20 09:24:17.533 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 09:24:17.674 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:24:17.715 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:24:17.804 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:25:24.317 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_2/index_3596_seed_23274_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 09:25:29.341 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 09:25:47.179 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:25:47.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:25:47.189 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:26:54.218 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_2/index_3596_seed_23275_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 09:26:58.397 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 09:27:16.853 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:27:18.656 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:27:18.707 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 09:27:19.689 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:29:12.958 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.15s (1.89min) +2026-06-20 09:29:13.686 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:31:13.177 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:38:13.228 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 137, train_steps: 5480 +2026-06-20 09:38:13.937 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 09:38:14.094 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:38:14.149 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:38:14.269 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:39:20.248 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000137/group_2/index_377_seed_26952_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 09:39:22.946 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 09:39:40.843 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:39:40.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:39:40.853 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:40:50.210 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000137/group_2/index_377_seed_26953_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 09:40:53.061 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 09:41:11.198 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:41:14.371 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:41:14.467 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 09:41:15.626 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:43:09.522 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-20 09:43:10.250 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:45:10.503 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 09:52:12.176 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 138, train_steps: 5520 +2026-06-20 09:52:12.865 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 09:52:13.013 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 09:52:13.063 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:52:13.160 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:53:19.292 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_2/index_442_seed_18004_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 09:53:22.035 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 09:53:40.079 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:53:40.088 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 09:53:40.088 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 09:54:49.341 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_2/index_442_seed_18005_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 09:54:54.367 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 09:55:12.301 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 09:55:13.383 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 09:55:13.479 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 09:55:14.692 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 09:57:10.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.46s (1.92min) +2026-06-20 09:57:10.921 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 09:59:11.150 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:06:12.678 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 139, train_steps: 5560 +2026-06-20 10:06:13.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 10:06:13.557 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:06:13.647 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:06:13.741 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:07:19.972 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_2/index_8340_seed_19065_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 10:07:24.223 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 10:07:42.450 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:07:42.460 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:07:42.460 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:08:49.998 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000139/group_2/index_8340_seed_19066_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 10:08:55.058 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 10:09:13.239 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:09:13.264 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:09:13.286 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 10:09:14.461 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:11:08.675 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.20s (1.90min) +2026-06-20 10:11:09.398 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:13:09.780 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:20:12.717 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 140, train_steps: 5600 +2026-06-20 10:20:49.865 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=140 (missing=0, unexpected=0) +2026-06-20 10:20:51.813 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 10:20:51.895 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:20:51.963 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:20:52.076 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:21:59.948 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000140/group_2/index_7900_seed_10363_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 10:22:04.134 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 10:22:22.697 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:22:22.707 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:22:22.707 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:23:28.955 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000140/group_2/index_7900_seed_10364_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 10:23:33.951 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 10:23:51.871 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:23:53.040 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:23:53.151 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 10:23:53.995 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:25:47.600 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.52s (1.89min) +2026-06-20 10:25:48.314 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:27:45.363 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:34:44.016 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 141, train_steps: 5640 +2026-06-20 10:34:44.687 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 10:34:44.821 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:34:44.864 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:34:44.954 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:35:51.171 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_2/index_6307_seed_18594_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 10:35:55.562 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 10:36:13.443 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:36:13.452 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:36:13.452 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:37:22.992 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_2/index_6307_seed_18595_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 10:37:28.066 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 10:37:46.037 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:37:46.846 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:37:46.955 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 10:37:47.816 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:39:41.415 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.51s (1.89min) +2026-06-20 10:39:42.148 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:41:39.461 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 10:48:38.470 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 142, train_steps: 5680 +2026-06-20 10:48:39.160 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 10:48:39.373 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 10:48:39.482 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:48:39.651 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:49:45.532 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000142/group_2/index_4811_seed_3058_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 10:49:50.249 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 10:50:07.784 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:50:07.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 10:50:07.794 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 10:51:19.097 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000142/group_2/index_4811_seed_3059_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 10:51:23.120 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 10:51:41.440 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 10:51:43.153 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 10:51:43.250 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 10:51:44.162 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 10:53:37.146 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.94s (1.88min) +2026-06-20 10:53:37.851 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 10:55:37.993 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:02:37.932 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 143, train_steps: 5720 +2026-06-20 11:02:38.633 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 11:02:38.827 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:02:38.923 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:02:39.019 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:03:45.179 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000143/group_2/index_6705_seed_32513_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 11:03:48.084 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 11:04:06.266 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:04:06.277 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:04:06.277 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:05:18.254 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000143/group_2/index_6705_seed_32514_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 11:05:21.091 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 11:05:39.411 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:05:42.169 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:05:42.320 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 11:05:43.345 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:07:36.589 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.22s (1.89min) +2026-06-20 11:07:37.312 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:09:36.843 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:16:37.114 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 144, train_steps: 5760 +2026-06-20 11:16:37.797 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 11:16:37.890 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:16:37.936 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:16:37.983 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:17:44.012 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_2/index_3703_seed_15961_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 11:17:47.022 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 11:18:05.178 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:18:05.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:18:05.188 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:19:13.181 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000144/group_2/index_3703_seed_15962_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 11:19:16.413 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 11:19:34.547 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:19:38.707 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:19:38.822 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 11:19:39.811 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:21:33.370 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.47s (1.89min) +2026-06-20 11:21:34.087 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:23:30.884 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:30:30.411 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 145, train_steps: 5800 +2026-06-20 11:30:31.100 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 11:30:31.306 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:30:31.411 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:30:31.506 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:31:37.459 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000145/group_2/index_2507_seed_25117_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 11:31:40.354 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 11:31:58.442 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:31:58.453 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:31:58.453 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:33:09.076 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000145/group_2/index_2507_seed_25118_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 11:33:14.454 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 11:33:33.197 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:33:33.207 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:33:33.226 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 11:33:34.180 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:35:28.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.89s (1.90min) +2026-06-20 11:35:28.900 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:37:25.686 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:44:23.924 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 146, train_steps: 5840 +2026-06-20 11:44:24.629 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 11:44:24.827 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:44:24.927 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:44:25.019 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:45:30.865 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_2/index_5961_seed_31248_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 11:45:35.935 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 11:45:54.049 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:45:54.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:45:54.059 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:46:59.371 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_2/index_5961_seed_31249_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 11:47:04.260 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 11:47:22.288 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:47:23.247 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 11:47:23.363 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 11:47:24.174 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 11:49:19.133 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.83s (1.91min) +2026-06-20 11:49:19.856 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 11:51:17.039 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 11:58:15.945 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 147, train_steps: 5880 +2026-06-20 11:58:16.619 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 11:58:16.809 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 11:58:16.912 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:58:17.009 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 11:59:23.881 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_2/index_9617_seed_19312_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 11:59:28.695 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 11:59:46.523 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 11:59:46.533 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 11:59:46.533 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:00:52.688 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_2/index_9617_seed_19313_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 12:00:57.629 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 12:01:16.002 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:01:16.232 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:01:16.303 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 12:01:17.280 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:03:12.586 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.29s (1.92min) +2026-06-20 12:03:13.293 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:05:12.904 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:12:12.714 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 148, train_steps: 5920 +2026-06-20 12:12:13.408 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 12:12:13.618 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:12:13.695 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:12:13.769 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:13:20.126 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000148/group_2/index_8969_seed_25725_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 12:13:24.380 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 12:13:42.744 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:13:42.754 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:13:42.755 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:14:49.643 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000148/group_2/index_8969_seed_25726_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 12:14:54.555 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 12:15:12.562 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:15:13.144 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:15:13.222 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 12:15:14.332 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:17:09.048 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.69s (1.91min) +2026-06-20 12:17:09.789 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:19:10.588 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:26:11.397 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 149, train_steps: 5960 +2026-06-20 12:26:12.137 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 12:26:12.358 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:26:12.451 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:26:12.547 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:27:18.431 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000149/group_2/index_1277_seed_22398_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 12:27:23.273 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 12:27:40.730 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:27:40.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:27:40.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:28:48.587 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000149/group_2/index_1277_seed_22399_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 12:28:53.547 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 12:29:11.186 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:29:13.029 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:29:13.141 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 12:29:14.011 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:31:08.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.71s (1.91min) +2026-06-20 12:31:09.533 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:33:09.831 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:40:10.994 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 150, train_steps: 6000 +2026-06-20 12:41:53.433 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=150 (missing=0, unexpected=0) +2026-06-20 12:41:56.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 12:41:56.282 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:41:56.401 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:41:56.500 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:43:04.250 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_2/index_5404_seed_32847_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 12:43:09.002 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 12:43:26.699 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:43:26.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:43:26.708 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:44:32.729 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_2/index_5404_seed_32848_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 12:44:36.563 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 12:44:55.017 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:44:59.744 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:44:59.894 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 12:45:00.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 12:46:55.516 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.70s (1.91min) +2026-06-20 12:46:56.236 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 12:48:52.988 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 12:55:51.686 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 151, train_steps: 6040 +2026-06-20 12:55:52.308 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 12:55:52.529 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 12:55:52.626 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:55:52.715 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:56:58.603 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_2/index_1145_seed_15287_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 12:57:03.766 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 12:57:21.384 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:57:21.392 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 12:57:21.392 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 12:58:27.270 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_2/index_1145_seed_15288_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 12:58:31.289 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 12:58:49.886 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 12:58:50.721 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 12:58:50.846 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 12:58:51.757 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:00:46.189 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.41s (1.91min) +2026-06-20 13:00:46.908 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:02:43.652 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:09:41.918 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 152, train_steps: 6080 +2026-06-20 13:09:42.598 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 13:09:42.782 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:09:42.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:09:42.975 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:10:49.257 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_2/index_9151_seed_2626_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 13:10:52.283 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 13:11:10.452 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:11:10.463 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:11:10.463 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:12:20.548 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_2/index_9151_seed_2627_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 13:12:24.797 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 13:12:42.970 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:12:43.910 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:12:43.958 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 13:12:44.936 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:14:38.764 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.79s (1.90min) +2026-06-20 13:14:39.492 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:16:38.549 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:23:37.234 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 153, train_steps: 6120 +2026-06-20 13:23:38.094 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 13:23:38.294 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:23:38.407 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:23:38.507 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:24:44.834 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_2/index_2147_seed_32760_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 13:24:47.630 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 13:25:05.890 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:25:05.901 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:25:05.901 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:26:15.022 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_2/index_2147_seed_32761_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 13:26:20.076 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 13:26:38.163 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:26:40.127 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:26:40.226 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 13:26:41.276 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:28:34.748 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.39s (1.89min) +2026-06-20 13:28:35.473 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:30:32.908 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:37:31.708 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 154, train_steps: 6160 +2026-06-20 13:37:32.394 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 13:37:32.607 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:37:32.703 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:37:32.783 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:38:38.706 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_2/index_8905_seed_19803_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 13:38:43.549 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 13:39:01.083 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:39:01.092 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:39:01.092 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:40:09.456 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_2/index_8905_seed_19804_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 13:40:13.784 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 13:40:33.773 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:40:33.788 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:40:33.801 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 13:40:34.828 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:42:29.329 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.47s (1.91min) +2026-06-20 13:42:30.062 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:44:27.161 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 13:51:24.984 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 155, train_steps: 6200 +2026-06-20 13:51:25.699 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 13:51:25.789 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 13:51:25.840 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:51:25.884 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:52:31.898 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_2/index_1242_seed_31152_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 13:52:34.679 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 13:52:53.114 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:52:53.124 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 13:52:53.125 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 13:54:01.084 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_2/index_1242_seed_31153_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 13:54:06.147 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 13:54:24.293 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 13:54:24.447 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 13:54:24.487 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 13:54:25.434 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 13:56:19.128 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.63s (1.89min) +2026-06-20 13:56:19.859 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 13:58:16.751 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:05:15.357 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 156, train_steps: 6240 +2026-06-20 14:05:16.035 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 14:05:16.252 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:05:16.384 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:05:16.483 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:06:23.615 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_2/index_4352_seed_15794_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 14:06:28.648 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 14:06:46.641 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:06:46.651 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:06:46.651 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:07:51.837 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_2/index_4352_seed_15795_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 14:07:56.800 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 14:08:15.921 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:08:15.931 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:08:15.944 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 14:08:16.960 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:10:11.722 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.74s (1.91min) +2026-06-20 14:10:12.457 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:12:11.103 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:19:12.251 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 157, train_steps: 6280 +2026-06-20 14:19:12.926 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 14:19:13.127 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:19:13.228 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:19:13.320 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:20:19.196 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000157/group_2/index_834_seed_31261_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 14:20:24.188 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 14:20:42.355 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:20:42.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:20:42.365 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:21:48.061 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000157/group_2/index_834_seed_31262_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 14:21:52.802 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 14:22:11.567 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:22:12.398 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:22:12.487 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 14:22:13.740 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:24:08.968 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.22s (1.92min) +2026-06-20 14:24:09.690 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:26:09.904 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:33:10.841 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 158, train_steps: 6320 +2026-06-20 14:33:11.519 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 14:33:11.720 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:33:11.837 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:33:11.917 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:34:17.538 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_2/index_4159_seed_31489_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 14:34:20.346 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 14:34:38.842 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:34:38.852 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:34:38.852 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:35:47.086 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_2/index_4159_seed_31490_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 14:35:52.098 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 14:36:10.923 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:36:10.935 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:36:11.052 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 14:36:12.058 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:38:07.056 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.88s (1.91min) +2026-06-20 14:38:07.784 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:40:05.257 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 14:47:04.575 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 159, train_steps: 6360 +2026-06-20 14:47:05.294 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 14:47:05.507 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 14:47:05.645 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:47:05.722 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:48:11.549 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_2/index_4177_seed_18631_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 14:48:16.941 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 14:48:34.964 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:48:34.973 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 14:48:34.974 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 14:49:40.251 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_2/index_4177_seed_18632_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 14:49:45.126 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 14:50:02.971 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 14:50:04.049 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 14:50:04.127 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 14:50:05.187 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 14:51:59.192 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.86s (1.90min) +2026-06-20 14:51:59.929 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 14:53:57.321 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:00:59.188 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 160, train_steps: 6400 +2026-06-20 15:01:37.509 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=160 (missing=0, unexpected=0) +2026-06-20 15:01:39.304 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 15:01:39.387 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:01:39.431 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:01:39.576 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:02:46.925 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_2/index_1860_seed_2114_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 15:02:51.636 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 15:03:09.643 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:03:09.650 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:03:09.650 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:04:18.845 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_2/index_1860_seed_2115_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 15:04:23.775 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 15:04:42.084 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:04:42.879 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:04:42.991 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 15:04:43.802 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:06:37.408 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.48s (1.89min) +2026-06-20 15:06:38.136 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:08:38.303 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:15:37.531 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 161, train_steps: 6440 +2026-06-20 15:15:38.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 15:15:38.410 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:15:38.515 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:15:38.611 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:16:44.559 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_2/index_8899_seed_591_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 15:16:48.471 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 15:17:06.777 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:17:06.784 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:17:06.784 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:18:16.635 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_2/index_8899_seed_592_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 15:18:21.830 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 15:18:40.365 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:18:40.375 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:18:40.388 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 15:18:41.496 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:20:35.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.87s (1.90min) +2026-06-20 15:20:36.332 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:22:36.147 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:29:36.904 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 162, train_steps: 6480 +2026-06-20 15:29:37.567 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 15:29:37.660 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:29:37.705 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:29:37.746 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:30:43.946 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_2/index_3382_seed_5764_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 15:30:48.784 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 15:31:06.755 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:31:06.763 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:31:06.763 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:32:12.001 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_2/index_3382_seed_5765_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 15:32:16.921 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 15:32:34.950 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:32:37.444 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:32:37.496 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 15:32:38.379 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:34:33.099 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.48s (1.91min) +2026-06-20 15:34:33.827 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:36:31.317 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:43:31.491 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 163, train_steps: 6520 +2026-06-20 15:43:32.172 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 15:43:32.371 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:43:32.498 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:43:32.595 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:44:38.649 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000163/group_2/index_5439_seed_7646_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 15:44:43.497 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 15:45:00.921 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:45:00.928 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:45:00.928 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:46:09.473 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000163/group_2/index_5439_seed_7647_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 15:46:13.449 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 15:46:33.497 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:46:33.676 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 15:46:33.743 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 15:46:34.714 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 15:48:29.208 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.47s (1.91min) +2026-06-20 15:48:29.934 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 15:50:26.537 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 15:57:25.516 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 164, train_steps: 6560 +2026-06-20 15:57:26.227 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 15:57:26.328 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 15:57:26.375 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:57:26.422 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 15:58:32.229 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_2/index_683_seed_24698_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 15:58:35.099 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 15:58:53.276 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 15:58:53.287 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 15:58:53.289 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:00:02.402 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_2/index_683_seed_24699_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 16:00:07.617 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 16:00:25.820 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:00:26.402 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:00:26.457 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 16:00:27.366 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:02:21.064 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.61s (1.89min) +2026-06-20 16:02:21.793 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:04:18.346 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:11:18.384 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 165, train_steps: 6600 +2026-06-20 16:11:19.088 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 16:11:19.300 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:11:19.418 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:11:19.499 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:12:25.793 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_2/index_5998_seed_32826_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 16:12:30.568 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 16:12:48.010 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:12:48.022 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:12:48.022 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:13:54.799 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_2/index_5998_seed_32827_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 16:13:59.651 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 16:14:17.562 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:14:19.106 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:14:19.170 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 16:14:20.180 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:16:14.145 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.75s (1.90min) +2026-06-20 16:16:14.870 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:18:14.125 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:25:13.935 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 166, train_steps: 6640 +2026-06-20 16:25:14.622 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 16:25:14.844 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:25:14.986 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:25:15.066 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:26:21.944 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_2/index_2108_seed_33223_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 16:26:26.831 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 16:26:44.280 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:26:44.290 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:26:44.290 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:27:50.060 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_2/index_2108_seed_33224_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 16:27:54.929 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 16:28:12.680 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:28:13.831 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:28:13.978 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 16:28:15.047 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:30:08.844 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.78s (1.90min) +2026-06-20 16:30:09.574 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:32:10.306 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:39:12.154 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 167, train_steps: 6680 +2026-06-20 16:39:12.849 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 16:39:13.037 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:39:13.122 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:39:13.213 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:40:19.565 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_2/index_3206_seed_19642_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 16:40:24.390 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 16:40:41.998 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:40:42.007 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:40:42.007 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:41:48.085 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_2/index_3206_seed_19643_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 16:41:52.381 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 16:42:10.629 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:42:12.095 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:42:12.195 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 16:42:13.362 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:44:08.745 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.36s (1.92min) +2026-06-20 16:44:09.477 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 16:46:10.019 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 16:53:11.389 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 168, train_steps: 6720 +2026-06-20 16:53:12.092 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 16:53:12.284 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 16:53:12.389 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:53:12.480 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:54:18.812 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_2/index_4575_seed_33079_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 16:54:21.627 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 16:54:39.692 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:54:39.701 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 16:54:39.701 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 16:55:47.500 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_2/index_4575_seed_33080_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 16:55:50.335 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 16:56:08.206 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 16:56:11.768 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 16:56:11.837 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 16:56:12.784 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 16:58:07.736 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 114.91s (1.92min) +2026-06-20 16:58:08.464 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:00:07.578 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:07:09.190 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 169, train_steps: 6760 +2026-06-20 17:07:09.878 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 17:07:10.018 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 17:07:10.082 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:07:10.218 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:08:16.467 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_2/index_866_seed_28904_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 17:08:21.580 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 17:08:39.244 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:08:39.253 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:08:39.255 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:09:47.724 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_2/index_866_seed_28905_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 17:09:51.761 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 17:10:10.179 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:10:11.904 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 17:10:11.989 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 17:10:12.933 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:12:08.254 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 115.20s (1.92min) +2026-06-20 17:12:08.969 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:14:09.318 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:21:11.057 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 170, train_steps: 6800 +2026-06-20 17:21:46.926 | INFO | __main__:maybe_update_ref_model_for_kl:801 - [Moving KL] Updated ref_model at update_step=170 (missing=0, unexpected=0) +2026-06-20 17:21:49.219 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 17:21:49.382 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 17:21:49.447 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:21:49.564 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:22:56.442 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_2/index_104_seed_20147_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 17:22:59.307 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 17:23:17.172 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:23:17.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:23:17.182 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:24:25.377 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_2/index_104_seed_20148_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 17:24:28.221 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 17:24:47.384 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:24:49.925 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 17:24:50.007 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 17:24:50.934 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:26:44.023 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 113.05s (1.88min) +2026-06-20 17:26:44.743 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-20 17:28:41.425 | INFO | hyvideo.utils.grpo_utils:train_one_step:2941 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] +2026-06-20 17:35:39.746 | INFO | hyvideo.utils.grpo_utils:train_one_step:3585 - Rank 8, sp_rank 0: update step: 171, train_steps: 6840 +2026-06-20 17:35:40.432 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1261 - [ParallelGroups] Rank 8, dp_rank 4: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=2, rank_in_group=0 +2026-06-20 17:35:40.646 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model to cpu before rollout. +2026-06-20 17:35:40.897 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:35:40.983 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:36:47.098 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000171/group_2/index_2111_seed_11896_dp_4_rank_8_sample_0_slot_0.mp4 +2026-06-20 17:36:50.093 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 17:37:08.278 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:37:08.289 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1399 - Rank 8 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 +2026-06-20 17:37:08.291 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1424 - Rank 8 using 'all' strategy (no mixgrpo): determistic=False for all timesteps +2026-06-20 17:38:19.258 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1571 - Rank 8: Saved video 1/1 to ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000171/group_2/index_2111_seed_11897_dp_4_rank_8_sample_1_slot_1.mp4 +2026-06-20 17:38:22.434 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Loaded VLM reward model onto cuda:0 for reward scoring. +2026-06-20 17:38:40.858 | INFO | hyvideo.models.reward_models.rewards:_log:1195 - [RewardDebug] Offloaded VLM reward model back to cpu after reward scoring. +2026-06-20 17:38:42.537 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1033 - [GatherRewards] Rank 8: Using local metric names without global sync: ['avg', 'vlm_logprob', 'vlm_num_questions', 'vlm_token_credit', 'vlm_yes_no_margin', 'vlm_yes_ratio'] +2026-06-20 17:38:42.603 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:1101 - [GatherRewards] Rank 8 saved 4 entries for group 2 +2026-06-20 17:38:43.734 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1819 - Rank 8: Pre-computing Reference Model statistics (Batch-wise)... +2026-06-20 17:40:36.661 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:1962 - Rank 8: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 112.89s (1.88min) +2026-06-20 17:40:37.378 | INFO | hyvideo.utils.grpo_utils:train_one_step:2713 - Rank 8, sp_rank 0: token_credit GRPO active with 10 valid question slots; loss_mode=video_adv, ratio_mode=weighted_logprob, spatial_granularity=window, advantage_source=token, shuffle_weights=False, uniform_weights=False. +2026-06-29 01:30:24.512 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 01:30:25.177 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 01:30:25.178 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 01:53:14.558 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 01:53:15.218 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 01:53:15.221 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 02:23:32.093 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 02:23:32.773 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 02:23:32.773 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 02:48:38.950 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 02:48:39.602 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 02:48:39.603 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:11:17.332 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 03:11:17.988 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:11:17.989 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 03:34:30.727 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 03:34:31.406 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 03:34:31.407 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:00:51.679 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 04:00:52.338 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:00:52.341 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:28:55.218 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 04:28:55.882 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:28:55.885 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 04:53:59.918 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 04:54:00.601 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 04:54:00.604 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 05:18:09.640 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 05:18:10.314 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 05:18:10.315 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 05:42:02.370 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 05:42:03.049 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 05:42:03.049 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:05:56.996 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 06:05:57.652 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:05:57.653 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:27:00.355 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 06:27:01.013 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:27:01.015 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 06:51:39.073 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 06:51:39.756 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 06:51:39.756 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 07:14:09.506 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 07:14:10.167 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 07:14:10.168 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 07:37:05.749 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 07:37:06.401 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 07:37:06.404 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:07:15.351 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 08:07:16.023 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:07:16.024 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 08:37:01.667 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 08:37:02.343 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 08:37:02.344 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:01:09.937 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 09:01:10.604 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:01:10.607 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:24:43.459 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 09:24:44.142 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:24:44.144 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 09:49:42.573 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 09:49:43.246 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 09:49:43.247 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 10:13:57.209 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 10:13:57.882 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 10:13:57.883 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 10:38:10.706 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 10:38:11.370 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 10:38:11.370 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:05:35.041 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 11:05:35.710 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:05:35.712 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:29:17.667 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 11:29:18.330 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:29:18.331 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 11:57:20.710 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 11:57:21.386 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 11:57:21.387 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 12:20:18.542 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 12:20:19.204 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 12:20:19.206 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 12:48:19.328 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 12:48:19.999 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 12:48:20.000 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 13:14:43.880 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 13:14:44.532 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 13:14:44.535 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 13:39:09.853 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 13:39:10.507 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 13:39:10.510 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:06:07.867 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 14:06:08.538 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:06:08.538 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:28:55.673 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 14:28:56.335 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:28:56.338 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 14:51:29.445 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 14:51:30.125 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 14:51:30.126 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 15:22:59.801 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 15:23:00.472 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 15:23:00.473 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 15:48:17.137 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 15:48:17.809 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 15:48:17.810 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 16:18:25.291 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 16:18:25.952 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 16:18:25.953 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 16:44:49.017 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 16:44:49.699 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 16:44:49.699 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:10:29.591 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 17:10:30.268 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:10:30.271 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:34:03.144 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 17:34:03.844 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:34:03.845 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 17:55:28.113 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 17:55:28.782 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 17:55:28.782 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 18:18:05.741 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 18:18:06.424 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 18:18:06.424 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done +2026-06-29 18:40:44.339 | INFO | __main__:__init__:936 - [RewardDebug] VLM reward active: model=snapvideo_vlm, path=/nfs/ywang29/ckpts/Qwen3.5-9B, family=qwen3_5, score_type=token_credit, num_frames=20, max_pixels=50176 +2026-06-29 18:40:44.996 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 +2026-06-29 18:40:44.997 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_0/index_tensor([7542, 7542])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_0/index_tensor([7542, 7542])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ab1d094cd3cb9a54a4c4838ce69e769e39ed2703 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_0/index_tensor([7542, 7542])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "tracking shot, shot from an extremely low angle, follows a heavy main battle tank as it rolls forward with a heavy, unstoppable momentum. The tank's massive steel armor, covered in a deep olive-green matte camouflage paint, is stained with splashing mud and streaks of rain. It traverses a rain-soaked dirt road. The diffuse light from the overcast sky casts a cold, shimmering light on the damp metal and the muddy ground. Each heavy rotation of its broad metal tracks deeply churns, squeezing, and tumbling the soft, slippery, dark brown earth beneath them, leaving behind two distinct, deep, and rapidly flooded ruts. Accompanied by the deafening roar of the engine and the low crunch of the tracks, black exhaust fumes billow from the rear exhaust pipe. The entire scene is imbued with a sense of oppression and cold military might, creating an atmosphere of austerity and unstoppable force.", + "seed": 28576, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_0/index_7542_seed_28576_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.09320950489491224, + "vlm_logprob": -0.09320950489491224, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.09320950489491224, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank rolling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's armor painted with a deep olive-green matte camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank's surface covered in splashing mud and streaks of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tank traverse a rain-soaked dirt road while churning deep, flooded ruts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03960883244872093 + ], + "opposite_logprob_tokens": [ + -3.5396087169647217 + ], + "logprob_reward": -0.03960883244872093, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03960883244872093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10703367739915848, + 0.2948492467403412, + 0.10256608575582504, + 0.08083843439817429, + 0.09167645126581192, + 0.05310014635324478, + 0.048495810478925705, + 0.06525830179452896, + 0.11115412414073944, + 0.04502775892615318 + ], + "credit_map": [ + 0.013473019003868103, + 0.007465356029570103, + 0.029916614294052124, + 0.013326291926205158, + 0.008346530608832836, + 0.010835591703653336, + 0.011859219521284103, + 0.011811040341854095, + 0.04541594162583351, + 0.022444700822234154, + 0.09291764348745346, + 0.04216228798031807, + 0.020749587565660477, + 0.017280494794249535, + 0.027614815160632133, + 0.02626376412808895, + 0.011443287134170532, + 0.008903087116777897, + 0.023656098172068596, + 0.017437679693102837, + 0.008749833330512047, + 0.006393773481249809, + 0.013556981459259987, + 0.012425337918102741, + 0.009997030720114708, + 0.007046017330139875, + 0.012385088950395584, + 0.017125336453318596, + 0.006185035686939955, + 0.005791362840682268, + 0.009708025492727757, + 0.012600532732903957, + 0.00979212298989296, + 0.007563863880932331, + 0.014399003237485886, + 0.006761349271982908, + 0.007310749497264624, + 0.007586752064526081, + 0.01393975131213665, + 0.024322854354977608, + 0.005787569098174572, + 0.007160666864365339, + 0.009965610690414906, + 0.005940134171396494, + 0.006253234576433897, + 0.00750091252848506, + 0.005235499702394009, + 0.005256514996290207, + 0.006226267199963331, + 0.005346256773918867, + 0.007727031596004963, + 0.0041791293770074844, + 0.008012821897864342, + 0.003599100513383746, + 0.007655009161680937, + 0.005750193726271391, + 0.009471507743000984, + 0.006547932047396898, + 0.009539599530398846, + 0.0055218529887497425, + 0.008228234015405178, + 0.009089699946343899, + 0.008089552633464336, + 0.008769919164478779, + 0.008026594296097755, + 0.009476855397224426, + 0.013777735643088818, + 0.010210602544248104, + 0.021407337859272957, + 0.010688886046409607, + 0.018793677911162376, + 0.01877242885529995, + 0.004454365000128746, + 0.004944927990436554, + 0.00472069950774312, + 0.003853829111903906, + 0.008510556071996689, + 0.008629289455711842, + 0.004568950738757849, + 0.0053451391868293285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.03960883244872093 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank rolling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's armor painted with a deep olive-green matte camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank's surface covered in splashing mud and streaks of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tank traverse a rain-soaked dirt road while churning deep, flooded ruts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17994770407676697 + ], + "opposite_logprob_tokens": [ + -1.8049477338790894 + ], + "logprob_reward": -0.17994770407676697, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17994770407676697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2119745910167694, + 0.1281745433807373, + 0.12222734838724136, + 0.030597809702157974, + 0.03775855153799057, + 0.03421039506793022, + 0.10153660178184509, + 0.10041381418704987, + 0.04864361509680748, + 0.184462770819664 + ], + "credit_map": [ + 0.024764055386185646, + 0.015236010774970055, + 0.04042164608836174, + 0.05408470332622528, + 0.015957362949848175, + 0.011307875625789165, + 0.024565650150179863, + 0.025637291371822357, + 0.015953995287418365, + 0.008616434410214424, + 0.03216981142759323, + 0.01704251579940319, + 0.010098756290972233, + 0.008667751215398312, + 0.018749665468931198, + 0.016875609755516052, + 0.01310677919536829, + 0.009813564829528332, + 0.02997644990682602, + 0.027560561895370483, + 0.008693106472492218, + 0.007362051401287317, + 0.012504310347139835, + 0.013210529461503029, + 0.003631801810115576, + 0.003261056262999773, + 0.005530419293791056, + 0.004105177707970142, + 0.0025653457269072533, + 0.0024418053217232227, + 0.0036211812403053045, + 0.005441022105515003, + 0.004071288742125034, + 0.0034391023218631744, + 0.005147571209818125, + 0.00391447264701128, + 0.0026400163769721985, + 0.0023684559855610132, + 0.009260277263820171, + 0.006917367223650217, + 0.004717656411230564, + 0.0029935091733932495, + 0.007087418343871832, + 0.004234121646732092, + 0.002501566894352436, + 0.0027265322860330343, + 0.004684154875576496, + 0.005265437066555023, + 0.01554158329963684, + 0.00975363701581955, + 0.013547856360673904, + 0.009131688624620438, + 0.011136529967188835, + 0.011514854617416859, + 0.01350115705281496, + 0.017409292981028557, + 0.01894969679415226, + 0.010129605419933796, + 0.015943273901939392, + 0.008778284303843975, + 0.011846302077174187, + 0.009223767556250095, + 0.013719867914915085, + 0.011823012493550777, + 0.0043396116234362125, + 0.005183939356356859, + 0.005835800897330046, + 0.00437470618635416, + 0.010947383008897305, + 0.003584343707188964, + 0.006070367060601711, + 0.008307462558150291, + 0.012564376927912235, + 0.0077750724740326405, + 0.010478077456355095, + 0.009975546039640903, + 0.03595192730426788, + 0.06326527148485184, + 0.016211841255426407, + 0.028240656480193138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.17994770407676697 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank rolling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's armor painted with a deep olive-green matte camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank's surface covered in splashing mud and streaks of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tank traverse a rain-soaked dirt road while churning deep, flooded ruts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011086412705481052 + ], + "opposite_logprob_tokens": [ + -4.511086463928223 + ], + "logprob_reward": -0.011086412705481052, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011086412705481052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08822621405124664, + 0.5635093450546265, + 0.04306923225522041, + 0.01878618448972702, + 0.02272970788180828, + 0.04759463667869568, + 0.10678096860647202, + 0.027652565389871597, + 0.024026690050959587, + 0.05762448534369469 + ], + "credit_map": [ + 0.01034395769238472, + 0.005565153434872627, + 0.024280909448862076, + 0.011874828487634659, + 0.006233528256416321, + 0.005490007810294628, + 0.00969802588224411, + 0.014739804901182652, + 0.09125024825334549, + 0.047095537185668945, + 0.18825440108776093, + 0.08128053694963455, + 0.04225083813071251, + 0.029815806075930595, + 0.04404992237687111, + 0.039512064307928085, + 0.0039013195782899857, + 0.003511171555146575, + 0.009713470004498959, + 0.010845066979527473, + 0.0029775055591017008, + 0.002368481131270528, + 0.004712651018053293, + 0.005039568990468979, + 0.0024264438543468714, + 0.001644589239731431, + 0.0033769619185477495, + 0.0033052819781005383, + 0.0016247292514890432, + 0.0014074743958190084, + 0.0020175252575427294, + 0.0029831775464117527, + 0.0024604075588285923, + 0.002899880986660719, + 0.003560195676982403, + 0.0026724040508270264, + 0.0018914035754278302, + 0.0022148676216602325, + 0.003847965970635414, + 0.0031825811602175236, + 0.007488900329917669, + 0.004156805109232664, + 0.006450083572417498, + 0.004201498348265886, + 0.0041703106835484505, + 0.0057114544324576855, + 0.007121899630874395, + 0.00829368270933628, + 0.020461715757846832, + 0.009744578041136265, + 0.016716204583644867, + 0.010224423371255398, + 0.009494024328887463, + 0.008849816396832466, + 0.013331837952136993, + 0.017958370968699455, + 0.004220152273774147, + 0.0024913367815315723, + 0.004319386091083288, + 0.0028505672235041857, + 0.0033247629180550575, + 0.0025070675183087587, + 0.003266272135078907, + 0.004673020448535681, + 0.0026365751400589943, + 0.0021779111120849848, + 0.0031175417825579643, + 0.0016950668068602681, + 0.005376526154577732, + 0.0021173059940338135, + 0.0030758308712393045, + 0.003829933935776353, + 0.0038897963240742683, + 0.0037806641776114702, + 0.005649795290082693, + 0.005220102146267891, + 0.011574642732739449, + 0.00945067685097456, + 0.007562838960438967, + 0.010495963506400585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.011086412705481052 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank rolling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's armor painted with a deep olive-green matte camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank's surface covered in splashing mud and streaks of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tank traverse a rain-soaked dirt road while churning deep, flooded ruts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22561785578727722 + ], + "opposite_logprob_tokens": [ + -1.6006178855895996 + ], + "logprob_reward": -0.22561785578727722, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22561785578727722, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057207148522138596, + 0.09804435074329376, + 0.059660639613866806, + 0.016056159511208534, + 0.03768390789628029, + 0.057013511657714844, + 0.2176768034696579, + 0.11069335043430328, + 0.29576054215431213, + 0.05020354688167572 + ], + "credit_map": [ + 0.007267637178301811, + 0.003372019389644265, + 0.012529302388429642, + 0.007633027154952288, + 0.003204306587576866, + 0.005271727219223976, + 0.007412200793623924, + 0.010516928508877754, + 0.015749560669064522, + 0.00802468042820692, + 0.03289945796132088, + 0.012793317437171936, + 0.007628006394952536, + 0.005123939830809832, + 0.00833733007311821, + 0.007488054223358631, + 0.006298467516899109, + 0.006072192452847958, + 0.0122104212641716, + 0.009099755436182022, + 0.005051273852586746, + 0.003592012682929635, + 0.009469966404139996, + 0.007866552099585533, + 0.0018335403874516487, + 0.0016018023015931249, + 0.002435204805806279, + 0.0031429878436028957, + 0.0014724263455718756, + 0.0012957717990502715, + 0.002147748600691557, + 0.0021266788244247437, + 0.0051773712038993835, + 0.003843069076538086, + 0.004918679129332304, + 0.003947199322283268, + 0.002924535423517227, + 0.003978492226451635, + 0.004508757498115301, + 0.008385805413126945, + 0.008144424296915531, + 0.00800185278058052, + 0.008966933935880661, + 0.004771812353283167, + 0.0049170153215527534, + 0.004635845776647329, + 0.0069509390741586685, + 0.010624689050018787, + 0.03191177174448967, + 0.020646357908844948, + 0.031471818685531616, + 0.018356934189796448, + 0.028315991163253784, + 0.01636425033211708, + 0.030115298926830292, + 0.0404943972826004, + 0.01440443005412817, + 0.009982379153370857, + 0.013747857883572578, + 0.009730607271194458, + 0.013003027066588402, + 0.01739676296710968, + 0.017362019047141075, + 0.01506627257913351, + 0.02946469746530056, + 0.026723135262727737, + 0.04178431257605553, + 0.029034728184342384, + 0.0478360652923584, + 0.028140483424067497, + 0.04897565022110939, + 0.04380147531628609, + 0.004971949849277735, + 0.004536163527518511, + 0.005736494902521372, + 0.004428913351148367, + 0.00841092225164175, + 0.0074887508526444435, + 0.008699057623744011, + 0.005931296851485968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.22561785578727722 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank rolling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's armor painted with a deep olive-green matte camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank's surface covered in splashing mud and streaks of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tank traverse a rain-soaked dirt road while churning deep, flooded ruts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978671945631504 + ], + "opposite_logprob_tokens": [ + -4.634786605834961 + ], + "logprob_reward": -0.00978671945631504, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978671945631504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10877668857574463, + 0.4480397403240204, + 0.0964886024594307, + 0.02695196121931076, + 0.019590789452195168, + 0.04930220544338226, + 0.053666193038225174, + 0.05061984434723854, + 0.07663880288600922, + 0.06992510706186295 + ], + "credit_map": [ + 0.010592733509838581, + 0.006594462785869837, + 0.024248355999588966, + 0.01776891015470028, + 0.01017291471362114, + 0.013154851272702217, + 0.01039817463606596, + 0.015846284106373787, + 0.0687331035733223, + 0.03768455237150192, + 0.13813982903957367, + 0.06338223069906235, + 0.0343024842441082, + 0.025264466181397438, + 0.042533304542303085, + 0.037999771535396576, + 0.010398993268609047, + 0.007922898977994919, + 0.02121557667851448, + 0.01801968924701214, + 0.006424003280699253, + 0.006001938134431839, + 0.01406847033649683, + 0.012437035329639912, + 0.0034907713998109102, + 0.002201620489358902, + 0.0051520466804504395, + 0.0051359012722969055, + 0.0020364834927022457, + 0.0023389444686472416, + 0.00280597060918808, + 0.0037902239710092545, + 0.0024808465968817472, + 0.0025320982094854116, + 0.003106271382421255, + 0.0021965622436255217, + 0.0015470328507944942, + 0.0021877274848520756, + 0.002837888663634658, + 0.0027023626025766134, + 0.00808049738407135, + 0.004492184612900019, + 0.006964037660509348, + 0.005014204885810614, + 0.004090816248208284, + 0.00611356645822525, + 0.006461671087890863, + 0.008085227571427822, + 0.007663334719836712, + 0.005407008808106184, + 0.009119974449276924, + 0.005516623146831989, + 0.004919824656099081, + 0.003966441377997398, + 0.006593428552150726, + 0.010479557327926159, + 0.00898679532110691, + 0.00453362287953496, + 0.008911315351724625, + 0.005164189729839563, + 0.005916394758969545, + 0.00415672967210412, + 0.006399942096322775, + 0.0065508512780070305, + 0.007049438077956438, + 0.006291447672992945, + 0.010925399139523506, + 0.005837472155690193, + 0.015765072777867317, + 0.0064603243954479694, + 0.010702953673899174, + 0.013606695458292961, + 0.008673862554132938, + 0.005965148564428091, + 0.006975315511226654, + 0.005727679468691349, + 0.014679338783025742, + 0.010521288961172104, + 0.00950764399021864, + 0.007874831557273865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.00978671945631504 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tracking shot, shot from an extremely low angle, follows a heavy main battle tank as it rolls forward with a heavy, unstoppable momentum. The tank's massive steel armor, covered in a deep olive-green matte camouflage paint, is stained with splashing mud and streaks of rain. It traverses a rain-soaked dirt road. The diffuse light from the overcast sky casts a cold, shimmering light on the damp metal and the muddy ground. Each heavy rotation of its broad metal tracks deeply churns, squeezing, and tumbling the soft, slippery, dark brown earth beneath them, leaving behind two distinct, deep, and rapidly flooded ruts. Accompanied by the deafening roar of the engine and the low crunch of the tracks, black exhaust fumes billow from the rear exhaust pipe. The entire scene is imbued with a sense of oppression and cold military might, creating an atmosphere of austerity and unstoppable force.", + "seed": 28577, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_0/index_7542_seed_28577_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05678782537579537, + "vlm_logprob": -0.05678782537579537, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.05678782537579537, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank rolling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's armor painted with a deep olive-green matte camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank's surface covered in splashing mud and streaks of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tank traverse a rain-soaked dirt road while churning deep, flooded ruts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09868581593036652 + ], + "opposite_logprob_tokens": [ + -2.4736857414245605 + ], + "logprob_reward": -0.09868581593036652, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09868581593036652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12391670048236847, + 0.13062617182731628, + 0.16213573515415192, + 0.07912798225879669, + 0.16304513812065125, + 0.07517630606889725, + 0.056332770735025406, + 0.06643551588058472, + 0.05026458203792572, + 0.09293908625841141 + ], + "credit_map": [ + 0.012300914153456688, + 0.01284504309296608, + 0.010815142653882504, + 0.013494756072759628, + 0.013431530445814133, + 0.013086200691759586, + 0.014281927607953548, + 0.03366119787096977, + 0.011235139332711697, + 0.016742926090955734, + 0.014371365308761597, + 0.0308942012488842, + 0.008357428014278412, + 0.008245456032454967, + 0.024738620966672897, + 0.016041036695241928, + 0.01738486811518669, + 0.0202699676156044, + 0.019263317808508873, + 0.018832096830010414, + 0.014047693461179733, + 0.01602870784699917, + 0.03260823339223862, + 0.023700859397649765, + 0.009424371644854546, + 0.013802207075059414, + 0.005133240483701229, + 0.00652918079867959, + 0.015151710249483585, + 0.009442990645766258, + 0.008527829311788082, + 0.011116456240415573, + 0.016619155183434486, + 0.03234140947461128, + 0.01117891725152731, + 0.008570604026317596, + 0.02167460322380066, + 0.03756129369139671, + 0.015803810209035873, + 0.01929536648094654, + 0.008261837996542454, + 0.015334236435592175, + 0.006006292067468166, + 0.00547216460108757, + 0.011764234863221645, + 0.014801518991589546, + 0.00639595091342926, + 0.0071400729939341545, + 0.006166873034089804, + 0.008592997677624226, + 0.007013039197772741, + 0.005295958369970322, + 0.010314466431736946, + 0.00673279631882906, + 0.004485012032091618, + 0.007731631398200989, + 0.0038587867747992277, + 0.0068626245483756065, + 0.01016248669475317, + 0.005486023612320423, + 0.013065172359347343, + 0.010576650500297546, + 0.006557553540915251, + 0.009866224601864815, + 0.002926002023741603, + 0.006806503981351852, + 0.009624497033655643, + 0.004248809535056353, + 0.005449480377137661, + 0.009235743433237076, + 0.005437841638922691, + 0.006535708904266357, + 0.007128630764782429, + 0.01077280193567276, + 0.007537445519119501, + 0.013633573427796364, + 0.013543669134378433, + 0.014330833218991756, + 0.009536584839224815, + 0.016455553472042084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.09868581593036652 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank rolling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's armor painted with a deep olive-green matte camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank's surface covered in splashing mud and streaks of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tank traverse a rain-soaked dirt road while churning deep, flooded ruts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05500134825706482 + ], + "opposite_logprob_tokens": [ + -2.9300012588500977 + ], + "logprob_reward": -0.05500134825706482, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05500134825706482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06898068636655807, + 0.02948513627052307, + 0.14410245418548584, + 0.12253774702548981, + 0.05125418305397034, + 0.14673659205436707, + 0.3066595792770386, + 0.018976425752043724, + 0.043375976383686066, + 0.0678912028670311 + ], + "credit_map": [ + 0.008379500359296799, + 0.00730523094534874, + 0.005642569623887539, + 0.006182059645652771, + 0.010059845633804798, + 0.007625400088727474, + 0.006877983920276165, + 0.016908107325434685, + 0.0030898095574229956, + 0.004062017425894737, + 0.0026841408107429743, + 0.0070066070184111595, + 0.0015853248769417405, + 0.001988187897950411, + 0.005429650656878948, + 0.0036393990740180016, + 0.015022281557321548, + 0.03664622828364372, + 0.01786038838326931, + 0.01702539250254631, + 0.013002890162169933, + 0.01403121929615736, + 0.017668116837739944, + 0.012845934368669987, + 0.012501518242061138, + 0.021319765597581863, + 0.010684476234018803, + 0.013269100338220596, + 0.01805974543094635, + 0.015368196181952953, + 0.013247127644717693, + 0.018087832257151604, + 0.005537407472729683, + 0.010729039087891579, + 0.004019089043140411, + 0.002807720098644495, + 0.008108326233923435, + 0.011066005565226078, + 0.0037032112013548613, + 0.005283385049551725, + 0.015311634168028831, + 0.016270607709884644, + 0.0107648316770792, + 0.006596051622182131, + 0.0290275476872921, + 0.04588325694203377, + 0.010119552724063396, + 0.012763130478560925, + 0.03704383969306946, + 0.055860839784145355, + 0.034624096006155014, + 0.029649406671524048, + 0.0474979504942894, + 0.04100935906171799, + 0.026429761201143265, + 0.034544333815574646, + 0.001337374560534954, + 0.0021708242129534483, + 0.002666800981387496, + 0.001779265352524817, + 0.003972526174038649, + 0.0029119839891791344, + 0.0015104220947250724, + 0.002627230016514659, + 0.0040772780776023865, + 0.0078097800724208355, + 0.005689237266778946, + 0.0044203512370586395, + 0.004146904684603214, + 0.007520132232457399, + 0.004320414271205664, + 0.0053918808698654175, + 0.004532604478299618, + 0.010120394639670849, + 0.007111030630767345, + 0.008015399798750877, + 0.010478299111127853, + 0.013687456026673317, + 0.006058715749531984, + 0.007887307554483414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 2 + ], + "selected_reward": -0.05500134825706482 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank rolling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's armor painted with a deep olive-green matte camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank's surface covered in splashing mud and streaks of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tank traverse a rain-soaked dirt road while churning deep, flooded ruts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020602626726031303 + ], + "opposite_logprob_tokens": [ + -3.8956027030944824 + ], + "logprob_reward": -0.020602626726031303, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020602626726031303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17989876866340637, + 0.18698610365390778, + 0.09383395314216614, + 0.03778667747974396, + 0.0428754948079586, + 0.036067135632038116, + 0.12356589734554291, + 0.10546322911977768, + 0.0626474916934967, + 0.13087528944015503 + ], + "credit_map": [ + 0.019696002826094627, + 0.02078239619731903, + 0.016338473185896873, + 0.0207276102155447, + 0.018709812313318253, + 0.017436224967241287, + 0.025906482711434364, + 0.040301743894815445, + 0.018571890890598297, + 0.027067219838500023, + 0.018779011443257332, + 0.039283689111471176, + 0.012598535977303982, + 0.011201220564544201, + 0.03571048006415367, + 0.023774024099111557, + 0.00955432653427124, + 0.016990743577480316, + 0.01414803322404623, + 0.012697950005531311, + 0.007190796080976725, + 0.007156759034842253, + 0.013987562619149685, + 0.01210777461528778, + 0.00408831425011158, + 0.006988551001995802, + 0.0028671431355178356, + 0.002900490304455161, + 0.0071167172864079475, + 0.004832802806049585, + 0.0038895816542208195, + 0.005103071685880423, + 0.0064657023176550865, + 0.006038295570760965, + 0.0036338609643280506, + 0.002795954467728734, + 0.009599060751497746, + 0.005938402842730284, + 0.003295480040833354, + 0.005108732730150223, + 0.004497216083109379, + 0.006525810342282057, + 0.0030113861430436373, + 0.002645047614350915, + 0.005107775330543518, + 0.006850431207567453, + 0.003067356301471591, + 0.004362108185887337, + 0.01668393984436989, + 0.028064820915460587, + 0.012452500872313976, + 0.010223600082099438, + 0.020760895684361458, + 0.01453369203954935, + 0.009439636953175068, + 0.01140680257230997, + 0.00633390573784709, + 0.012060609646141529, + 0.01242851559072733, + 0.008441178128123283, + 0.01995542272925377, + 0.01641487516462803, + 0.012982192449271679, + 0.016846517100930214, + 0.004321814980357885, + 0.008069569244980812, + 0.010910269804298878, + 0.009212221018970013, + 0.007390924263745546, + 0.006376295816153288, + 0.008822535164654255, + 0.007543855346739292, + 0.0077326358295977116, + 0.017826957628130913, + 0.012742938473820686, + 0.015867264941334724, + 0.016495447605848312, + 0.030899429693818092, + 0.013291397131979465, + 0.01601920649409294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.020602626726031303 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank rolling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's armor painted with a deep olive-green matte camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank's surface covered in splashing mud and streaks of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tank traverse a rain-soaked dirt road while churning deep, flooded ruts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906247466802597 + ], + "opposite_logprob_tokens": [ + -2.464062452316284 + ], + "logprob_reward": -0.08906247466802597, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906247466802597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10064690560102463, + 0.09048675000667572, + 0.08084118366241455, + 0.08061084151268005, + 0.0787031427025795, + 0.19361069798469543, + 0.043897926807403564, + 0.10684043914079666, + 0.07521405071020126, + 0.149148091673851 + ], + "credit_map": [ + 0.012980444356799126, + 0.01234268955886364, + 0.007982620969414711, + 0.011305454187095165, + 0.010376673191785812, + 0.01098158210515976, + 0.011286442168056965, + 0.0233909972012043, + 0.008701052516698837, + 0.011494731530547142, + 0.00802003312855959, + 0.026586437597870827, + 0.005368570331484079, + 0.006019150372594595, + 0.013588465750217438, + 0.010708313435316086, + 0.009990361519157887, + 0.010979949496686459, + 0.012504530139267445, + 0.010952796787023544, + 0.007432535756379366, + 0.005192222539335489, + 0.01344247069209814, + 0.010346313007175922, + 0.00873066671192646, + 0.01223448570817709, + 0.008973532356321812, + 0.010184530168771744, + 0.011700187809765339, + 0.006997859571129084, + 0.009811046533286572, + 0.011978531256318092, + 0.009555741213262081, + 0.017369724810123444, + 0.006643299479037523, + 0.0049691107124090195, + 0.011929504573345184, + 0.015162010677158833, + 0.004187882877886295, + 0.008885867893695831, + 0.023890996351838112, + 0.03410375863313675, + 0.01892365701496601, + 0.015483622439205647, + 0.02524525858461857, + 0.04115729406476021, + 0.017693374305963516, + 0.0171127300709486, + 0.004698998294770718, + 0.005258762743324041, + 0.005483037326484919, + 0.0031541085336357355, + 0.01100135874003172, + 0.00623357342556119, + 0.003539217868819833, + 0.004528871271759272, + 0.011198965832591057, + 0.015109091997146606, + 0.01887046918272972, + 0.009854074567556381, + 0.012613809667527676, + 0.01468820683658123, + 0.009724323637783527, + 0.014781494624912739, + 0.00718965008854866, + 0.0158015675842762, + 0.011312422342598438, + 0.00829283520579338, + 0.008927201852202415, + 0.01193649135529995, + 0.0057138483971357346, + 0.006040036678314209, + 0.009901419281959534, + 0.01993739977478981, + 0.015689311549067497, + 0.02048436924815178, + 0.016719218343496323, + 0.025158746168017387, + 0.01571502350270748, + 0.02554260939359665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 7 + ], + "selected_reward": -0.08906247466802597 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank rolling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's armor painted with a deep olive-green matte camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank's surface covered in splashing mud and streaks of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tank traverse a rain-soaked dirt road while churning deep, flooded ruts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020586861297488213 + ], + "opposite_logprob_tokens": [ + -3.8955869674682617 + ], + "logprob_reward": -0.020586861297488213, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020586861297488213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19417300820350647, + 0.06051451712846756, + 0.05944223701953888, + 0.07017045468091965, + 0.13715805113315582, + 0.0979871079325676, + 0.0784778818488121, + 0.1317131370306015, + 0.1127704605460167, + 0.057593099772930145 + ], + "credit_map": [ + 0.030266311019659042, + 0.02347947098314762, + 0.01992380991578102, + 0.023113105446100235, + 0.024964338168501854, + 0.01750364527106285, + 0.02881755493581295, + 0.02610478363931179, + 0.0053977943025529385, + 0.008803865872323513, + 0.007005424238741398, + 0.012091083452105522, + 0.003718709573149681, + 0.0036806545685976744, + 0.011167607270181179, + 0.00864938460290432, + 0.005672798026353121, + 0.011931774206459522, + 0.00628624577075243, + 0.007243217434734106, + 0.0051528108306229115, + 0.004535440821200609, + 0.009042870253324509, + 0.009577086195349693, + 0.00954275019466877, + 0.012455225922167301, + 0.005907122045755386, + 0.0069450740702450275, + 0.011963617987930775, + 0.005573972128331661, + 0.00786393042653799, + 0.009918766096234322, + 0.0184672512114048, + 0.02431149035692215, + 0.010521906428039074, + 0.0077348253689706326, + 0.025119539350271225, + 0.02504783868789673, + 0.008965110406279564, + 0.01699010282754898, + 0.009336414746940136, + 0.014381990768015385, + 0.007534970995038748, + 0.007394467946141958, + 0.01451084390282631, + 0.025115132331848145, + 0.010702910833060741, + 0.009010382927954197, + 0.009543283842504025, + 0.014876987785100937, + 0.008590605109930038, + 0.006578075233846903, + 0.013940238393843174, + 0.00973275862634182, + 0.00658345315605402, + 0.008632482960820198, + 0.009386515244841576, + 0.012000852264463902, + 0.017145095393061638, + 0.010786265134811401, + 0.024239245802164078, + 0.01984412595629692, + 0.01573348604142666, + 0.022577553987503052, + 0.006780415773391724, + 0.01436134334653616, + 0.02264401502907276, + 0.015024065971374512, + 0.013082817196846008, + 0.011124799028038979, + 0.015337205491960049, + 0.014415806159377098, + 0.0037036449648439884, + 0.00868659932166338, + 0.00539009366184473, + 0.0055639734491705894, + 0.007920795120298862, + 0.013302999548614025, + 0.005811444018036127, + 0.0072135538794100285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.020586861297488213 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tracking shot, shot from an extremely low angle, follows a heavy main battle tank as it rolls forward with a heavy, unstoppable momentum. The tank's massive steel armor, covered in a deep olive-green matte camouflage paint, is stained with splashing mud and streaks of rain. It traverses a rain-soaked dirt road. The diffuse light from the overcast sky casts a cold, shimmering light on the damp metal and the muddy ground. Each heavy rotation of its broad metal tracks deeply churns, squeezing, and tumbling the soft, slippery, dark brown earth beneath them, leaving behind two distinct, deep, and rapidly flooded ruts. Accompanied by the deafening roar of the engine and the low crunch of the tracks, black exhaust fumes billow from the rear exhaust pipe. The entire scene is imbued with a sense of oppression and cold military might, creating an atmosphere of austerity and unstoppable force.", + "seed": 28578, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_0/index_7542_seed_28578_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04625261202454567, + "vlm_logprob": -0.04625261202454567, + "vlm_yes_no_margin": 3.3249999046325684, + "vlm_token_credit": -0.04625261202454567, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank rolling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's armor painted with a deep olive-green matte camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank's surface covered in splashing mud and streaks of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tank traverse a rain-soaked dirt road while churning deep, flooded ruts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06470508128404617 + ], + "opposite_logprob_tokens": [ + -2.9397051334381104 + ], + "logprob_reward": -0.06470508128404617, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06470508128404617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12805265188217163, + 0.1972101479768753, + 0.13429798185825348, + 0.08504912257194519, + 0.08558181673288345, + 0.03669367730617523, + 0.07949996739625931, + 0.11377999186515808, + 0.09460214525461197, + 0.045232564210891724 + ], + "credit_map": [ + 0.014695634134113789, + 0.025798438116908073, + 0.01511123962700367, + 0.012711646035313606, + 0.01652372255921364, + 0.012022406794130802, + 0.01896709017455578, + 0.01222246140241623, + 0.014656654559075832, + 0.031017370522022247, + 0.019269319251179695, + 0.02687959000468254, + 0.025556344538927078, + 0.012000160291790962, + 0.03660818561911583, + 0.031222499907016754, + 0.010755530558526516, + 0.01996353641152382, + 0.01842477172613144, + 0.009344376623630524, + 0.025284666568040848, + 0.008883915841579437, + 0.021564539521932602, + 0.020076628774404526, + 0.008863127790391445, + 0.020796610042452812, + 0.005265375599265099, + 0.005994620267301798, + 0.009333057329058647, + 0.005524784792214632, + 0.008813139982521534, + 0.0204584039747715, + 0.010435772128403187, + 0.013106108643114567, + 0.012292912229895592, + 0.014092469587922096, + 0.008270373567938805, + 0.007626222912222147, + 0.011154416017234325, + 0.008603538386523724, + 0.0035591633059084415, + 0.00256290752440691, + 0.0050789667293429375, + 0.005583282560110092, + 0.006240068469196558, + 0.005038811359554529, + 0.0032674800604581833, + 0.005362994968891144, + 0.007806208450347185, + 0.007297482807189226, + 0.013415185734629631, + 0.012542097829282284, + 0.00820205733180046, + 0.007721489295363426, + 0.009957910515367985, + 0.012557527981698513, + 0.011602476239204407, + 0.010998486541211605, + 0.019024675711989403, + 0.012268699705600739, + 0.012803480960428715, + 0.014594954438507557, + 0.00965868216007948, + 0.022828521206974983, + 0.010567892342805862, + 0.010215633548796177, + 0.01372407004237175, + 0.00884181447327137, + 0.023949990049004555, + 0.00624126335605979, + 0.007364868186414242, + 0.01369659323245287, + 0.0060356007888913155, + 0.0046242219395935535, + 0.004386184271425009, + 0.0046720923855900764, + 0.008560160174965858, + 0.0029991157352924347, + 0.004144737962633371, + 0.009810445830225945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.06470508128404617 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank rolling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's armor painted with a deep olive-green matte camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank's surface covered in splashing mud and streaks of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tank traverse a rain-soaked dirt road while churning deep, flooded ruts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029828058555722237 + ], + "opposite_logprob_tokens": [ + -3.5298280715942383 + ], + "logprob_reward": -0.029828058555722237, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029828058555722237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10098571330308914, + 0.15230877697467804, + 0.046163614839315414, + 0.040126606822013855, + 0.17634762823581696, + 0.05584866553544998, + 0.024753978475928307, + 0.08833134919404984, + 0.16757097840309143, + 0.14756271243095398 + ], + "credit_map": [ + 0.007667940575629473, + 0.020452167838811874, + 0.01116618886590004, + 0.015253948979079723, + 0.009892535395920277, + 0.0077040228061378, + 0.015549678355455399, + 0.01329922117292881, + 0.01141322311013937, + 0.03271044045686722, + 0.015224463306367397, + 0.01692778244614601, + 0.019486786797642708, + 0.013404776342213154, + 0.0201512910425663, + 0.0229900274425745, + 0.0036293102893978357, + 0.0055801039561629295, + 0.00508627574890852, + 0.004824021365493536, + 0.009867350570857525, + 0.002970317844301462, + 0.006471866741776466, + 0.007734367623925209, + 0.005002535879611969, + 0.0072172158397734165, + 0.0035576927475631237, + 0.0044818357564508915, + 0.00451244181022048, + 0.0048857261426746845, + 0.004969473462551832, + 0.005499681923538446, + 0.02061435952782631, + 0.027330413460731506, + 0.023157961666584015, + 0.01663903333246708, + 0.02334393374621868, + 0.02155354991555214, + 0.027278728783130646, + 0.01642964407801628, + 0.0052657597698271275, + 0.005998380482196808, + 0.008450455032289028, + 0.005809037480503321, + 0.010481174103915691, + 0.008066854439675808, + 0.005513230804353952, + 0.00626377435401082, + 0.002387160900980234, + 0.0031389230862259865, + 0.0022503877989947796, + 0.002019425854086876, + 0.006328051444143057, + 0.0025451299734413624, + 0.003302555298432708, + 0.002782344352453947, + 0.011489438824355602, + 0.009306933730840683, + 0.015888100489974022, + 0.009857933968305588, + 0.012555499561131, + 0.010463746264576912, + 0.00718690175563097, + 0.01158280111849308, + 0.013423026539385319, + 0.017443932592868805, + 0.01423634309321642, + 0.010615047067403793, + 0.06206737458705902, + 0.013517782092094421, + 0.012223017401993275, + 0.024044446647167206, + 0.01614248752593994, + 0.014524487778544426, + 0.019619451835751534, + 0.012146830558776855, + 0.04091878980398178, + 0.012339185923337936, + 0.012049755081534386, + 0.019821718335151672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 1 + ], + "selected_reward": -0.029828058555722237 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank rolling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's armor painted with a deep olive-green matte camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank's surface covered in splashing mud and streaks of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tank traverse a rain-soaked dirt road while churning deep, flooded ruts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018200064077973366 + ], + "opposite_logprob_tokens": [ + -4.018199920654297 + ], + "logprob_reward": -0.018200064077973366, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018200064077973366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16420382261276245, + 0.13882768154144287, + 0.06381365656852722, + 0.15337379276752472, + 0.18396781384944916, + 0.04904855415225029, + 0.02644994854927063, + 0.0349867083132267, + 0.09567950665950775, + 0.08964853733778 + ], + "credit_map": [ + 0.015257992781698704, + 0.035152606666088104, + 0.017921891063451767, + 0.021219613030552864, + 0.01620177924633026, + 0.011646566912531853, + 0.0239446721971035, + 0.022858701646327972, + 0.009659969247877598, + 0.018826723098754883, + 0.013868200592696667, + 0.01384296640753746, + 0.01792777143418789, + 0.0158421378582716, + 0.02911989949643612, + 0.01974000595510006, + 0.004340567160397768, + 0.009426245465874672, + 0.009631142020225525, + 0.005799420177936554, + 0.011246786452829838, + 0.0036136654671281576, + 0.008425421081483364, + 0.01133040338754654, + 0.01615193672478199, + 0.03504570201039314, + 0.010478462092578411, + 0.009022698737680912, + 0.013336806558072567, + 0.009775277227163315, + 0.014610910788178444, + 0.04495198652148247, + 0.01926044374704361, + 0.02154393121600151, + 0.02750338613986969, + 0.0232012290507555, + 0.020853186026215553, + 0.0257717315107584, + 0.026009678840637207, + 0.01982422173023224, + 0.004516598302870989, + 0.006609820295125246, + 0.00804615393280983, + 0.0076256487518548965, + 0.004780526738613844, + 0.005473889410495758, + 0.004547633696347475, + 0.007448284886777401, + 0.0025082856882363558, + 0.003470703726634383, + 0.0032797346357256174, + 0.003494884120300412, + 0.006286100018769503, + 0.0020455175545066595, + 0.002590223914012313, + 0.002774498425424099, + 0.003818223485723138, + 0.0034187522251158953, + 0.003656739601865411, + 0.0031596359331160784, + 0.008495901711285114, + 0.003736990736797452, + 0.0033303198870271444, + 0.005370143335312605, + 0.011278358288109303, + 0.01205221377313137, + 0.013657180592417717, + 0.010161628015339375, + 0.01706685498356819, + 0.006966296583414078, + 0.008873403072357178, + 0.01562357135117054, + 0.009851595386862755, + 0.013361534103751183, + 0.010998979210853577, + 0.008873829618096352, + 0.020224692299962044, + 0.007949627935886383, + 0.008312869817018509, + 0.010075412690639496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.018200064077973366 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank rolling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's armor painted with a deep olive-green matte camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank's surface covered in splashing mud and streaks of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tank traverse a rain-soaked dirt road while churning deep, flooded ruts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10034266859292984 + ], + "opposite_logprob_tokens": [ + -2.3503427505493164 + ], + "logprob_reward": -0.10034266859292984, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10034266859292984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10123080015182495, + 0.21507027745246887, + 0.06742553412914276, + 0.0360199436545372, + 0.06316094100475311, + 0.0796719491481781, + 0.06469139456748962, + 0.15945395827293396, + 0.16691721975803375, + 0.04635798931121826 + ], + "credit_map": [ + 0.008666936308145523, + 0.0183077584952116, + 0.014036919921636581, + 0.013149652630090714, + 0.00944608822464943, + 0.007901701144874096, + 0.017907166853547096, + 0.01181457843631506, + 0.015473807230591774, + 0.033030930906534195, + 0.03151911497116089, + 0.027058275416493416, + 0.02266716957092285, + 0.01742587424814701, + 0.036549802869558334, + 0.03134531155228615, + 0.00878830999135971, + 0.008404604159295559, + 0.009790394455194473, + 0.005826069973409176, + 0.010619161650538445, + 0.005198129452764988, + 0.008821774274110794, + 0.009977092035114765, + 0.004192113410681486, + 0.004444099497050047, + 0.003914158791303635, + 0.004516344051808119, + 0.004800165072083473, + 0.0025194240733981133, + 0.006067886482924223, + 0.005565756931900978, + 0.0070973834954202175, + 0.01232893019914627, + 0.006031286437064409, + 0.008716437965631485, + 0.009828941896557808, + 0.005852291360497475, + 0.007199106737971306, + 0.006106555927544832, + 0.006974856369197369, + 0.008371446281671524, + 0.010779865086078644, + 0.018723979592323303, + 0.00937587022781372, + 0.005052161403000355, + 0.009928450919687748, + 0.010465318337082863, + 0.005617112386971712, + 0.007029240485280752, + 0.011781612411141396, + 0.009126666933298111, + 0.008277825079858303, + 0.006295858416706324, + 0.007049460429698229, + 0.009513617493212223, + 0.020034320652484894, + 0.01675027422606945, + 0.03171880170702934, + 0.01725178398191929, + 0.017392225563526154, + 0.01960662193596363, + 0.013563393615186214, + 0.02313653565943241, + 0.021344609558582306, + 0.020614290609955788, + 0.016426118090748787, + 0.0151980584487319, + 0.03386460244655609, + 0.014021198265254498, + 0.011174830608069897, + 0.034273505210876465, + 0.0047691999934613705, + 0.0050260089337825775, + 0.004976286087185144, + 0.00503486767411232, + 0.011772713623940945, + 0.004385452717542648, + 0.0049440814182162285, + 0.005449377931654453 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.10034266859292984 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank rolling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's armor painted with a deep olive-green matte camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank's surface covered in splashing mud and streaks of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tank traverse a rain-soaked dirt road while churning deep, flooded ruts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018187187612056732 + ], + "opposite_logprob_tokens": [ + -4.018187046051025 + ], + "logprob_reward": -0.018187187612056732, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018187187612056732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1805693507194519, + 0.15519870817661285, + 0.09549620747566223, + 0.09952182322740555, + 0.08561918884515762, + 0.11790934950113297, + 0.07302939891815186, + 0.06861969828605652, + 0.06407393515110016, + 0.05996231734752655 + ], + "credit_map": [ + 0.014691447839140892, + 0.035286929458379745, + 0.022375237196683884, + 0.02348921447992325, + 0.01503513753414154, + 0.013627317734062672, + 0.029496068134903908, + 0.02656800113618374, + 0.011270018294453621, + 0.028041670098900795, + 0.01597139798104763, + 0.01781208999454975, + 0.02179541625082493, + 0.012936627492308617, + 0.024904267862439156, + 0.022467218339443207, + 0.00654529919847846, + 0.017083637416362762, + 0.009210938587784767, + 0.008736279793083668, + 0.016790524125099182, + 0.005751293618232012, + 0.012212066911160946, + 0.019166165962815285, + 0.009977148845791817, + 0.02425714023411274, + 0.006717895157635212, + 0.006419303361326456, + 0.010884551331400871, + 0.005886099301278591, + 0.009941598400473595, + 0.025438087061047554, + 0.00814459566026926, + 0.012353201396763325, + 0.009736020117998123, + 0.009800235740840435, + 0.01135463360697031, + 0.013882866129279137, + 0.012037434615194798, + 0.008310196921229362, + 0.011351522989571095, + 0.016415927559137344, + 0.019391480833292007, + 0.013131855987012386, + 0.019613705575466156, + 0.008695829659700394, + 0.011765610426664352, + 0.01754341647028923, + 0.007431610953062773, + 0.008333160541951656, + 0.010303545743227005, + 0.009648114442825317, + 0.01040872186422348, + 0.007319096475839615, + 0.008083775639533997, + 0.011501369997859001, + 0.0062696547247469425, + 0.006002387497574091, + 0.008123287931084633, + 0.007634470239281654, + 0.009442256763577461, + 0.008633541874587536, + 0.006430001929402351, + 0.016084099188447, + 0.007600038778036833, + 0.009250737726688385, + 0.011856677941977978, + 0.006631666794419289, + 0.008418175391852856, + 0.006220668088644743, + 0.0057874093763530254, + 0.008308561518788338, + 0.006409636698663235, + 0.0058862934820353985, + 0.007465770468115807, + 0.00868441816419363, + 0.007841040380299091, + 0.0045170909725129604, + 0.006588594987988472, + 0.012569471262395382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.018187187612056732 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tracking shot, shot from an extremely low angle, follows a heavy main battle tank as it rolls forward with a heavy, unstoppable momentum. The tank's massive steel armor, covered in a deep olive-green matte camouflage paint, is stained with splashing mud and streaks of rain. It traverses a rain-soaked dirt road. The diffuse light from the overcast sky casts a cold, shimmering light on the damp metal and the muddy ground. Each heavy rotation of its broad metal tracks deeply churns, squeezing, and tumbling the soft, slippery, dark brown earth beneath them, leaving behind two distinct, deep, and rapidly flooded ruts. Accompanied by the deafening roar of the engine and the low crunch of the tracks, black exhaust fumes billow from the rear exhaust pipe. The entire scene is imbued with a sense of oppression and cold military might, creating an atmosphere of austerity and unstoppable force.", + "seed": 28579, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_0/index_7542_seed_28579_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04748139604926109, + "vlm_logprob": -0.04748139604926109, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.04748139604926109, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank rolling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's armor painted with a deep olive-green matte camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank's surface covered in splashing mud and streaks of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tank traverse a rain-soaked dirt road while churning deep, flooded ruts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06100994348526001 + ], + "opposite_logprob_tokens": [ + -3.0610098838806152 + ], + "logprob_reward": -0.06100994348526001, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.06100994348526001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1393251270055771, + 0.12069631367921829, + 0.04656803235411644, + 0.10591770708560944, + 0.07293004542589188, + 0.06892065703868866, + 0.11796130239963531, + 0.13389582931995392, + 0.07857948541641235, + 0.11520548909902573 + ], + "credit_map": [ + 0.012379510328173637, + 0.015603837557137012, + 0.02148147113621235, + 0.018254293128848076, + 0.018530355766415596, + 0.013974012807011604, + 0.01695994660258293, + 0.022141700610518456, + 0.016179509460926056, + 0.00938479881733656, + 0.017967060208320618, + 0.014869581907987595, + 0.019399961456656456, + 0.01583542674779892, + 0.014908120967447758, + 0.012151855044066906, + 0.005332421511411667, + 0.007151109632104635, + 0.008214899338781834, + 0.005283358506858349, + 0.005981555674225092, + 0.005067707039415836, + 0.004769685212522745, + 0.004767296370118856, + 0.01026184856891632, + 0.010767064057290554, + 0.014099599793553352, + 0.022888274863362312, + 0.012253159657120705, + 0.010394670069217682, + 0.01051456481218338, + 0.014738529920578003, + 0.006443653721362352, + 0.008071686141192913, + 0.009372124448418617, + 0.010293488390743732, + 0.012478938326239586, + 0.010331675410270691, + 0.008768676780164242, + 0.007169797085225582, + 0.004028111230581999, + 0.004349542781710625, + 0.008736047893762589, + 0.01203002966940403, + 0.0062053874135017395, + 0.009606340900063515, + 0.013731013052165508, + 0.010234180837869644, + 0.008694135583937168, + 0.013687812723219395, + 0.025323903188109398, + 0.016760803759098053, + 0.010546647943556309, + 0.02508297748863697, + 0.008869309909641743, + 0.008995709009468555, + 0.009364837780594826, + 0.01644306816160679, + 0.02449456788599491, + 0.024149851873517036, + 0.014750535599887371, + 0.020527133718132973, + 0.013301377184689045, + 0.010864445939660072, + 0.007370592560619116, + 0.008648921735584736, + 0.02088654786348343, + 0.009792724624276161, + 0.0064414795488119125, + 0.011074733920395374, + 0.00747845321893692, + 0.006886026356369257, + 0.008322447538375854, + 0.008803257718682289, + 0.009695501998066902, + 0.013506240211427212, + 0.014909226447343826, + 0.03519020602107048, + 0.013193673454225063, + 0.011584939435124397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.06100994348526001 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank rolling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's armor painted with a deep olive-green matte camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank's surface covered in splashing mud and streaks of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tank traverse a rain-soaked dirt road while churning deep, flooded ruts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062075175344944 + ], + "opposite_logprob_tokens": [ + -2.812075138092041 + ], + "logprob_reward": -0.062075175344944, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062075175344944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048491448163986206, + 0.1581314653158188, + 0.11011356860399246, + 0.02423674985766411, + 0.0703645795583725, + 0.03819857910275459, + 0.17832620441913605, + 0.03506233170628548, + 0.2659170627593994, + 0.07115806639194489 + ], + "credit_map": [ + 0.005692497361451387, + 0.006507753860205412, + 0.004424679093062878, + 0.006830061320215464, + 0.007878482341766357, + 0.006239522714167833, + 0.004802767187356949, + 0.006115679629147053, + 0.016820961609482765, + 0.019915858283638954, + 0.022168217226862907, + 0.01695955917239189, + 0.035611752420663834, + 0.01621011458337307, + 0.016642184928059578, + 0.013802808709442616, + 0.012347450479865074, + 0.01579900085926056, + 0.014273777604103088, + 0.016347743570804596, + 0.014624016359448433, + 0.015486586838960648, + 0.010524354875087738, + 0.010710623115301132, + 0.0023762646596878767, + 0.0018890632782131433, + 0.002513219602406025, + 0.005028989631682634, + 0.0034446087665855885, + 0.001797414617612958, + 0.0032545174472033978, + 0.0039326706901192665, + 0.006537078879773617, + 0.006614144891500473, + 0.010056166909635067, + 0.009593280963599682, + 0.017204616218805313, + 0.008535877801477909, + 0.0070891063660383224, + 0.004734306130558252, + 0.0025676407385617495, + 0.002314570127055049, + 0.005338285584002733, + 0.006976663134992123, + 0.003764674300327897, + 0.005914459470659494, + 0.00674236286431551, + 0.004579918924719095, + 0.012886513024568558, + 0.01808994822204113, + 0.039025336503982544, + 0.02706402726471424, + 0.01593945361673832, + 0.03977115452289581, + 0.013556482270359993, + 0.011993258260190487, + 0.0033458757679909468, + 0.004522024188190699, + 0.005016938783228397, + 0.0043811495415866375, + 0.006649572402238846, + 0.00572873605415225, + 0.002806254895403981, + 0.0026117749512195587, + 0.024777788668870926, + 0.03006501868367195, + 0.044281844049692154, + 0.02398761734366417, + 0.02917877770960331, + 0.07405715435743332, + 0.019172469154000282, + 0.02039634995162487, + 0.004993837792426348, + 0.0067444141022861, + 0.006995099596679211, + 0.007480967789888382, + 0.008919661864638329, + 0.01966695487499237, + 0.007202235050499439, + 0.009154890663921833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 1 + ], + "selected_reward": -0.062075175344944 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank rolling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's armor painted with a deep olive-green matte camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank's surface covered in splashing mud and streaks of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tank traverse a rain-soaked dirt road while churning deep, flooded ruts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202821999788284 + ], + "opposite_logprob_tokens": [ + -4.26420259475708 + ], + "logprob_reward": -0.014202821999788284, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202821999788284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15388399362564087, + 0.08877908438444138, + 0.08011233061552048, + 0.034521471709012985, + 0.2789002060890198, + 0.03488617762923241, + 0.11774810403585434, + 0.03169960156083107, + 0.10149897634983063, + 0.0779699832201004 + ], + "credit_map": [ + 0.014008458703756332, + 0.016773464158177376, + 0.015840929001569748, + 0.01871372200548649, + 0.015531261451542377, + 0.01658651977777481, + 0.026016205549240112, + 0.030413424596190453, + 0.010585678741335869, + 0.00910962000489235, + 0.010765811428427696, + 0.013136504217982292, + 0.01096346601843834, + 0.009748810902237892, + 0.011445228941738605, + 0.013023968786001205, + 0.009072446264326572, + 0.011708660051226616, + 0.010148652829229832, + 0.014881155453622341, + 0.00807862263172865, + 0.008772430941462517, + 0.009275715798139572, + 0.008174650371074677, + 0.003339211456477642, + 0.0019909546244889498, + 0.006006188690662384, + 0.006422976963222027, + 0.005341663956642151, + 0.0037243275437504053, + 0.003324786201119423, + 0.004371362738311291, + 0.024155976250767708, + 0.021115677431225777, + 0.05468928441405296, + 0.040667466819286346, + 0.03730910271406174, + 0.046799153089523315, + 0.02998436987400055, + 0.024179207161068916, + 0.0026774443686008453, + 0.0033117220737040043, + 0.004871418233960867, + 0.006308496464043856, + 0.003001454286277294, + 0.005869683809578419, + 0.005072895437479019, + 0.0037730636540800333, + 0.007999259047210217, + 0.011358065530657768, + 0.026681119576096535, + 0.01748880185186863, + 0.010856947861611843, + 0.02475464716553688, + 0.009430917911231518, + 0.009178350679576397, + 0.0024294042959809303, + 0.0036641964688897133, + 0.004441126715391874, + 0.005382508505135775, + 0.0038747212383896112, + 0.003458714345470071, + 0.005078834015876055, + 0.0033700959756970406, + 0.00847635231912136, + 0.008541992865502834, + 0.015795178711414337, + 0.01009699422866106, + 0.011114438064396381, + 0.030531952157616615, + 0.008075215853750706, + 0.008866852149367332, + 0.006194073241204023, + 0.008184361271560192, + 0.0059174466878175735, + 0.007779892534017563, + 0.009443899616599083, + 0.026010451838374138, + 0.0074416399002075195, + 0.006998215802013874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.014202821999788284 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank rolling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's armor painted with a deep olive-green matte camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank's surface covered in splashing mud and streaks of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tank traverse a rain-soaked dirt road while churning deep, flooded ruts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904524147510529 + ], + "opposite_logprob_tokens": [ + -2.464045286178589 + ], + "logprob_reward": -0.08904524147510529, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904524147510529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13305173814296722, + 0.050859417766332626, + 0.04664018005132675, + 0.06787122786045074, + 0.07288271188735962, + 0.09343960881233215, + 0.1253223568201065, + 0.2764148712158203, + 0.09886566549539566, + 0.03465213626623154 + ], + "credit_map": [ + 0.008898363448679447, + 0.01014027651399374, + 0.01590702123939991, + 0.018653107807040215, + 0.013019527308642864, + 0.014198492281138897, + 0.023297946900129318, + 0.02893701381981373, + 0.005730119068175554, + 0.006282083224505186, + 0.006932212505489588, + 0.0071654547937214375, + 0.00757810240611434, + 0.005362676456570625, + 0.005697184707969427, + 0.006111592520028353, + 0.005265583284199238, + 0.0053103589452803135, + 0.006450457498431206, + 0.007979687303304672, + 0.007271579001098871, + 0.00482269236817956, + 0.004925521090626717, + 0.004614304751157761, + 0.006655828561633825, + 0.006325867027044296, + 0.011220461688935757, + 0.011240986175835133, + 0.010581247508525848, + 0.005519284401088953, + 0.0074808355420827866, + 0.008846723474562168, + 0.007234959397464991, + 0.006867059972137213, + 0.008287479169666767, + 0.010024686343967915, + 0.015299629420042038, + 0.011765753850340843, + 0.007625817786902189, + 0.005777334328740835, + 0.006569329183548689, + 0.007586257066577673, + 0.01422898843884468, + 0.01765144057571888, + 0.00885742250829935, + 0.019511716440320015, + 0.010918454267084599, + 0.008116009645164013, + 0.0088035324588418, + 0.013407351449131966, + 0.027795366942882538, + 0.016485139727592468, + 0.014795508235692978, + 0.022058716043829918, + 0.009867895394563675, + 0.012108854949474335, + 0.020406581461429596, + 0.034139201045036316, + 0.05114039033651352, + 0.0468621551990509, + 0.031127270311117172, + 0.03888186812400818, + 0.028000978752970695, + 0.02585645206272602, + 0.009396090172231197, + 0.012716894969344139, + 0.02380811981856823, + 0.011296641081571579, + 0.011097792536020279, + 0.012284839525818825, + 0.008782651275396347, + 0.009482640773057938, + 0.0023822514340281487, + 0.004374761134386063, + 0.002926976652815938, + 0.0038621118292212486, + 0.003933428321033716, + 0.010664667934179306, + 0.003313259920105338, + 0.0031946829985827208 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.08904524147510529 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank rolling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's armor painted with a deep olive-green matte camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank's surface covered in splashing mud and streaks of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tank traverse a rain-soaked dirt road while churning deep, flooded ruts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011073797941207886 + ], + "opposite_logprob_tokens": [ + -4.511073589324951 + ], + "logprob_reward": -0.011073797941207886, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011073797941207886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09970589727163315, + 0.05876389145851135, + 0.08358995616436005, + 0.07047528773546219, + 0.14299346506595612, + 0.18220968544483185, + 0.0967664048075676, + 0.05166035145521164, + 0.17303845286369324, + 0.04079660773277283 + ], + "credit_map": [ + 0.009568553417921066, + 0.010923516005277634, + 0.014794639311730862, + 0.013529233634471893, + 0.009664446115493774, + 0.01106224488466978, + 0.014741743914783001, + 0.015421513468027115, + 0.006491960026323795, + 0.006251025479286909, + 0.007674331311136484, + 0.0077980803325772285, + 0.010094800032675266, + 0.0063879541121423244, + 0.007641254924237728, + 0.006424487102776766, + 0.008917367085814476, + 0.014736436307430267, + 0.008529230952262878, + 0.012982170097529888, + 0.012987187132239342, + 0.010512172244489193, + 0.007670949213206768, + 0.007254444528371096, + 0.0052704340778291225, + 0.005295892246067524, + 0.006119682919234037, + 0.015735596418380737, + 0.012754932977259159, + 0.007133079692721367, + 0.007920218631625175, + 0.010245449841022491, + 0.011477697640657425, + 0.012009602971374989, + 0.029148640111088753, + 0.020110290497541428, + 0.016662977635860443, + 0.024605298414826393, + 0.01586752012372017, + 0.013111425563693047, + 0.013582244515419006, + 0.012101520784199238, + 0.023585038259625435, + 0.03244933485984802, + 0.017506418749690056, + 0.033924516290426254, + 0.026896333321928978, + 0.022164268419146538, + 0.006805934477597475, + 0.009935121983289719, + 0.020265312865376472, + 0.01572917401790619, + 0.008668274618685246, + 0.019433842971920967, + 0.00804220512509346, + 0.007886548526585102, + 0.004768176935613155, + 0.005950570106506348, + 0.007633513770997524, + 0.007630871143192053, + 0.004892132710665464, + 0.011121190153062344, + 0.005196377635002136, + 0.004467515740543604, + 0.014272640459239483, + 0.01252326462417841, + 0.023960789665579796, + 0.017508745193481445, + 0.019990023225545883, + 0.055826738476753235, + 0.013386698439717293, + 0.015569556504487991, + 0.0035955498460680246, + 0.003861087141558528, + 0.003091687336564064, + 0.005755000282078981, + 0.0054304394870996475, + 0.009574460797011852, + 0.004945729393512011, + 0.004542653914541006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 4 + ], + "selected_reward": -0.011073797941207886 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_1/index_tensor([8214, 8214])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_1/index_tensor([8214, 8214])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7635b00978f193edd668be000955d420cb839c85 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_1/index_tensor([8214, 8214])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot from a slightly upward angle, the camera slowly and steadily moves forward across the crystal-clear azure waters. Several clear beams of light, radiating from the shimmering surface, cast dancing, shifting patterns of light across the water. At the center of the frame, a mermaid lies gracefully on her side on a vibrant, vibrant coral reef. Her cascading silver hair gently flows in the current, shimmering with light. She possesses delicate features and radiant skin. Her deep, ocean-like turquoise eyes are half-closed, her expression captivated and serene, as if completely absorbed in her own world. Her lower body is a massive tail covered in iridescent, turquoise scales, each scale shifting in color as the light refracts. Her enormous caudal fin is as thin as a cicada's wing, its edges tinged with a faint golden sheen. Her red lips parted, as if in silent song. Streams of tiny, shimmering bubbles slowly escaped from her lips, drifting upward and dissipating in a column of light. Around her, schools of orange clownfish and vibrant coral shrimp darted leisurely among the diverse staghorn and brain corals. The entire scene, bathed in the soft underwater light and shadows, evoked a tranquil, dreamy, and even slightly sacred atmosphere.", + "seed": 1872, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_1/index_8214_seed_1872_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0552823749370873, + "vlm_logprob": -0.0552823749370873, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.0552823749370873, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with crystal-clear azure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mermaid lying on her side on a vibrant coral reef at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mermaid have cascading silver hair, turquoise eyes, and a massive tail with iridescent turquoise scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there schools of orange clownfish and vibrant coral shrimp swimming among staghorn and brain corals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10241696238517761 + ], + "opposite_logprob_tokens": [ + -2.3524169921875 + ], + "logprob_reward": -0.10241696238517761, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10241696238517761, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10591466724872589, + 0.3043815493583679, + 0.02777794562280178, + 0.04558275640010834, + 0.028247160837054253, + 0.024073051288723946, + 0.0317562036216259, + 0.039549052715301514, + 0.3475511372089386, + 0.04516639932990074 + ], + "credit_map": [ + 0.010840095579624176, + 0.019799068570137024, + 0.010923809371888638, + 0.014303850941359997, + 0.013829542323946953, + 0.014042678289115429, + 0.011634021997451782, + 0.010541616939008236, + 0.028721215203404427, + 0.059853482991456985, + 0.02563302405178547, + 0.029884885996580124, + 0.03730609640479088, + 0.04799138382077217, + 0.03431970626115799, + 0.04067177698016167, + 0.004078637808561325, + 0.003161246655508876, + 0.0035878203343600035, + 0.00308794598095119, + 0.002938286866992712, + 0.0030443889554589987, + 0.0035018327180296183, + 0.004377787932753563, + 0.005199217703193426, + 0.005035795737057924, + 0.0033400217071175575, + 0.006465753074735403, + 0.005379785317927599, + 0.006001675501465797, + 0.007529704365879297, + 0.006630806252360344, + 0.003609008388593793, + 0.004435448441654444, + 0.002953639253973961, + 0.0040549119003117085, + 0.0036525139585137367, + 0.0033925347961485386, + 0.0025183933321386576, + 0.003630711929872632, + 0.00473000667989254, + 0.0030329874716699123, + 0.003084602765738964, + 0.0029682996682822704, + 0.002977884840220213, + 0.0027823192067444324, + 0.0020559828262776136, + 0.002440971089527011, + 0.004666854161769152, + 0.003987905569374561, + 0.004981758538633585, + 0.002312979893758893, + 0.0036513940431177616, + 0.0061113410629332066, + 0.002389950677752495, + 0.003654023865237832, + 0.005501011852174997, + 0.005303612444549799, + 0.0038498425856232643, + 0.0029246644116938114, + 0.005353147629648447, + 0.007602803874760866, + 0.0041882977820932865, + 0.004825678188353777, + 0.04163321107625961, + 0.04299261420965195, + 0.07457321137189865, + 0.01689876616001129, + 0.036876361817121506, + 0.08792900294065475, + 0.021953454241156578, + 0.02469453774392605, + 0.007224190514534712, + 0.0037681825924664736, + 0.005224182736128569, + 0.0032030073925852776, + 0.006592962425202131, + 0.012870511040091515, + 0.003437822451815009, + 0.00284554623067379 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.10241696238517761 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with crystal-clear azure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mermaid lying on her side on a vibrant coral reef at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mermaid have cascading silver hair, turquoise eyes, and a massive tail with iridescent turquoise scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there schools of orange clownfish and vibrant coral shrimp swimming among staghorn and brain corals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076359291560947895 + ], + "opposite_logprob_tokens": [ + -4.882636070251465 + ], + "logprob_reward": -0.0076359291560947895, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076359291560947895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09055435657501221, + 0.10084497183561325, + 0.058811552822589874, + 0.044108279049396515, + 0.03672774136066437, + 0.02058500610291958, + 0.06479570269584656, + 0.033299919217824936, + 0.18681597709655762, + 0.3634565472602844 + ], + "credit_map": [ + 0.008773108012974262, + 0.01671598106622696, + 0.00915945041924715, + 0.01286510843783617, + 0.014363323338329792, + 0.010551436804234982, + 0.00978444516658783, + 0.008341499604284763, + 0.009687989018857479, + 0.012429207563400269, + 0.016554702073335648, + 0.0114469388499856, + 0.015980839729309082, + 0.012710994109511375, + 0.008033900521695614, + 0.014000392518937588, + 0.006425539031624794, + 0.011960401199758053, + 0.005562199279665947, + 0.0047690607607364655, + 0.007054077461361885, + 0.010320217348635197, + 0.006108780391514301, + 0.006611272692680359, + 0.007127286866307259, + 0.0053868587128818035, + 0.004188220947980881, + 0.005141490139067173, + 0.004536260385066271, + 0.005468577146530151, + 0.006416274234652519, + 0.00584330502897501, + 0.005055440589785576, + 0.0077997115440666676, + 0.003165970090776682, + 0.0038084546104073524, + 0.00611840607598424, + 0.004223203286528587, + 0.002999688731506467, + 0.0035568629391491413, + 0.002883210312575102, + 0.0025445022620260715, + 0.0020783902145922184, + 0.0022529398556798697, + 0.003365212818607688, + 0.0025284599978476763, + 0.002197012770920992, + 0.0027352767065167427, + 0.009716170839965343, + 0.007533487398177385, + 0.011612718924880028, + 0.00511064101010561, + 0.005949606653302908, + 0.011728435754776001, + 0.005133114289492369, + 0.008011522702872753, + 0.0038168481551110744, + 0.004413079470396042, + 0.005130994599312544, + 0.0033270264975726604, + 0.005296492483466864, + 0.0042093065567314625, + 0.00393390329554677, + 0.00317226629704237, + 0.01618105173110962, + 0.016144581139087677, + 0.03746528550982475, + 0.01727389544248581, + 0.026905395090579987, + 0.03499603644013405, + 0.015341892838478088, + 0.022507837042212486, + 0.07732376456260681, + 0.04450894147157669, + 0.02419310435652733, + 0.01440794114023447, + 0.0366785041987896, + 0.13111436367034912, + 0.017303386703133583, + 0.017926499247550964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.0076359291560947895 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with crystal-clear azure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mermaid lying on her side on a vibrant coral reef at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mermaid have cascading silver hair, turquoise eyes, and a massive tail with iridescent turquoise scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there schools of orange clownfish and vibrant coral shrimp swimming among staghorn and brain corals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016062384471297264 + ], + "opposite_logprob_tokens": [ + -4.141062259674072 + ], + "logprob_reward": -0.016062384471297264, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016062384471297264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11443506926298141, + 0.15499432384967804, + 0.05716025084257126, + 0.04282157123088837, + 0.04660980403423309, + 0.0706295445561409, + 0.15064556896686554, + 0.12167289108037949, + 0.09977317601442337, + 0.1412578523159027 + ], + "credit_map": [ + 0.013000466860830784, + 0.01654311828315258, + 0.011742089875042439, + 0.015051900409162045, + 0.012666996568441391, + 0.01278528943657875, + 0.019376717507839203, + 0.013268493115901947, + 0.015584824606776237, + 0.032088786363601685, + 0.012443514540791512, + 0.012612704187631607, + 0.01762755960226059, + 0.025945419445633888, + 0.019479379057884216, + 0.019212141633033752, + 0.005424555391073227, + 0.00969708152115345, + 0.004562203772366047, + 0.006500999443233013, + 0.008828382007777691, + 0.009419205598533154, + 0.007479043677449226, + 0.005248775705695152, + 0.005322934128344059, + 0.004127561114728451, + 0.003812682582065463, + 0.0046851858496665955, + 0.004664276726543903, + 0.005461771506816149, + 0.007584173232316971, + 0.007162987254559994, + 0.00746051874011755, + 0.008871279656887054, + 0.0045163594186306, + 0.004991333466023207, + 0.006664067506790161, + 0.005730525124818087, + 0.0034744665026664734, + 0.004901253152638674, + 0.00925713125616312, + 0.012605118565261364, + 0.006912230048328638, + 0.008341390639543533, + 0.013901800848543644, + 0.007429884746670723, + 0.005684556905180216, + 0.00649743527173996, + 0.019625650718808174, + 0.019400520250201225, + 0.019658299162983894, + 0.011623980477452278, + 0.014296908862888813, + 0.029128454625606537, + 0.01862339861690998, + 0.018288351595401764, + 0.014962056651711464, + 0.015619687736034393, + 0.010336865670979023, + 0.009229481220245361, + 0.02131112478673458, + 0.025956423953175545, + 0.013198641128838062, + 0.011058619245886803, + 0.013696788810193539, + 0.012473548762500286, + 0.017603415995836258, + 0.006233315449208021, + 0.013269436545670033, + 0.017853446304798126, + 0.009791675955057144, + 0.008851536549627781, + 0.02930002100765705, + 0.019934168085455894, + 0.01524843368679285, + 0.007000261917710304, + 0.010223204270005226, + 0.046081703156232834, + 0.006841938942670822, + 0.00662811566144228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 9 + ], + "selected_reward": -0.016062384471297264 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with crystal-clear azure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mermaid lying on her side on a vibrant coral reef at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mermaid have cascading silver hair, turquoise eyes, and a massive tail with iridescent turquoise scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there schools of orange clownfish and vibrant coral shrimp swimming among staghorn and brain corals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12702611088752747 + ], + "opposite_logprob_tokens": [ + -2.127026081085205 + ], + "logprob_reward": -0.12702611088752747, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12702611088752747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1435638815164566, + 0.20825093984603882, + 0.12232031673192978, + 0.051447805017232895, + 0.026566367596387863, + 0.02635050192475319, + 0.04192519560456276, + 0.03463294729590416, + 0.2429574728012085, + 0.10198456048965454 + ], + "credit_map": [ + 0.013282779604196548, + 0.016857583075761795, + 0.016584934666752815, + 0.013853561133146286, + 0.02539662830531597, + 0.021216217428445816, + 0.01642073132097721, + 0.019951453432440758, + 0.020307747647166252, + 0.0428277924656868, + 0.020579272881150246, + 0.018085403367877007, + 0.031669020652770996, + 0.02691231481730938, + 0.022204428911209106, + 0.025664959102869034, + 0.01433201227337122, + 0.024550922214984894, + 0.009551686234772205, + 0.010891599580645561, + 0.012796246446669102, + 0.025653500109910965, + 0.01340324617922306, + 0.011141098104417324, + 0.00600736727938056, + 0.00952198263257742, + 0.0035978362429887056, + 0.007324393838644028, + 0.011193996295332909, + 0.0061025177128612995, + 0.003692132653668523, + 0.004007578361779451, + 0.00397528288885951, + 0.005377164110541344, + 0.0022859699092805386, + 0.0024637109600007534, + 0.005476815160363913, + 0.0023221576120704412, + 0.0018226666143164039, + 0.002842601854354143, + 0.003983458038419485, + 0.00398279121145606, + 0.0029326092917472124, + 0.00256248633377254, + 0.004101312253624201, + 0.0032475509215146303, + 0.0020232400856912136, + 0.0035170544870197773, + 0.006812920793890953, + 0.0057978881523013115, + 0.007711935322731733, + 0.004069484770298004, + 0.005132659338414669, + 0.004703948274254799, + 0.0032763772178441286, + 0.004419982898980379, + 0.004656960256397724, + 0.00394417392089963, + 0.005300870165228844, + 0.0026752310805022717, + 0.004490380175411701, + 0.006637335289269686, + 0.0031155990436673164, + 0.003812400158494711, + 0.02666311524808407, + 0.026095645502209663, + 0.057423390448093414, + 0.017990875989198685, + 0.026708275079727173, + 0.05124365910887718, + 0.0170837864279747, + 0.019748728722333908, + 0.02501838095486164, + 0.01269624475389719, + 0.0062429821118712425, + 0.003043363569304347, + 0.008980359882116318, + 0.03799706697463989, + 0.0034419139847159386, + 0.004564247559756041 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.12702611088752747 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with crystal-clear azure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mermaid lying on her side on a vibrant coral reef at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mermaid have cascading silver hair, turquoise eyes, and a massive tail with iridescent turquoise scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there schools of orange clownfish and vibrant coral shrimp swimming among staghorn and brain corals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023270487785339355 + ], + "opposite_logprob_tokens": [ + -3.773270606994629 + ], + "logprob_reward": -0.023270487785339355, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023270487785339355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14860877394676208, + 0.131461963057518, + 0.12113694846630096, + 0.040148451924324036, + 0.029994048178195953, + 0.04114111140370369, + 0.048237815499305725, + 0.044704683125019073, + 0.211827352643013, + 0.18273888528347015 + ], + "credit_map": [ + 0.012215555645525455, + 0.020031604915857315, + 0.012795384041965008, + 0.01955474726855755, + 0.02171843685209751, + 0.01845991425216198, + 0.03092300519347191, + 0.012910119257867336, + 0.01285501103848219, + 0.030864575877785683, + 0.009105232544243336, + 0.00985720381140709, + 0.01964142732322216, + 0.024224886670708656, + 0.012833178974688053, + 0.012080445885658264, + 0.008707170374691486, + 0.03188924118876457, + 0.006106134969741106, + 0.00820605456829071, + 0.011226555332541466, + 0.03288477659225464, + 0.012593748979270458, + 0.009523266926407814, + 0.004379126708954573, + 0.006145930849015713, + 0.003860312979668379, + 0.0053032380528748035, + 0.0070765577256679535, + 0.004664127249270678, + 0.004712795838713646, + 0.004006359726190567, + 0.004627464804798365, + 0.004495399538427591, + 0.0037112131249159575, + 0.002663690596818924, + 0.004194939509034157, + 0.003608578583225608, + 0.00293991481885314, + 0.0037528472021222115, + 0.005983097944408655, + 0.005838872864842415, + 0.004985658451914787, + 0.0058569032698869705, + 0.006020957138389349, + 0.005066479090601206, + 0.0033638281747698784, + 0.004025313537567854, + 0.006308028474450111, + 0.006615661550313234, + 0.005744625348597765, + 0.004163000732660294, + 0.00616804976016283, + 0.007018298842012882, + 0.004252426326274872, + 0.007967725396156311, + 0.005730677396059036, + 0.006840157322585583, + 0.006279513705521822, + 0.0032564960420131683, + 0.0046805571764707565, + 0.008199256844818592, + 0.004731988534331322, + 0.004986033774912357, + 0.028340620920062065, + 0.02304074354469776, + 0.027227504178881645, + 0.024911286309361458, + 0.03097105585038662, + 0.03633028268814087, + 0.014520268887281418, + 0.02648557908833027, + 0.03681238740682602, + 0.02761848457157612, + 0.020575040951371193, + 0.012714643962681293, + 0.018777405843138695, + 0.045950595289468765, + 0.010362200438976288, + 0.009928139857947826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.023270487785339355 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a slightly upward angle, the camera slowly and steadily moves forward across the crystal-clear azure waters. Several clear beams of light, radiating from the shimmering surface, cast dancing, shifting patterns of light across the water. At the center of the frame, a mermaid lies gracefully on her side on a vibrant, vibrant coral reef. Her cascading silver hair gently flows in the current, shimmering with light. She possesses delicate features and radiant skin. Her deep, ocean-like turquoise eyes are half-closed, her expression captivated and serene, as if completely absorbed in her own world. Her lower body is a massive tail covered in iridescent, turquoise scales, each scale shifting in color as the light refracts. Her enormous caudal fin is as thin as a cicada's wing, its edges tinged with a faint golden sheen. Her red lips parted, as if in silent song. Streams of tiny, shimmering bubbles slowly escaped from her lips, drifting upward and dissipating in a column of light. Around her, schools of orange clownfish and vibrant coral shrimp darted leisurely among the diverse staghorn and brain corals. The entire scene, bathed in the soft underwater light and shadows, evoked a tranquil, dreamy, and even slightly sacred atmosphere.", + "seed": 1873, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_1/index_8214_seed_1873_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.011863014753907919, + "vlm_logprob": -0.011863014753907919, + "vlm_yes_no_margin": 4.624999952316284, + "vlm_token_credit": -0.011863014753907919, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with crystal-clear azure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mermaid lying on her side on a vibrant coral reef at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mermaid have cascading silver hair, turquoise eyes, and a massive tail with iridescent turquoise scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there schools of orange clownfish and vibrant coral shrimp swimming among staghorn and brain corals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004953374620527029 + ], + "opposite_logprob_tokens": [ + -5.629953384399414 + ], + "logprob_reward": -0.004953374620527029, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.004953374620527029, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11660460382699966, + 0.10042857378721237, + 0.13017764687538147, + 0.034993961453437805, + 0.05416873097419739, + 0.15473408997058868, + 0.035320576280355453, + 0.1506357491016388, + 0.17083220183849335, + 0.052103906869888306 + ], + "credit_map": [ + 0.010511736385524273, + 0.009258309379220009, + 0.017481978982686996, + 0.012233424000442028, + 0.019127095118165016, + 0.01985408551990986, + 0.018644854426383972, + 0.00949312187731266, + 0.014600600115954876, + 0.009303171187639236, + 0.010177168995141983, + 0.009564475156366825, + 0.018444014713168144, + 0.00929249357432127, + 0.020354190841317177, + 0.00869246106594801, + 0.014950214885175228, + 0.006860793102532625, + 0.03455869108438492, + 0.005925762001425028, + 0.040518276393413544, + 0.011538952589035034, + 0.008216745220124722, + 0.007608209270983934, + 0.0028247206937521696, + 0.003247450804337859, + 0.0033824562560766935, + 0.0025852450635284185, + 0.007451602723449469, + 0.008349480107426643, + 0.004454729612916708, + 0.0026982747949659824, + 0.006582897622138262, + 0.004987749271094799, + 0.008264024741947651, + 0.004716984927654266, + 0.011583666317164898, + 0.006162457168102264, + 0.0058184158988296986, + 0.006052535492926836, + 0.009617846459150314, + 0.010730070061981678, + 0.015349967405200005, + 0.01432852353900671, + 0.028735261410474777, + 0.021005017682909966, + 0.03670663386583328, + 0.018260769546031952, + 0.0033533985260874033, + 0.002679120749235153, + 0.0027660003397613764, + 0.002114108530804515, + 0.006396138109266758, + 0.010859492234885693, + 0.0029750512912869453, + 0.004177268128842115, + 0.0067283716052770615, + 0.006716998293995857, + 0.007709130644798279, + 0.006053413730114698, + 0.0570356659591198, + 0.011467700824141502, + 0.007535512093454599, + 0.0473889596760273, + 0.007261483930051327, + 0.0074128624983131886, + 0.008179602213203907, + 0.011723751202225685, + 0.06478647887706757, + 0.013538538478314877, + 0.008107146248221397, + 0.04982234165072441, + 0.008347846567630768, + 0.004878041800111532, + 0.005152236204594374, + 0.004946423228830099, + 0.009792828932404518, + 0.0055979518219828606, + 0.004021253436803818, + 0.009367327205836773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.004953374620527029 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with crystal-clear azure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mermaid lying on her side on a vibrant coral reef at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mermaid have cascading silver hair, turquoise eyes, and a massive tail with iridescent turquoise scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there schools of orange clownfish and vibrant coral shrimp swimming among staghorn and brain corals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005949762649834156 + ], + "opposite_logprob_tokens": [ + -5.130949974060059 + ], + "logprob_reward": -0.005949762649834156, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005949762649834156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05518893897533417, + 0.06880255043506622, + 0.1355789750814438, + 0.08222945034503937, + 0.06572037190198898, + 0.04613545909523964, + 0.035064928233623505, + 0.036679707467556, + 0.3914099931716919, + 0.08318961411714554 + ], + "credit_map": [ + 0.009725878946483135, + 0.0054573314264416695, + 0.007312606554478407, + 0.004520415794104338, + 0.008440925739705563, + 0.0071624089032411575, + 0.0077219377271831036, + 0.004847429692745209, + 0.0148870088160038, + 0.008950483053922653, + 0.005734999198466539, + 0.0061755916103720665, + 0.016890615224838257, + 0.004527953919023275, + 0.006732403300702572, + 0.004903496243059635, + 0.01061149500310421, + 0.005961626768112183, + 0.041044387966394424, + 0.006784546189010143, + 0.04437913000583649, + 0.010577021166682243, + 0.009023621678352356, + 0.007197136525064707, + 0.010363338515162468, + 0.00979527086019516, + 0.010196718387305737, + 0.009905721992254257, + 0.016820071265101433, + 0.01050154771655798, + 0.00856025330722332, + 0.006086525972932577, + 0.009899989701807499, + 0.00520756421610713, + 0.009527353569865227, + 0.007464360911399126, + 0.011431332677602768, + 0.00752502353861928, + 0.007624107412993908, + 0.007040639873594046, + 0.0056979176588356495, + 0.0054157571867108345, + 0.004249243531376123, + 0.005134759470820427, + 0.009495829232037067, + 0.005623675882816315, + 0.006797921843826771, + 0.0037203519605100155, + 0.004150243941694498, + 0.0038753298576921225, + 0.005166627001017332, + 0.004207388963550329, + 0.004585717339068651, + 0.006613885518163443, + 0.004009554162621498, + 0.0024561802856624126, + 0.004348928574472666, + 0.005414946936070919, + 0.004801700357347727, + 0.004960538353770971, + 0.005215799435973167, + 0.003392194863408804, + 0.004767838399857283, + 0.0037777612451463938, + 0.014552443288266659, + 0.013401873409748077, + 0.016628041863441467, + 0.023237433284521103, + 0.1599246859550476, + 0.029957905411720276, + 0.01577656902372837, + 0.1179310604929924, + 0.009069321677088737, + 0.006899780593812466, + 0.009483332745730877, + 0.0056830202229321, + 0.01729678548872471, + 0.009078414179384708, + 0.005073730833828449, + 0.020605234429240227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 9 + ], + "selected_reward": -0.005949762649834156 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with crystal-clear azure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mermaid lying on her side on a vibrant coral reef at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mermaid have cascading silver hair, turquoise eyes, and a massive tail with iridescent turquoise scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there schools of orange clownfish and vibrant coral shrimp swimming among staghorn and brain corals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014188130386173725 + ], + "opposite_logprob_tokens": [ + -4.264188289642334 + ], + "logprob_reward": -0.014188130386173725, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014188130386173725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1552790403366089, + 0.11275118589401245, + 0.16043421626091003, + 0.028831815347075462, + 0.08906466513872147, + 0.043280381709337234, + 0.08346309512853622, + 0.1548008918762207, + 0.09825873374938965, + 0.07383600622415543 + ], + "credit_map": [ + 0.03719694912433624, + 0.011050845496356487, + 0.008951127529144287, + 0.012616799212992191, + 0.04723108932375908, + 0.012493251822888851, + 0.015480952337384224, + 0.010258031077682972, + 0.022068679332733154, + 0.010336073115468025, + 0.007043137680739164, + 0.008382167667150497, + 0.03508417680859566, + 0.007720362860709429, + 0.012953820638358593, + 0.009162766858935356, + 0.010410462506115437, + 0.007434164639562368, + 0.043004997074604034, + 0.009388997219502926, + 0.05584613233804703, + 0.014695815742015839, + 0.010000068694353104, + 0.009653564542531967, + 0.003365573473274708, + 0.003425631672143936, + 0.0025024674832820892, + 0.002612243639305234, + 0.003712930716574192, + 0.004907169379293919, + 0.004626212641596794, + 0.0036795856431126595, + 0.012054438702762127, + 0.006279478315263987, + 0.013987572863698006, + 0.00859268568456173, + 0.013419823721051216, + 0.009482732973992825, + 0.013482947833836079, + 0.011764978058636189, + 0.005876451265066862, + 0.006715525407344103, + 0.004556772764772177, + 0.005128799006342888, + 0.007221879903227091, + 0.0042555322870612144, + 0.005199001636356115, + 0.004326418042182922, + 0.006890594959259033, + 0.005717509426176548, + 0.008438554592430592, + 0.004953391384333372, + 0.018896523863077164, + 0.024544714018702507, + 0.007572902366518974, + 0.006448907777667046, + 0.014944135211408138, + 0.015502703376114368, + 0.015295957215130329, + 0.019330445677042007, + 0.024935845285654068, + 0.014988958835601807, + 0.02717137150466442, + 0.022631462663412094, + 0.006320736836642027, + 0.00470695598050952, + 0.005318517331033945, + 0.006461738608777523, + 0.034970059990882874, + 0.00781199149787426, + 0.0045736306346952915, + 0.028095101937651634, + 0.013057801872491837, + 0.011671746149659157, + 0.009508966468274593, + 0.006843936163932085, + 0.01203226763755083, + 0.00519249401986599, + 0.0076705412939190865, + 0.007858249358832836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.014188130386173725 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with crystal-clear azure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mermaid lying on her side on a vibrant coral reef at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mermaid have cascading silver hair, turquoise eyes, and a massive tail with iridescent turquoise scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there schools of orange clownfish and vibrant coral shrimp swimming among staghorn and brain corals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016056636348366737 + ], + "opposite_logprob_tokens": [ + -4.141056537628174 + ], + "logprob_reward": -0.016056636348366737, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016056636348366737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27750447392463684, + 0.0768277570605278, + 0.11461035907268524, + 0.1004866436123848, + 0.1636980175971985, + 0.055510956794023514, + 0.028794342651963234, + 0.04171157628297806, + 0.0315348245203495, + 0.10932102054357529 + ], + "credit_map": [ + 0.08169329911470413, + 0.015415618196129799, + 0.01023392379283905, + 0.013746577315032482, + 0.1024627760052681, + 0.022595738992094994, + 0.02222628891468048, + 0.009130246937274933, + 0.01570114679634571, + 0.009515197947621346, + 0.007885425351560116, + 0.006889706943184137, + 0.015803448855876923, + 0.005610576365143061, + 0.00907092448323965, + 0.006351334508508444, + 0.010209404863417149, + 0.005480821710079908, + 0.023350069299340248, + 0.0054081943817436695, + 0.04253210872411728, + 0.011069882661104202, + 0.007788856513798237, + 0.008771030232310295, + 0.012578543275594711, + 0.012506701052188873, + 0.011104277335107327, + 0.010763011872768402, + 0.01559457927942276, + 0.016742713749408722, + 0.011445313692092896, + 0.009751513600349426, + 0.03336353227496147, + 0.018507273867726326, + 0.019397106021642685, + 0.0225163996219635, + 0.0243267510086298, + 0.014010127633810043, + 0.016585543751716614, + 0.014991283416748047, + 0.005756720434874296, + 0.005021155346184969, + 0.006474201567471027, + 0.005443023983389139, + 0.009599925018846989, + 0.0069561186246573925, + 0.010734625160694122, + 0.005525194574147463, + 0.004560993518680334, + 0.0036553647369146347, + 0.0030201382469385862, + 0.004000370856374502, + 0.0031775583047419786, + 0.0032999333925545216, + 0.004675463307648897, + 0.0024045207537710667, + 0.002329971408471465, + 0.00261922599747777, + 0.0026375751476734877, + 0.0020824952516704798, + 0.013813878409564495, + 0.002881707390770316, + 0.0035726267378777266, + 0.011774098500609398, + 0.002282131928950548, + 0.001884045428596437, + 0.002056736033409834, + 0.0022151335142552853, + 0.009940831921994686, + 0.0023104713764041662, + 0.0021278532221913338, + 0.008717622607946396, + 0.010987849906086922, + 0.006952253170311451, + 0.00998995453119278, + 0.007137700449675322, + 0.025709450244903564, + 0.014158090576529503, + 0.007415801286697388, + 0.02696993015706539 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.016056636348366737 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with crystal-clear azure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mermaid lying on her side on a vibrant coral reef at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mermaid have cascading silver hair, turquoise eyes, and a massive tail with iridescent turquoise scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there schools of orange clownfish and vibrant coral shrimp swimming among staghorn and brain corals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018167169764637947 + ], + "opposite_logprob_tokens": [ + -4.018167018890381 + ], + "logprob_reward": -0.018167169764637947, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018167169764637947, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08302035927772522, + 0.03722166270017624, + 0.054314445704221725, + 0.03449960798025131, + 0.0391387939453125, + 0.08758511394262314, + 0.11061470955610275, + 0.09202749282121658, + 0.42999112606048584, + 0.03158668801188469 + ], + "credit_map": [ + 0.026051055639982224, + 0.006093962118029594, + 0.006338066887110472, + 0.0030659250915050507, + 0.02755100280046463, + 0.005715995095670223, + 0.005706542171537876, + 0.002497816924005747, + 0.004078228026628494, + 0.003646084340289235, + 0.003177742240950465, + 0.003799568861722946, + 0.005212392657995224, + 0.004886467009782791, + 0.00920447800308466, + 0.0032167008612304926, + 0.007924914360046387, + 0.00524770375341177, + 0.011578579433262348, + 0.004475187975913286, + 0.006564943585544825, + 0.004657129757106304, + 0.006681591738015413, + 0.007184397894889116, + 0.004023998510092497, + 0.0033463244326412678, + 0.005351896397769451, + 0.0032632688526064157, + 0.004250373225659132, + 0.004631263203918934, + 0.0048701041378080845, + 0.0047623771242797375, + 0.004028565250337124, + 0.002789074555039406, + 0.006066548638045788, + 0.003142490517348051, + 0.004867205861955881, + 0.004522955976426601, + 0.007242563180625439, + 0.006479393225163221, + 0.0074060698971152306, + 0.007062980439513922, + 0.00883982703089714, + 0.008300275541841984, + 0.014599795453250408, + 0.011798305436968803, + 0.018678296357393265, + 0.010899564251303673, + 0.007457761559635401, + 0.007002672180533409, + 0.007959654554724693, + 0.0043517909944057465, + 0.02987314760684967, + 0.04083571583032608, + 0.006755676586180925, + 0.006378294434398413, + 0.006822173949331045, + 0.008033309131860733, + 0.007204113062471151, + 0.008188270963728428, + 0.022044328972697258, + 0.007926777005195618, + 0.008408648893237114, + 0.023399872705340385, + 0.016808783635497093, + 0.01422605849802494, + 0.016111792996525764, + 0.02589428424835205, + 0.1708763688802719, + 0.031375836580991745, + 0.01696479506790638, + 0.13773316144943237, + 0.003945967648178339, + 0.0036255198065191507, + 0.0031680529937148094, + 0.0032994200009852648, + 0.006122034043073654, + 0.003751220414415002, + 0.002852864796295762, + 0.004821608774363995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.018167169764637947 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a slightly upward angle, the camera slowly and steadily moves forward across the crystal-clear azure waters. Several clear beams of light, radiating from the shimmering surface, cast dancing, shifting patterns of light across the water. At the center of the frame, a mermaid lies gracefully on her side on a vibrant, vibrant coral reef. Her cascading silver hair gently flows in the current, shimmering with light. She possesses delicate features and radiant skin. Her deep, ocean-like turquoise eyes are half-closed, her expression captivated and serene, as if completely absorbed in her own world. Her lower body is a massive tail covered in iridescent, turquoise scales, each scale shifting in color as the light refracts. Her enormous caudal fin is as thin as a cicada's wing, its edges tinged with a faint golden sheen. Her red lips parted, as if in silent song. Streams of tiny, shimmering bubbles slowly escaped from her lips, drifting upward and dissipating in a column of light. Around her, schools of orange clownfish and vibrant coral shrimp darted leisurely among the diverse staghorn and brain corals. The entire scene, bathed in the soft underwater light and shadows, evoked a tranquil, dreamy, and even slightly sacred atmosphere.", + "seed": 1874, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_1/index_8214_seed_1874_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02062043212354183, + "vlm_logprob": -0.02062043212354183, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.02062043212354183, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with crystal-clear azure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mermaid lying on her side on a vibrant coral reef at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mermaid have cascading silver hair, turquoise eyes, and a massive tail with iridescent turquoise scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there schools of orange clownfish and vibrant coral shrimp swimming among staghorn and brain corals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03542123734951019 + ], + "opposite_logprob_tokens": [ + -3.410421133041382 + ], + "logprob_reward": -0.03542123734951019, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03542123734951019, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0631541982293129, + 0.17317979037761688, + 0.15221381187438965, + 0.11206046491861343, + 0.07798384130001068, + 0.04215405136346817, + 0.00981068704277277, + 0.05321284011006355, + 0.027758345007896423, + 0.2884719669818878 + ], + "credit_map": [ + 0.014978771097958088, + 0.005277944728732109, + 0.009404580108821392, + 0.0038105722051113844, + 0.004222966264933348, + 0.014758105389773846, + 0.005015767179429531, + 0.0056854854337871075, + 0.06913913041353226, + 0.005923538934439421, + 0.008498146198689938, + 0.004184192977845669, + 0.005111183039844036, + 0.06778071075677872, + 0.003495974699035287, + 0.009046902880072594, + 0.02164078690111637, + 0.031030496582388878, + 0.012524382211267948, + 0.012838575057685375, + 0.02674069069325924, + 0.020900709554553032, + 0.013491852208971977, + 0.013046308420598507, + 0.018197976052761078, + 0.013832819648087025, + 0.009017464704811573, + 0.007001874502748251, + 0.02347586490213871, + 0.02268804982304573, + 0.00907844863831997, + 0.008767953142523766, + 0.01154082641005516, + 0.01008028257638216, + 0.008743366226553917, + 0.006485106889158487, + 0.011323239654302597, + 0.015145637094974518, + 0.006518447771668434, + 0.008146928623318672, + 0.006851571146398783, + 0.005199091043323278, + 0.0051506394520401955, + 0.004013109486550093, + 0.005634946282953024, + 0.005793052259832621, + 0.00319476006552577, + 0.006316877901554108, + 0.001569082261994481, + 0.0011672205291688442, + 0.0008061923435889184, + 0.0009765570284798741, + 0.0019020314794033766, + 0.0016155701596289873, + 0.0009561558836139739, + 0.0008178766001947224, + 0.008128555491566658, + 0.007538566831499338, + 0.005172498989850283, + 0.00753184175118804, + 0.005095968022942543, + 0.006753277499228716, + 0.006937829777598381, + 0.006054294295608997, + 0.004362588748335838, + 0.003034142777323723, + 0.0033999953884631395, + 0.0019703523721545935, + 0.003317788476124406, + 0.0033983970060944557, + 0.0037976661697030067, + 0.004477411508560181, + 0.09640337526798248, + 0.012946804985404015, + 0.011155436746776104, + 0.0065848687663674355, + 0.021457470953464508, + 0.11911159008741379, + 0.00788092240691185, + 0.012931476347148418 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 2 + ], + "selected_reward": -0.03542123734951019 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with crystal-clear azure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mermaid lying on her side on a vibrant coral reef at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mermaid have cascading silver hair, turquoise eyes, and a massive tail with iridescent turquoise scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there schools of orange clownfish and vibrant coral shrimp swimming among staghorn and brain corals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864844024181366 + ], + "opposite_logprob_tokens": [ + -4.75864839553833 + ], + "logprob_reward": -0.00864844024181366, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864844024181366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09018703550100327, + 0.5485283136367798, + 0.07096488773822784, + 0.05298271030187607, + 0.046525705605745316, + 0.029285674914717674, + 0.025239868089556694, + 0.026977015659213066, + 0.020510233938694, + 0.08879860490560532 + ], + "credit_map": [ + 0.020473351702094078, + 0.007786285597831011, + 0.012374352663755417, + 0.005936013534665108, + 0.006283979397267103, + 0.021677035838365555, + 0.006494004279375076, + 0.009162016212940216, + 0.22930829226970673, + 0.02015622891485691, + 0.014818704687058926, + 0.009335965849459171, + 0.013310879468917847, + 0.2204759269952774, + 0.012240397743880749, + 0.028881913051009178, + 0.010269866324961185, + 0.01508876122534275, + 0.008614128455519676, + 0.007494070567190647, + 0.008890008553862572, + 0.008071382530033588, + 0.005306222941726446, + 0.007230452727526426, + 0.0075323618948459625, + 0.007081764750182629, + 0.00565233314409852, + 0.0030756776686757803, + 0.010633527301251888, + 0.007812769152224064, + 0.0032249942887574434, + 0.007969287224113941, + 0.006607450544834137, + 0.008640378713607788, + 0.004631572403013706, + 0.007078084163367748, + 0.00412054592743516, + 0.0047820741310715675, + 0.006222521420568228, + 0.004443084355443716, + 0.004350539296865463, + 0.0040002865716814995, + 0.003122099209576845, + 0.004071009811013937, + 0.0031934913713485003, + 0.00433064391836524, + 0.0028870829846709967, + 0.003330521984025836, + 0.003551417961716652, + 0.003006342798471451, + 0.0029786911327391863, + 0.00411367230117321, + 0.002428591251373291, + 0.003365668933838606, + 0.0023064501583576202, + 0.0034890349488705397, + 0.004562194459140301, + 0.005309893283993006, + 0.002194745233282447, + 0.00340269529260695, + 0.002547170501202345, + 0.00297702569514513, + 0.0030248211696743965, + 0.0029584746807813644, + 0.0035088639706373215, + 0.002343427622690797, + 0.0025970982387661934, + 0.0017693836707621813, + 0.0016655795043334365, + 0.002510334365069866, + 0.0022904241923242807, + 0.003825124353170395, + 0.022814447060227394, + 0.011413764208555222, + 0.004280432127416134, + 0.006063336040824652, + 0.00856267660856247, + 0.025267397984862328, + 0.006188509054481983, + 0.004208053927868605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.00864844024181366 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with crystal-clear azure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mermaid lying on her side on a vibrant coral reef at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mermaid have cascading silver hair, turquoise eyes, and a massive tail with iridescent turquoise scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there schools of orange clownfish and vibrant coral shrimp swimming among staghorn and brain corals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072147637605667 + ], + "opposite_logprob_tokens": [ + -4.511072158813477 + ], + "logprob_reward": -0.011072147637605667, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072147637605667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06650535017251968, + 0.22059860825538635, + 0.15007443726062775, + 0.07165347784757614, + 0.08478173613548279, + 0.05868900939822197, + 0.03281824290752411, + 0.05521279573440552, + 0.06561178714036942, + 0.19405454397201538 + ], + "credit_map": [ + 0.015094191767275333, + 0.005887324921786785, + 0.008220857940614223, + 0.0041794804856181145, + 0.004931593779474497, + 0.01677507348358631, + 0.005803081206977367, + 0.005613751709461212, + 0.08898705989122391, + 0.00975070334970951, + 0.007520835381001234, + 0.0046827117912471294, + 0.006675274111330509, + 0.08671415597200394, + 0.004368636291474104, + 0.011899236589670181, + 0.020267680287361145, + 0.03642541170120239, + 0.012074178084731102, + 0.013870741240680218, + 0.025404350832104683, + 0.016279995441436768, + 0.01290185283869505, + 0.012850229628384113, + 0.01260900404304266, + 0.008324172347784042, + 0.004756020847707987, + 0.005915776360780001, + 0.014610376209020615, + 0.01267348974943161, + 0.007116429507732391, + 0.0056482101790606976, + 0.013721122406423092, + 0.010314716026186943, + 0.007908374071121216, + 0.011861893348395824, + 0.011676639318466187, + 0.01391972042620182, + 0.008618445135653019, + 0.006760823540389538, + 0.008704489096999168, + 0.008860458619892597, + 0.004088294692337513, + 0.008751979097723961, + 0.009260992519557476, + 0.0070090945810079575, + 0.006364400964230299, + 0.0056493026204407215, + 0.00511590763926506, + 0.00432680593803525, + 0.003317925613373518, + 0.002816276391968131, + 0.006783731747418642, + 0.0037580898497253656, + 0.003842927049845457, + 0.002856577979400754, + 0.00655542965978384, + 0.010373392142355442, + 0.005015142261981964, + 0.006152496673166752, + 0.009357750415802002, + 0.006754003930836916, + 0.00642950925976038, + 0.00457506999373436, + 0.010685982182621956, + 0.009904790669679642, + 0.006895992439240217, + 0.006578973028808832, + 0.005579023156315088, + 0.006839071866124868, + 0.008570719510316849, + 0.010557231493294239, + 0.027388624846935272, + 0.031143154948949814, + 0.013264693319797516, + 0.01674683578312397, + 0.038622453808784485, + 0.022722981870174408, + 0.027294358238577843, + 0.016871441155672073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 2 + ], + "selected_reward": -0.011072147637605667 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with crystal-clear azure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mermaid lying on her side on a vibrant coral reef at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mermaid have cascading silver hair, turquoise eyes, and a massive tail with iridescent turquoise scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there schools of orange clownfish and vibrant coral shrimp swimming among staghorn and brain corals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978871762752533 + ], + "opposite_logprob_tokens": [ + -3.5297887325286865 + ], + "logprob_reward": -0.02978871762752533, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978871762752533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10835704207420349, + 0.24772807955741882, + 0.0695166289806366, + 0.05361181125044823, + 0.07620852440595627, + 0.0628545880317688, + 0.02833208255469799, + 0.12559695541858673, + 0.026824669912457466, + 0.2009696215391159 + ], + "credit_map": [ + 0.026881815865635872, + 0.007919390685856342, + 0.013797417283058167, + 0.006798822898417711, + 0.008156951516866684, + 0.029557963833212852, + 0.008090762421488762, + 0.007153920829296112, + 0.1035156399011612, + 0.00998018030077219, + 0.006998325232416391, + 0.004251845646649599, + 0.0063501750119030476, + 0.09791126102209091, + 0.004056714009493589, + 0.014663949608802795, + 0.010274948552250862, + 0.009970457293093204, + 0.007357660681009293, + 0.005020321812480688, + 0.012463053688406944, + 0.010198027826845646, + 0.005188316106796265, + 0.009043856523931026, + 0.008623985573649406, + 0.005854891613125801, + 0.004896285943686962, + 0.004046991001814604, + 0.011150501668453217, + 0.009624630212783813, + 0.005209319293498993, + 0.004205209668725729, + 0.011814836412668228, + 0.01051813829690218, + 0.008502064272761345, + 0.007082885131239891, + 0.009285778738558292, + 0.013868623413145542, + 0.006586928386241198, + 0.008549274876713753, + 0.009194638580083847, + 0.0071233296766877174, + 0.006653104443103075, + 0.0055003114975988865, + 0.01174550037831068, + 0.009525795467197895, + 0.006311855278909206, + 0.006800058297812939, + 0.003948055673390627, + 0.0030908845365047455, + 0.004151328932493925, + 0.0026706906501203775, + 0.0031976185273379087, + 0.004449485335499048, + 0.003232622053474188, + 0.003591400571167469, + 0.0185441542416811, + 0.01465161144733429, + 0.00876696314662695, + 0.007092295680195093, + 0.03596378490328789, + 0.018216870725154877, + 0.011928282678127289, + 0.01043300423771143, + 0.005039426032453775, + 0.0025927904061973095, + 0.003176298923790455, + 0.0018577601294964552, + 0.003216503420844674, + 0.003151854733005166, + 0.004064907785505056, + 0.003725132904946804, + 0.06656694412231445, + 0.011108157224953175, + 0.007903339341282845, + 0.00552354846149683, + 0.013887854292988777, + 0.08089394867420197, + 0.006236831657588482, + 0.008849014528095722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 7 + ], + "selected_reward": -0.02978871762752533 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with crystal-clear azure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mermaid lying on her side on a vibrant coral reef at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mermaid have cascading silver hair, turquoise eyes, and a massive tail with iridescent turquoise scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there schools of orange clownfish and vibrant coral shrimp swimming among staghorn and brain corals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01817161776125431 + ], + "opposite_logprob_tokens": [ + -4.018171787261963 + ], + "logprob_reward": -0.01817161776125431, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01817161776125431, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09242178499698639, + 0.1988440603017807, + 0.13240359723567963, + 0.045730218291282654, + 0.048766858875751495, + 0.02292407490313053, + 0.014138022437691689, + 0.028689393773674965, + 0.04786362498998642, + 0.3682183027267456 + ], + "credit_map": [ + 0.022839132696390152, + 0.01001298613846302, + 0.007895040325820446, + 0.005855841096490622, + 0.006552542559802532, + 0.023414472118020058, + 0.007268926128745079, + 0.00858285091817379, + 0.08271023631095886, + 0.008757106959819794, + 0.005499835126101971, + 0.003402742324396968, + 0.005227313376963139, + 0.07837870717048645, + 0.004100315738469362, + 0.010767818428575993, + 0.018456919118762016, + 0.03394020348787308, + 0.0108176963403821, + 0.013515901751816273, + 0.01972142979502678, + 0.013855907134711742, + 0.010630789212882519, + 0.011464754119515419, + 0.007182089611887932, + 0.007490837015211582, + 0.003330904059112072, + 0.004841395188122988, + 0.006354300305247307, + 0.006847207900136709, + 0.005715899635106325, + 0.00396758783608675, + 0.005653879605233669, + 0.009290146641433239, + 0.005255820229649544, + 0.005645372439175844, + 0.004758292809128761, + 0.008056918159127235, + 0.004740995820611715, + 0.00536543782800436, + 0.0038856754545122385, + 0.0033973667304962873, + 0.002005190821364522, + 0.002880035899579525, + 0.0027302675880491734, + 0.0024133757688105106, + 0.0022494979202747345, + 0.003362665418535471, + 0.002429882762953639, + 0.0020885418634861708, + 0.0010802929755300283, + 0.0010277180699631572, + 0.0031446190550923347, + 0.0019680194091051817, + 0.001235964591614902, + 0.0011629837099462748, + 0.004103188402950764, + 0.003382965689525008, + 0.002571455202996731, + 0.002404272323474288, + 0.0067631336860358715, + 0.0034029141534119844, + 0.00258051510900259, + 0.003480951301753521, + 0.008240771479904652, + 0.006225768476724625, + 0.005916516296565533, + 0.0030988138169050217, + 0.006475675385445356, + 0.005306490231305361, + 0.005942807998508215, + 0.006656781770288944, + 0.12283048033714294, + 0.016158679500222206, + 0.018155930563807487, + 0.012233050540089607, + 0.026744918897747993, + 0.14311710000038147, + 0.011383676901459694, + 0.01759445108473301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 2 + ], + "selected_reward": -0.01817161776125431 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a slightly upward angle, the camera slowly and steadily moves forward across the crystal-clear azure waters. Several clear beams of light, radiating from the shimmering surface, cast dancing, shifting patterns of light across the water. At the center of the frame, a mermaid lies gracefully on her side on a vibrant, vibrant coral reef. Her cascading silver hair gently flows in the current, shimmering with light. She possesses delicate features and radiant skin. Her deep, ocean-like turquoise eyes are half-closed, her expression captivated and serene, as if completely absorbed in her own world. Her lower body is a massive tail covered in iridescent, turquoise scales, each scale shifting in color as the light refracts. Her enormous caudal fin is as thin as a cicada's wing, its edges tinged with a faint golden sheen. Her red lips parted, as if in silent song. Streams of tiny, shimmering bubbles slowly escaped from her lips, drifting upward and dissipating in a column of light. Around her, schools of orange clownfish and vibrant coral shrimp darted leisurely among the diverse staghorn and brain corals. The entire scene, bathed in the soft underwater light and shadows, evoked a tranquil, dreamy, and even slightly sacred atmosphere.", + "seed": 1875, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_1/index_8214_seed_1875_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.012700105272233485, + "vlm_logprob": -0.012700105272233485, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.012700105272233485, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with crystal-clear azure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mermaid lying on her side on a vibrant coral reef at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mermaid have cascading silver hair, turquoise eyes, and a massive tail with iridescent turquoise scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there schools of orange clownfish and vibrant coral shrimp swimming among staghorn and brain corals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013020847924053669 + ], + "opposite_logprob_tokens": [ + -4.513020992279053 + ], + "logprob_reward": -0.013020847924053669, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.013020847924053669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5299471616744995, + 0.03749880939722061, + 0.06345284730195999, + 0.1936454474925995, + 0.04707931727170944, + 0.01687975972890854, + 0.02373558282852173, + 0.02763432078063488, + 0.02966717816889286, + 0.03045956790447235 + ], + "credit_map": [ + 0.019059721380472183, + 0.009375058114528656, + 0.017144570127129555, + 0.012867609038949013, + 0.015434995293617249, + 0.20810925960540771, + 0.018845049664378166, + 0.2291109412908554, + 0.0028557642363011837, + 0.0028995403554290533, + 0.004214792512357235, + 0.007283959072083235, + 0.0035718607250601053, + 0.006312450859695673, + 0.003854809794574976, + 0.006505632773041725, + 0.0071093132719397545, + 0.006886887364089489, + 0.007811141200363636, + 0.007077912800014019, + 0.007855969481170177, + 0.009769047610461712, + 0.007571585476398468, + 0.009370990097522736, + 0.023816412314772606, + 0.014467881992459297, + 0.03233613818883896, + 0.01622919552028179, + 0.03885739669203758, + 0.02930646575987339, + 0.020891007035970688, + 0.017740944400429726, + 0.004911290947347879, + 0.004816531203687191, + 0.007397484499961138, + 0.0038682441227138042, + 0.010671763680875301, + 0.008726263418793678, + 0.0026224907487630844, + 0.00406524958088994, + 0.0024240422062575817, + 0.001210322487168014, + 0.003073933767154813, + 0.0012244089739397168, + 0.0029072891920804977, + 0.002231249585747719, + 0.0019860598258674145, + 0.0018224528757855296, + 0.0021784822456538677, + 0.001896655303426087, + 0.0034669768065214157, + 0.0015861380379647017, + 0.003224361455067992, + 0.005617485847324133, + 0.003633958986029029, + 0.002131523797288537, + 0.0025917883031070232, + 0.00238632271066308, + 0.0029313608538359404, + 0.0017762487987056375, + 0.006625039037317038, + 0.00518249673768878, + 0.0037126706447452307, + 0.0024283919483423233, + 0.00275815324857831, + 0.001484112348407507, + 0.0020243178587406874, + 0.0016563823446631432, + 0.008573885075747967, + 0.004816250875592232, + 0.004828155972063541, + 0.003525921842083335, + 0.0036196529399603605, + 0.002206405159085989, + 0.0048950607888400555, + 0.002666276879608631, + 0.004443324636667967, + 0.00480162538588047, + 0.0034171324223279953, + 0.004410089924931526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.013020847924053669 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with crystal-clear azure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mermaid lying on her side on a vibrant coral reef at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mermaid have cascading silver hair, turquoise eyes, and a massive tail with iridescent turquoise scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there schools of orange clownfish and vibrant coral shrimp swimming among staghorn and brain corals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036187181249260902 + ], + "opposite_logprob_tokens": [ + -5.6286187171936035 + ], + "logprob_reward": -0.0036187181249260902, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036187181249260902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3778630793094635, + 0.053451281040906906, + 0.10808169841766357, + 0.1737249195575714, + 0.04361413046717644, + 0.04691389575600624, + 0.05873415619134903, + 0.011247620917856693, + 0.040785204619169235, + 0.08558400720357895 + ], + "credit_map": [ + 0.008412127383053303, + 0.006548388861119747, + 0.007850718684494495, + 0.008135904558002949, + 0.009547090157866478, + 0.15354801714420319, + 0.015279230661690235, + 0.16854159533977509, + 0.010029883123934269, + 0.007502926979213953, + 0.008439050987362862, + 0.0053438120521605015, + 0.006700323894619942, + 0.004635718651115894, + 0.00397152453660965, + 0.006828040815889835, + 0.014278152957558632, + 0.011856799945235252, + 0.01121519599109888, + 0.014652380719780922, + 0.0181675273925066, + 0.010741975158452988, + 0.015132774598896503, + 0.012036886066198349, + 0.02470938302576542, + 0.015261347405612469, + 0.026913437992334366, + 0.014636687934398651, + 0.03506847098469734, + 0.023625340312719345, + 0.01793036051094532, + 0.015579880215227604, + 0.004813058767467737, + 0.004063645843416452, + 0.007521430961787701, + 0.004598261788487434, + 0.007330172695219517, + 0.006997722201049328, + 0.003238269593566656, + 0.00505156721919775, + 0.006135717500001192, + 0.0034982627257704735, + 0.010217870585620403, + 0.003817146411165595, + 0.007836291566491127, + 0.005998768378049135, + 0.004338833037763834, + 0.005071001127362251, + 0.005580205004662275, + 0.004463594872504473, + 0.011220155283808708, + 0.004474916961044073, + 0.007663309574127197, + 0.013860832899808884, + 0.006903373170644045, + 0.004567769356071949, + 0.0012864028103649616, + 0.0008796000038273633, + 0.0012956359423696995, + 0.001042724004946649, + 0.0020568796899169683, + 0.0018862781580537558, + 0.0015988112427294254, + 0.0012012887746095657, + 0.004244852811098099, + 0.004639258608222008, + 0.006426718086004257, + 0.0032517418731004, + 0.007402495946735144, + 0.004545059520751238, + 0.005867317318916321, + 0.004407760687172413, + 0.007008178625255823, + 0.005848762113600969, + 0.009288270026445389, + 0.0074764154851436615, + 0.01411451492458582, + 0.012972690165042877, + 0.01882576383650303, + 0.010049412027001381 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.0036187181249260902 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with crystal-clear azure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mermaid lying on her side on a vibrant coral reef at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mermaid have cascading silver hair, turquoise eyes, and a massive tail with iridescent turquoise scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there schools of orange clownfish and vibrant coral shrimp swimming among staghorn and brain corals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012531821615993977 + ], + "opposite_logprob_tokens": [ + -4.387531757354736 + ], + "logprob_reward": -0.012531821615993977, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012531821615993977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4375980496406555, + 0.07315332442522049, + 0.07798042893409729, + 0.18700815737247467, + 0.07163014262914658, + 0.023194119334220886, + 0.027451174333691597, + 0.019518041983246803, + 0.048190850764513016, + 0.03427576646208763 + ], + "credit_map": [ + 0.010583944618701935, + 0.006406068801879883, + 0.008730375207960606, + 0.00849909707903862, + 0.011231694370508194, + 0.17861542105674744, + 0.016900615766644478, + 0.1966308206319809, + 0.01329326443374157, + 0.010241981595754623, + 0.010705157183110714, + 0.008586660027503967, + 0.010414186865091324, + 0.006023680325597525, + 0.006880536209791899, + 0.007007855921983719, + 0.010857337154448032, + 0.008297606371343136, + 0.01013485249131918, + 0.009062260389328003, + 0.012248491868376732, + 0.008202767930924892, + 0.010912613943219185, + 0.008264492265880108, + 0.029887605458498, + 0.019271323457360268, + 0.03132552653551102, + 0.016880307346582413, + 0.032119862735271454, + 0.021974707022309303, + 0.019328638911247253, + 0.01622016914188862, + 0.00736946240067482, + 0.0070649199187755585, + 0.01433822512626648, + 0.005419726017862558, + 0.012924833223223686, + 0.013788933865725994, + 0.004007353913038969, + 0.006716682109981775, + 0.002514335559681058, + 0.0019448744133114815, + 0.004945757333189249, + 0.0022907289676368237, + 0.0033764548134058714, + 0.003186657326295972, + 0.002423362573608756, + 0.002511946251615882, + 0.0022522227372974157, + 0.0022219771053642035, + 0.004025761969387531, + 0.001698606531135738, + 0.004495697095990181, + 0.005319079849869013, + 0.004981220234185457, + 0.0024566049687564373, + 0.0023379449266940355, + 0.0016963055822998285, + 0.0019099022028967738, + 0.0012866181787103415, + 0.0051754848100245, + 0.0029845640528947115, + 0.0019749500788748264, + 0.0021522692404687405, + 0.004629597533494234, + 0.004259842913597822, + 0.005059809889644384, + 0.004927785601466894, + 0.008516600355505943, + 0.005972293205559254, + 0.0067683677189052105, + 0.008056549355387688, + 0.0036213984712958336, + 0.0022814495023339987, + 0.004378414247184992, + 0.0031296522356569767, + 0.005175767466425896, + 0.006130985449999571, + 0.00563855841755867, + 0.003919536247849464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.012531821615993977 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with crystal-clear azure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mermaid lying on her side on a vibrant coral reef at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mermaid have cascading silver hair, turquoise eyes, and a massive tail with iridescent turquoise scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there schools of orange clownfish and vibrant coral shrimp swimming among staghorn and brain corals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011063659563660622 + ], + "opposite_logprob_tokens": [ + -4.511063575744629 + ], + "logprob_reward": -0.011063659563660622, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011063659563660622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24389833211898804, + 0.09816954284906387, + 0.06962399184703827, + 0.12046170234680176, + 0.09837565571069717, + 0.09247562289237976, + 0.022084183990955353, + 0.03538288548588753, + 0.15948812663555145, + 0.0600399374961853 + ], + "credit_map": [ + 0.04063607007265091, + 0.022217288613319397, + 0.03664731606841087, + 0.044454991817474365, + 0.023952819406986237, + 0.02338968589901924, + 0.024775858968496323, + 0.027824293822050095, + 0.011742406524717808, + 0.008995119482278824, + 0.018510593101382256, + 0.010070664808154106, + 0.015434710308909416, + 0.011262086220085621, + 0.00987349171191454, + 0.012280458584427834, + 0.012109887786209583, + 0.008114434778690338, + 0.006949027068912983, + 0.007755028083920479, + 0.01131124421954155, + 0.00666268402710557, + 0.009383869357407093, + 0.007337816059589386, + 0.016592275351285934, + 0.012070036493241787, + 0.018095381557941437, + 0.010083052329719067, + 0.023166336119174957, + 0.016462184488773346, + 0.01204058900475502, + 0.011951838620007038, + 0.012288105674088001, + 0.008678057231009007, + 0.01974952593445778, + 0.009401385672390461, + 0.015057167038321495, + 0.014215266332030296, + 0.006755550391972065, + 0.01223058719187975, + 0.012584012933075428, + 0.0068786838091909885, + 0.018695535138249397, + 0.0078062815591692924, + 0.01570352166891098, + 0.012405715882778168, + 0.008815052919089794, + 0.009586801752448082, + 0.002266600728034973, + 0.0019355807453393936, + 0.004189108964055777, + 0.002032369840890169, + 0.0032124104909598827, + 0.0035699820145964622, + 0.0022879026364535093, + 0.002590226475149393, + 0.004629873670637608, + 0.002677205018699169, + 0.003531151218339801, + 0.0022204003762453794, + 0.010734324343502522, + 0.00544791528955102, + 0.0031762965954840183, + 0.002965714316815138, + 0.015850551426410675, + 0.012348674237728119, + 0.02164321579039097, + 0.016180701553821564, + 0.02828618884086609, + 0.02133154310286045, + 0.021388491615653038, + 0.022458748891949654, + 0.006835313979536295, + 0.004761037416756153, + 0.00608019158244133, + 0.006889045238494873, + 0.007958348840475082, + 0.010259663686156273, + 0.010343359783291817, + 0.006912972312420607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.011063659563660622 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with crystal-clear azure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mermaid lying on her side on a vibrant coral reef at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mermaid have cascading silver hair, turquoise eyes, and a massive tail with iridescent turquoise scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there schools of orange clownfish and vibrant coral shrimp swimming among staghorn and brain corals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023265479132533073 + ], + "opposite_logprob_tokens": [ + -3.7732653617858887 + ], + "logprob_reward": -0.023265479132533073, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023265479132533073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3561525344848633, + 0.09041302651166916, + 0.05552988499403, + 0.12452509254217148, + 0.07192645221948624, + 0.056029293686151505, + 0.06251035630702972, + 0.03954290971159935, + 0.07991502434015274, + 0.06345538049936295 + ], + "credit_map": [ + 0.01857544481754303, + 0.016118045896291733, + 0.028807537630200386, + 0.03245969116687775, + 0.020668823271989822, + 0.09976381063461304, + 0.025052549317479134, + 0.11470665037631989, + 0.014125015586614609, + 0.007620980963110924, + 0.010640439577400684, + 0.017336459830403328, + 0.009152643382549286, + 0.009179276414215565, + 0.012166501022875309, + 0.010191718116402626, + 0.009034317918121815, + 0.005842908285558224, + 0.006980575155466795, + 0.005547080654650927, + 0.009669255465269089, + 0.005363485310226679, + 0.006879794877022505, + 0.006212468259036541, + 0.012222224846482277, + 0.011113651096820831, + 0.021380633115768433, + 0.010604353621602058, + 0.022423187270760536, + 0.020768696442246437, + 0.012942381203174591, + 0.013069979846477509, + 0.0049365367740392685, + 0.0044493163004517555, + 0.01405302993953228, + 0.005872054025530815, + 0.013737387023866177, + 0.014571591280400753, + 0.005794820841401815, + 0.008511724881827831, + 0.008217276073992252, + 0.004241634625941515, + 0.01055489107966423, + 0.004871077835559845, + 0.008869502693414688, + 0.007795234676450491, + 0.004714221693575382, + 0.006765455473214388, + 0.0060509019531309605, + 0.005978979170322418, + 0.010112840682268143, + 0.006492783781141043, + 0.008025445975363255, + 0.013777438551187515, + 0.0069112591445446014, + 0.005160715896636248, + 0.005524680949747562, + 0.002970409346744418, + 0.004484403412789106, + 0.002110604429617524, + 0.00838707759976387, + 0.0073972903192043304, + 0.004953252617269754, + 0.0037151926662772894, + 0.007580001372843981, + 0.006921085063368082, + 0.013822045177221298, + 0.006525115109980106, + 0.012220021337270737, + 0.010526678524911404, + 0.011164266616106033, + 0.01115582138299942, + 0.007948569022119045, + 0.0052488562650978565, + 0.008814314380288124, + 0.0058680507354438305, + 0.009325142949819565, + 0.009802618063986301, + 0.008907780982553959, + 0.00754005229100585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.023265479132533073 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_10/index_tensor([2048, 2048])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_10/index_tensor([2048, 2048])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0114d50be70cdba93376bac9d5a3c43c00261d40 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_10/index_tensor([2048, 2048])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "textured, frame-by-frame animated shot, captured from a first-person perspective, looks down from a very low angle. A fluffy, predominantly gray and white cat paw takes center stage. This paw, focused yet clumsily, attempts to propel a vibrantly colored rubber dog toy ball—brightly striped with red and yellow—into a pile of fluffy cat litter, composed of fine white particles, resembling miniature sand dunes. Each movement is marked by a distinct pause. The paw first laboriously pushes the ball forward, then, with its pink pads and slightly open claws, persistently and persistently scrapes the litter from the side, attempting to cover the incongruous colored ball. Soft, warm indoor light streams from above and side, illuminating the edges of the cat's fur and casting subtle, yet distinct, shadows on the litter. The entire scene, shrouded in this clumsy effort, is imbued with an air of innocent, endearing cuteness.", + "seed": 1798, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_10/index_2048_seed_1798_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.038744383258745074, + "vlm_logprob": -0.038744383258745074, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.038744383258745074, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person perspective looking down from a very low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, predominantly gray and white cat paw visible taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrantly colored rubber dog toy ball, brightly striped with red and yellow, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pile of fluffy white cat litter resembling miniature sand dunes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paw push the ball forward and then scrape the litter from the side to cover it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06111457943916321 + ], + "opposite_logprob_tokens": [ + -3.186114549636841 + ], + "logprob_reward": -0.06111457943916321, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.06111457943916321, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24590963125228882, + 0.05289246886968613, + 0.04282930865883827, + 0.08047006279230118, + 0.061913590878248215, + 0.051644280552864075, + 0.2046787142753601, + 0.1422269642353058, + 0.07299628108739853, + 0.04443874582648277 + ], + "credit_map": [ + 0.04056287556886673, + 0.035631660372018814, + 0.044771257787942886, + 0.019792873412370682, + 0.014830036088824272, + 0.029515940696001053, + 0.023918671533465385, + 0.0368863008916378, + 0.0038151652552187443, + 0.01006064098328352, + 0.0051919808611273766, + 0.005966925993561745, + 0.003962809685617685, + 0.014569873921573162, + 0.004507462028414011, + 0.00481761246919632, + 0.0036986807826906443, + 0.004224351607263088, + 0.006841450463980436, + 0.005638463888317347, + 0.005649799946695566, + 0.007093280088156462, + 0.003986061085015535, + 0.005697217769920826, + 0.009674648754298687, + 0.017766427248716354, + 0.010907950811088085, + 0.005042389500886202, + 0.009999134577810764, + 0.009918765164911747, + 0.005729339085519314, + 0.011431405320763588, + 0.007625273894518614, + 0.01549670659005642, + 0.0050512016750872135, + 0.00391464913263917, + 0.008099292404949665, + 0.008390570059418678, + 0.005767452530562878, + 0.0075684417970478535, + 0.005498470738530159, + 0.010268124751746655, + 0.008153891190886497, + 0.004884774796664715, + 0.007247780915349722, + 0.005185224581509829, + 0.004604007117450237, + 0.00580200832337141, + 0.019918441772460938, + 0.041057854890823364, + 0.016454419121146202, + 0.034040551632642746, + 0.03415491804480553, + 0.019762037321925163, + 0.021157879382371902, + 0.018132619559764862, + 0.02035556733608246, + 0.01296958141028881, + 0.014247135259211063, + 0.03593600541353226, + 0.016570979729294777, + 0.01339622214436531, + 0.01653401181101799, + 0.012217454612255096, + 0.006542842835187912, + 0.01492715161293745, + 0.011101425625383854, + 0.009960891678929329, + 0.009461971931159496, + 0.006398929748684168, + 0.006715225987136364, + 0.007887840270996094, + 0.0048140729777514935, + 0.009561797603964806, + 0.005677833687514067, + 0.006346946116536856, + 0.005612495355308056, + 0.004221300128847361, + 0.003744505112990737, + 0.00445979181677103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.06111457943916321 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person perspective looking down from a very low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, predominantly gray and white cat paw visible taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrantly colored rubber dog toy ball, brightly striped with red and yellow, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pile of fluffy white cat litter resembling miniature sand dunes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paw push the ball forward and then scrape the litter from the side to cover it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012563489377498627 + ], + "opposite_logprob_tokens": [ + -4.387563705444336 + ], + "logprob_reward": -0.012563489377498627, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012563489377498627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12400557100772858, + 0.12115279585123062, + 0.1538514494895935, + 0.11329420655965805, + 0.06779477000236511, + 0.045355115085840225, + 0.07898729294538498, + 0.1706075221300125, + 0.07144918292760849, + 0.05350204184651375 + ], + "credit_map": [ + 0.013431607745587826, + 0.013083594851195812, + 0.013126282952725887, + 0.011819684877991676, + 0.008738788776099682, + 0.017124494537711143, + 0.022771112620830536, + 0.02391001582145691, + 0.011120175942778587, + 0.02434084564447403, + 0.02048506774008274, + 0.014318378642201424, + 0.010245241224765778, + 0.014565701596438885, + 0.014767508953809738, + 0.011309891939163208, + 0.010526823811233044, + 0.016130270436406136, + 0.019763469696044922, + 0.021258631721138954, + 0.0201556459069252, + 0.024176251143217087, + 0.018537400290369987, + 0.023302968591451645, + 0.011364422738552094, + 0.024273276329040527, + 0.011312388814985752, + 0.005836554802954197, + 0.012160315178334713, + 0.026264270767569542, + 0.006812713108956814, + 0.015270279720425606, + 0.007101166527718306, + 0.007729229051619768, + 0.005319108720868826, + 0.00696657644584775, + 0.011784086935222149, + 0.008491287007927895, + 0.006615053862333298, + 0.013788262382149696, + 0.004507109522819519, + 0.005724283400923014, + 0.009396933019161224, + 0.0027882985305041075, + 0.005346549674868584, + 0.005276753567159176, + 0.0035886967089027166, + 0.008726492524147034, + 0.005222269333899021, + 0.00819320697337389, + 0.008718151599168777, + 0.006892302073538303, + 0.021777456626296043, + 0.011897595599293709, + 0.006696776952594519, + 0.0095895417034626, + 0.03428931534290314, + 0.01734180562198162, + 0.018076272681355476, + 0.04564213752746582, + 0.017937839031219482, + 0.014574686996638775, + 0.010867428965866566, + 0.011878042481839657, + 0.007761831395328045, + 0.010548966936767101, + 0.008925279602408409, + 0.009201722219586372, + 0.012748859822750092, + 0.006315015256404877, + 0.007005389779806137, + 0.00894212257117033, + 0.006540077272802591, + 0.00893146637827158, + 0.009825211949646473, + 0.005133745260536671, + 0.005264929961413145, + 0.004506698343902826, + 0.004622355103492737, + 0.008677560836076736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.012563489377498627 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person perspective looking down from a very low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, predominantly gray and white cat paw visible taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrantly colored rubber dog toy ball, brightly striped with red and yellow, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pile of fluffy white cat litter resembling miniature sand dunes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paw push the ball forward and then scrape the litter from the side to cover it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024953915271908045 + ], + "opposite_logprob_tokens": [ + -6.002495288848877 + ], + "logprob_reward": -0.0024953915271908045, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024953915271908045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2209305763244629, + 0.05177508294582367, + 0.08240525424480438, + 0.08602878451347351, + 0.06580593436956406, + 0.109714575111866, + 0.24775001406669617, + 0.0657990574836731, + 0.033113088458776474, + 0.03667763993144035 + ], + "credit_map": [ + 0.026691462844610214, + 0.04778566583991051, + 0.04094862565398216, + 0.016994725912809372, + 0.022110041230916977, + 0.027508921921253204, + 0.016848089173436165, + 0.022043030709028244, + 0.0035414076410233974, + 0.008523263968527317, + 0.004294675309211016, + 0.0048341224901378155, + 0.009215399622917175, + 0.01043054647743702, + 0.005243122112005949, + 0.005692543927580118, + 0.005633625201880932, + 0.008830485865473747, + 0.009164595976471901, + 0.012467253021895885, + 0.010356093756854534, + 0.014539706520736217, + 0.010710722766816616, + 0.010702772997319698, + 0.0075956047512590885, + 0.01531666237860918, + 0.006854756269603968, + 0.00897955521941185, + 0.02308936044573784, + 0.010899183340370655, + 0.005869308952242136, + 0.0074243503622710705, + 0.0063458215445280075, + 0.00970705971121788, + 0.009929399937391281, + 0.00744086317718029, + 0.008287457749247551, + 0.007390356156975031, + 0.004826249089092016, + 0.011878730729222298, + 0.008143816143274307, + 0.012866363860666752, + 0.0168034378439188, + 0.012558361515402794, + 0.022544974461197853, + 0.009798937477171421, + 0.014250160194933414, + 0.012748530134558678, + 0.018527910113334656, + 0.02417980134487152, + 0.02082657441496849, + 0.03867432475090027, + 0.08038891106843948, + 0.030633723363280296, + 0.01441218052059412, + 0.020106563344597816, + 0.007276941556483507, + 0.005575762130320072, + 0.007116968277841806, + 0.014144936576485634, + 0.008476595394313335, + 0.007209387607872486, + 0.00908085610717535, + 0.0069176084361970425, + 0.0035392134450376034, + 0.007027909159660339, + 0.004225424490869045, + 0.0032344015780836344, + 0.004540642723441124, + 0.004497140180319548, + 0.0028589870780706406, + 0.003189369570463896, + 0.004246814176440239, + 0.003660868387669325, + 0.0037205691915005445, + 0.005554853938519955, + 0.005169008858501911, + 0.004692657385021448, + 0.0056501878425478935, + 0.003982678055763245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.0024953915271908045 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person perspective looking down from a very low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, predominantly gray and white cat paw visible taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrantly colored rubber dog toy ball, brightly striped with red and yellow, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pile of fluffy white cat litter resembling miniature sand dunes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paw push the ball forward and then scrape the litter from the side to cover it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633998032659292 + ], + "opposite_logprob_tokens": [ + -5.379633903503418 + ], + "logprob_reward": -0.004633998032659292, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633998032659292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13204199075698853, + 0.09653067588806152, + 0.08319728076457977, + 0.47956565022468567, + 0.08261974900960922, + 0.030550727620720863, + 0.02925102226436138, + 0.02337033487856388, + 0.026675308123230934, + 0.016197239980101585 + ], + "credit_map": [ + 0.013430705294013023, + 0.015727881342172623, + 0.01547954324632883, + 0.010113788768649101, + 0.015325285494327545, + 0.01670130155980587, + 0.0220197644084692, + 0.023243727162480354, + 0.010103967040777206, + 0.021337540820240974, + 0.018328635022044182, + 0.012250407598912716, + 0.006915952078998089, + 0.011901916936039925, + 0.00777343288064003, + 0.00791882537305355, + 0.00525590730831027, + 0.013461836613714695, + 0.010978671722114086, + 0.015654059126973152, + 0.00774565851315856, + 0.008859447203576565, + 0.01292805839329958, + 0.008313640020787716, + 0.04687187075614929, + 0.10409349948167801, + 0.06343772262334824, + 0.025402652099728584, + 0.037230491638183594, + 0.10650493204593658, + 0.028235184028744698, + 0.06778930872678757, + 0.009985269047319889, + 0.01221715472638607, + 0.010725889354944229, + 0.0048845368437469006, + 0.009388815611600876, + 0.010496378876268864, + 0.007688096724450588, + 0.01723361946642399, + 0.0025840154848992825, + 0.003818417899310589, + 0.005407529417425394, + 0.0028081326745450497, + 0.004585664253681898, + 0.0031459075398743153, + 0.00359550304710865, + 0.004605557303875685, + 0.0024422763381153345, + 0.006569456774741411, + 0.0036884562578052282, + 0.0033261922653764486, + 0.004797489382326603, + 0.0035504151601344347, + 0.0026599185075610876, + 0.0022168199066072702, + 0.002446732483804226, + 0.002592594362795353, + 0.0024153748527169228, + 0.0043610516004264355, + 0.0031744723673909903, + 0.002582348883152008, + 0.002948484383523464, + 0.002849278738722205, + 0.002483965363353491, + 0.004746810533106327, + 0.003232231130823493, + 0.003819543169811368, + 0.0034425724297761917, + 0.003684964030981064, + 0.002116973977535963, + 0.003148248651996255, + 0.001985112903639674, + 0.0028769387863576412, + 0.002924600150436163, + 0.0021258846390992403, + 0.0021101252641528845, + 0.0012796791270375252, + 0.0013646823354065418, + 0.0015302174724638462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.004633998032659292 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person perspective looking down from a very low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, predominantly gray and white cat paw visible taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrantly colored rubber dog toy ball, brightly striped with red and yellow, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pile of fluffy white cat litter resembling miniature sand dunes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paw push the ball forward and then scrape the litter from the side to cover it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291445791721344 + ], + "opposite_logprob_tokens": [ + -2.237914562225342 + ], + "logprob_reward": -0.11291445791721344, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291445791721344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16407090425491333, + 0.06321887671947479, + 0.03564750775694847, + 0.051433488726615906, + 0.03794175758957863, + 0.060242846608161926, + 0.2656911015510559, + 0.1720070242881775, + 0.0363609679043293, + 0.11338554322719574 + ], + "credit_map": [ + 0.02544480562210083, + 0.029812533408403397, + 0.0214737206697464, + 0.011062120087444782, + 0.010545150376856327, + 0.023161696270108223, + 0.0169586893171072, + 0.02561219409108162, + 0.004478130489587784, + 0.00799779687076807, + 0.0061794463545084, + 0.007560332305729389, + 0.011648686602711678, + 0.008583049289882183, + 0.008601648733019829, + 0.008169782347977161, + 0.0025110668502748013, + 0.003968433011323214, + 0.006774297449737787, + 0.004088758956640959, + 0.004087110515683889, + 0.0061663235537707806, + 0.00356265134178102, + 0.004488867241889238, + 0.00533665157854557, + 0.01057576946914196, + 0.005378542467951775, + 0.004369061905890703, + 0.009405170567333698, + 0.0055705346167087555, + 0.004553791601210833, + 0.006243965122848749, + 0.0038649383932352066, + 0.008635890670120716, + 0.00459480844438076, + 0.002916118362918496, + 0.005396687425673008, + 0.004790067672729492, + 0.002387422602623701, + 0.005355821456760168, + 0.006800269708037376, + 0.011973045766353607, + 0.01189409289509058, + 0.0042846896685659885, + 0.004326569382101297, + 0.007247431669384241, + 0.004494090098887682, + 0.009222650900483131, + 0.016788236796855927, + 0.04432939738035202, + 0.01920178346335888, + 0.06937798112630844, + 0.05528906732797623, + 0.020061016082763672, + 0.01709628477692604, + 0.023547329008579254, + 0.02483881264925003, + 0.015020495280623436, + 0.01938272826373577, + 0.04230504110455513, + 0.018056942149996758, + 0.01766004040837288, + 0.020963571965694427, + 0.013779381290078163, + 0.004146295133978128, + 0.006306574679911137, + 0.005901968106627464, + 0.004661045037209988, + 0.004076638724654913, + 0.0040999287739396095, + 0.003077867440879345, + 0.004090646747499704, + 0.011220064014196396, + 0.024782124906778336, + 0.015614821575582027, + 0.012020561844110489, + 0.011171640828251839, + 0.00885805580765009, + 0.012880911119282246, + 0.016837354749441147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.11291445791721344 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "textured, frame-by-frame animated shot, captured from a first-person perspective, looks down from a very low angle. A fluffy, predominantly gray and white cat paw takes center stage. This paw, focused yet clumsily, attempts to propel a vibrantly colored rubber dog toy ball—brightly striped with red and yellow—into a pile of fluffy cat litter, composed of fine white particles, resembling miniature sand dunes. Each movement is marked by a distinct pause. The paw first laboriously pushes the ball forward, then, with its pink pads and slightly open claws, persistently and persistently scrapes the litter from the side, attempting to cover the incongruous colored ball. Soft, warm indoor light streams from above and side, illuminating the edges of the cat's fur and casting subtle, yet distinct, shadows on the litter. The entire scene, shrouded in this clumsy effort, is imbued with an air of innocent, endearing cuteness.", + "seed": 1799, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_10/index_2048_seed_1799_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06691768642049283, + "vlm_logprob": -0.06691768642049283, + "vlm_yes_no_margin": 3.924999976158142, + "vlm_token_credit": -0.06691768642049283, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person perspective looking down from a very low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, predominantly gray and white cat paw visible taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrantly colored rubber dog toy ball, brightly striped with red and yellow, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pile of fluffy white cat litter resembling miniature sand dunes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paw push the ball forward and then scrape the litter from the side to cover it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11423958837985992 + ], + "opposite_logprob_tokens": [ + -2.3642396926879883 + ], + "logprob_reward": -0.11423958837985992, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11423958837985992, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14428775012493134, + 0.06365307420492172, + 0.05249203369021416, + 0.05677036568522453, + 0.02648288942873478, + 0.3876034617424011, + 0.10329063981771469, + 0.05916731059551239, + 0.059697236865758896, + 0.04655524343252182 + ], + "credit_map": [ + 0.023081641644239426, + 0.018786940723657608, + 0.028339127078652382, + 0.014260535128414631, + 0.014987945556640625, + 0.016691336408257484, + 0.01214852649718523, + 0.0159916989505291, + 0.005915527697652578, + 0.011557333171367645, + 0.007248478941619396, + 0.00681047048419714, + 0.007063128985464573, + 0.01030662376433611, + 0.0041053113527596, + 0.010646197944879532, + 0.0043424987234175205, + 0.010852211154997349, + 0.010188567452132702, + 0.005328943952918053, + 0.0059956144541502, + 0.006508446764200926, + 0.003728229086846113, + 0.005547523032873869, + 0.004894640296697617, + 0.010754958726465702, + 0.006396063137799501, + 0.009833836928009987, + 0.007476824801415205, + 0.006201147101819515, + 0.004326464608311653, + 0.0068864356726408005, + 0.0023746206425130367, + 0.003576589748263359, + 0.005280682351440191, + 0.0030101777520030737, + 0.0024241763167083263, + 0.004318840801715851, + 0.0021799548994749784, + 0.0033178457524627447, + 0.05035775527358055, + 0.15852400660514832, + 0.03104441799223423, + 0.014251204207539558, + 0.02001849003136158, + 0.03333042562007904, + 0.01620483212172985, + 0.0638723075389862, + 0.00661483034491539, + 0.016058562323451042, + 0.011222348548471928, + 0.01040260773152113, + 0.023721810430288315, + 0.011002276092767715, + 0.01272872556000948, + 0.011539486236870289, + 0.005440083798021078, + 0.00642519211396575, + 0.014542453922331333, + 0.006608179304748774, + 0.006774927955120802, + 0.004754525609314442, + 0.007623541634529829, + 0.006998407654464245, + 0.0051462906412780285, + 0.007756514474749565, + 0.007660900708287954, + 0.007407168857753277, + 0.006585969123989344, + 0.006799713708460331, + 0.009259800426661968, + 0.00908088032156229, + 0.004043548367917538, + 0.007601527962833643, + 0.011591759510338306, + 0.005793440155684948, + 0.004011986311525106, + 0.006432168185710907, + 0.0030890151392668486, + 0.00399179570376873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.11423958837985992 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person perspective looking down from a very low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, predominantly gray and white cat paw visible taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrantly colored rubber dog toy ball, brightly striped with red and yellow, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pile of fluffy white cat litter resembling miniature sand dunes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paw push the ball forward and then scrape the litter from the side to cover it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906683325767517 + ], + "opposite_logprob_tokens": [ + -2.464066743850708 + ], + "logprob_reward": -0.08906683325767517, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906683325767517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33993712067604065, + 0.1043066680431366, + 0.04968133941292763, + 0.05118565261363983, + 0.07228004187345505, + 0.0881495401263237, + 0.10510988533496857, + 0.06680287420749664, + 0.06359115242958069, + 0.05895572528243065 + ], + "credit_map": [ + 0.054818328469991684, + 0.05812469497323036, + 0.03662728890776634, + 0.039822231978178024, + 0.03681844845414162, + 0.047742512077093124, + 0.026719773188233376, + 0.03926384449005127, + 0.011801126413047314, + 0.015000530518591404, + 0.008984599262475967, + 0.01346510462462902, + 0.011944006197154522, + 0.02007538639008999, + 0.007278990000486374, + 0.015756919980049133, + 0.003737344406545162, + 0.008094332180917263, + 0.005257259588688612, + 0.008055978454649448, + 0.005280357785522938, + 0.00781464483588934, + 0.00428685313090682, + 0.007154572289437056, + 0.0035071794409304857, + 0.007115927990525961, + 0.005728051997721195, + 0.006326592992991209, + 0.01141136884689331, + 0.0059659164398908615, + 0.004452995490282774, + 0.006677617784589529, + 0.007466998882591724, + 0.016174525022506714, + 0.016263199970126152, + 0.005483584478497505, + 0.00608835369348526, + 0.010837488807737827, + 0.0038530987221747637, + 0.006112797185778618, + 0.00852627120912075, + 0.02877655439078808, + 0.006595185492187738, + 0.005891883745789528, + 0.007615400478243828, + 0.007788780145347118, + 0.004405500832945108, + 0.01854996383190155, + 0.008703770115971565, + 0.013609195128083229, + 0.013879300095140934, + 0.014973529614508152, + 0.014659211970865726, + 0.012332976795732975, + 0.013639736920595169, + 0.013312163762748241, + 0.0050813681446015835, + 0.008172709494829178, + 0.008807484060525894, + 0.006027448456734419, + 0.009288109838962555, + 0.010110058821737766, + 0.008546535857021809, + 0.010769156739115715, + 0.004721546545624733, + 0.009564357809722424, + 0.006174231413751841, + 0.013184083625674248, + 0.0059234206564724445, + 0.007113310042768717, + 0.008179149590432644, + 0.008731050416827202, + 0.0058570606634020805, + 0.011180220171809196, + 0.014066765084862709, + 0.006350203882902861, + 0.004219398833811283, + 0.007433007936924696, + 0.004353215452283621, + 0.005495856050401926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.08906683325767517 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person perspective looking down from a very low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, predominantly gray and white cat paw visible taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrantly colored rubber dog toy ball, brightly striped with red and yellow, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pile of fluffy white cat litter resembling miniature sand dunes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paw push the ball forward and then scrape the litter from the side to cover it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006413786904886365 + ], + "opposite_logprob_tokens": [ + -7.375641345977783 + ], + "logprob_reward": -0.0006413786904886365, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006413786904886365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10867173224687576, + 0.06879091262817383, + 0.1087309792637825, + 0.0641903281211853, + 0.027960818260908127, + 0.16085729002952576, + 0.11541592329740524, + 0.04900487884879112, + 0.23527710139751434, + 0.06110003963112831 + ], + "credit_map": [ + 0.009453644044697285, + 0.01694635860621929, + 0.011541860178112984, + 0.015880856662988663, + 0.01442964468151331, + 0.014763451181352139, + 0.013808169402182102, + 0.01184774748980999, + 0.008675592020154, + 0.014431701973080635, + 0.005331286229193211, + 0.007326080463826656, + 0.006878300104290247, + 0.010117288678884506, + 0.00467555969953537, + 0.011355103924870491, + 0.011428605765104294, + 0.0359969362616539, + 0.011941476725041866, + 0.00993514247238636, + 0.009973407723009586, + 0.014100625179708004, + 0.005675086751580238, + 0.00967970211058855, + 0.006501669529825449, + 0.012782757170498371, + 0.009204484522342682, + 0.005299526732414961, + 0.008324623107910156, + 0.008010593242943287, + 0.0030729828868061304, + 0.010993688367307186, + 0.0029357471503317356, + 0.004128372762352228, + 0.005507537163794041, + 0.004044371657073498, + 0.001969758188351989, + 0.0032570427283644676, + 0.0021359098609536886, + 0.00398207874968648, + 0.018257800489664078, + 0.06286697089672089, + 0.011294946074485779, + 0.007157426793128252, + 0.00934728141874075, + 0.014421377331018448, + 0.006794743239879608, + 0.03071674145758152, + 0.008332236669957638, + 0.0157035943120718, + 0.013144348748028278, + 0.01708090864121914, + 0.02464284375309944, + 0.011251275427639484, + 0.01566391810774803, + 0.009596799500286579, + 0.003258221782743931, + 0.005862329620867968, + 0.005058969836682081, + 0.0053909579291939735, + 0.009455275721848011, + 0.006258591543883085, + 0.00823668297380209, + 0.005483848042786121, + 0.01825919933617115, + 0.03650585934519768, + 0.023152437061071396, + 0.03491190820932388, + 0.02992420271039009, + 0.021479783579707146, + 0.04516474902629852, + 0.025878947228193283, + 0.004254390485584736, + 0.006796891335397959, + 0.006435402203351259, + 0.013181059621274471, + 0.0072168996557593346, + 0.007776849437505007, + 0.007751434575766325, + 0.007687109522521496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 6 + ], + "selected_reward": -0.0006413786904886365 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person perspective looking down from a very low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, predominantly gray and white cat paw visible taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrantly colored rubber dog toy ball, brightly striped with red and yellow, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pile of fluffy white cat litter resembling miniature sand dunes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paw push the ball forward and then scrape the litter from the side to cover it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003613254288211465 + ], + "opposite_logprob_tokens": [ + -5.628613471984863 + ], + "logprob_reward": -0.003613254288211465, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003613254288211465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10335336625576019, + 0.15671604871749878, + 0.05232173204421997, + 0.055535346269607544, + 0.03527237847447395, + 0.06501046568155289, + 0.2696707546710968, + 0.19343455135822296, + 0.036782339215278625, + 0.03190302848815918 + ], + "credit_map": [ + 0.01619667559862137, + 0.012140614911913872, + 0.015005565248429775, + 0.012021264992654324, + 0.01004378218203783, + 0.014797580428421497, + 0.009560031816363335, + 0.013587851077318192, + 0.009913572110235691, + 0.03647467866539955, + 0.011046884581446648, + 0.02019580639898777, + 0.018399059772491455, + 0.019595671445131302, + 0.01547516230493784, + 0.025615213438868523, + 0.005399678833782673, + 0.01700730249285698, + 0.006653334945440292, + 0.005749060306698084, + 0.0037822111044079065, + 0.006498116999864578, + 0.003247265936806798, + 0.00398475956171751, + 0.005559924989938736, + 0.015414491295814514, + 0.007036798633635044, + 0.004840061068534851, + 0.005249539390206337, + 0.007339127827435732, + 0.0025313058868050575, + 0.007564101368188858, + 0.003484095213934779, + 0.004286044742912054, + 0.00737380376085639, + 0.0051697916351258755, + 0.004116789437830448, + 0.0046944888308644295, + 0.0026774315629154444, + 0.003469934919849038, + 0.008230114355683327, + 0.026092050597071648, + 0.004494466353207827, + 0.002705461112782359, + 0.0034723402932286263, + 0.006286857184022665, + 0.0028689587488770485, + 0.010860213078558445, + 0.015924349427223206, + 0.026828665286302567, + 0.02159878984093666, + 0.018723463639616966, + 0.057633791118860245, + 0.050886787474155426, + 0.031150424852967262, + 0.046924516558647156, + 0.011594375595450401, + 0.020252536982297897, + 0.016246702522039413, + 0.01644255593419075, + 0.038735248148441315, + 0.035075072199106216, + 0.024452803656458855, + 0.030635254457592964, + 0.0027588121592998505, + 0.005433077923953533, + 0.004582637920975685, + 0.0033616176806390285, + 0.0038849362172186375, + 0.005594492424279451, + 0.003910238854587078, + 0.0072565264999866486, + 0.0031102169305086136, + 0.004611116833984852, + 0.007819889113307, + 0.003085531760007143, + 0.00254400004632771, + 0.005761716980487108, + 0.0019112376030534506, + 0.003059317823499441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 1 + ], + "selected_reward": -0.003613254288211465 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person perspective looking down from a very low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, predominantly gray and white cat paw visible taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrantly colored rubber dog toy ball, brightly striped with red and yellow, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pile of fluffy white cat litter resembling miniature sand dunes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paw push the ball forward and then scrape the litter from the side to cover it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12702737748622894 + ], + "opposite_logprob_tokens": [ + -2.1270272731781006 + ], + "logprob_reward": -0.12702737748622894, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12702737748622894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11768986284732819, + 0.1200437918305397, + 0.09803663939237595, + 0.08053051680326462, + 0.04019785672426224, + 0.14886386692523956, + 0.09071699529886246, + 0.1427803486585617, + 0.12505342066287994, + 0.03608664870262146 + ], + "credit_map": [ + 0.015085109509527683, + 0.015216090716421604, + 0.013955028727650642, + 0.01643783040344715, + 0.012243625707924366, + 0.021064138039946556, + 0.011409331113100052, + 0.012278713285923004, + 0.01394388172775507, + 0.031200554221868515, + 0.011285101994872093, + 0.015618515200912952, + 0.012079145759344101, + 0.012791956774890423, + 0.012635444290935993, + 0.0104892048984766, + 0.009059793315827847, + 0.015055910684168339, + 0.015398121438920498, + 0.013857307843863964, + 0.011774742044508457, + 0.010968568734824657, + 0.008751538582146168, + 0.013170662336051464, + 0.008198065683245659, + 0.01286252774298191, + 0.01137094758450985, + 0.01253560371696949, + 0.009181883186101913, + 0.010672022588551044, + 0.006071405950933695, + 0.009638061746954918, + 0.0037545771338045597, + 0.006156447343528271, + 0.005219524260610342, + 0.006000943947583437, + 0.004969453904777765, + 0.0064729549922049046, + 0.0027049912605434656, + 0.004918965976685286, + 0.015420441515743732, + 0.0532233901321888, + 0.009479871019721031, + 0.008133252151310444, + 0.010379754938185215, + 0.012697925791144371, + 0.006319383159279823, + 0.033209867775440216, + 0.010179713368415833, + 0.012682362459599972, + 0.015352597460150719, + 0.015335814096033573, + 0.009149989113211632, + 0.007399624213576317, + 0.011656371876597404, + 0.008960524573922157, + 0.010172918438911438, + 0.016229476779699326, + 0.03027196228504181, + 0.013491885736584663, + 0.021395865827798843, + 0.016321608796715736, + 0.019443519413471222, + 0.01545313186943531, + 0.010363563895225525, + 0.01513590943068266, + 0.014235028997063637, + 0.020153431221842766, + 0.013044684194028378, + 0.015235341154038906, + 0.018282048404216766, + 0.01860341802239418, + 0.0033061630092561245, + 0.005444675218313932, + 0.005384158343076706, + 0.0045977020636200905, + 0.0042968625202775, + 0.006860417779535055, + 0.002674535382539034, + 0.0035221383441239595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 8 + ], + "selected_reward": -0.12702737748622894 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "textured, frame-by-frame animated shot, captured from a first-person perspective, looks down from a very low angle. A fluffy, predominantly gray and white cat paw takes center stage. This paw, focused yet clumsily, attempts to propel a vibrantly colored rubber dog toy ball—brightly striped with red and yellow—into a pile of fluffy cat litter, composed of fine white particles, resembling miniature sand dunes. Each movement is marked by a distinct pause. The paw first laboriously pushes the ball forward, then, with its pink pads and slightly open claws, persistently and persistently scrapes the litter from the side, attempting to cover the incongruous colored ball. Soft, warm indoor light streams from above and side, illuminating the edges of the cat's fur and casting subtle, yet distinct, shadows on the litter. The entire scene, shrouded in this clumsy effort, is imbued with an air of innocent, endearing cuteness.", + "seed": 1800, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_10/index_2048_seed_1800_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05080451511312276, + "vlm_logprob": -0.05080451511312276, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.05080451511312276, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person perspective looking down from a very low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, predominantly gray and white cat paw visible taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrantly colored rubber dog toy ball, brightly striped with red and yellow, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pile of fluffy white cat litter resembling miniature sand dunes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paw push the ball forward and then scrape the litter from the side to cover it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15644878149032593 + ], + "opposite_logprob_tokens": [ + -2.0314488410949707 + ], + "logprob_reward": -0.15644878149032593, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.15644878149032593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0949067547917366, + 0.10949472337961197, + 0.17657150328159332, + 0.05735734850168228, + 0.13627244532108307, + 0.09039656072854996, + 0.06488652527332306, + 0.07491160929203033, + 0.13496904075145721, + 0.0602334588766098 + ], + "credit_map": [ + 0.01061189454048872, + 0.008517970331013203, + 0.012098981998860836, + 0.011710667982697487, + 0.01665457896888256, + 0.013624824583530426, + 0.010319682769477367, + 0.011368146166205406, + 0.009394819848239422, + 0.012435520067811012, + 0.01770014315843582, + 0.010538018308579922, + 0.01685239188373089, + 0.018130965530872345, + 0.012343406677246094, + 0.012099460698664188, + 0.007640534546226263, + 0.07399924844503403, + 0.005280455574393272, + 0.008323987945914268, + 0.010977217927575111, + 0.004928707145154476, + 0.008793535642325878, + 0.05662781745195389, + 0.004136921837925911, + 0.006576229818165302, + 0.009740883484482765, + 0.006369773764163256, + 0.010229237377643585, + 0.007447593845427036, + 0.004749320447444916, + 0.008107381872832775, + 0.0140615776181221, + 0.013982249423861504, + 0.031287286430597305, + 0.013830752111971378, + 0.014845459721982479, + 0.011859574355185032, + 0.017573459073901176, + 0.01883208379149437, + 0.010422144085168839, + 0.008322449401021004, + 0.01245210226625204, + 0.016209591180086136, + 0.011334067210555077, + 0.010037630796432495, + 0.010477027855813503, + 0.01114155538380146, + 0.0056974864564836025, + 0.005596010480076075, + 0.0066907103173434734, + 0.007939495146274567, + 0.012317300774157047, + 0.00967894122004509, + 0.006906193681061268, + 0.0100603848695755, + 0.007980764843523502, + 0.0065196119248867035, + 0.009969820268452168, + 0.00951207522302866, + 0.012690084986388683, + 0.012986496090888977, + 0.00561992172151804, + 0.009632828645408154, + 0.009943592362105846, + 0.013673161156475544, + 0.03574470430612564, + 0.010394327342510223, + 0.015874965116381645, + 0.013454779982566833, + 0.009053236804902554, + 0.02683028019964695, + 0.0062064556404948235, + 0.006065450608730316, + 0.010548543184995651, + 0.004693209659308195, + 0.012676824815571308, + 0.005585941486060619, + 0.004958872217684984, + 0.009498164989054203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 8 + ], + "selected_reward": -0.15644878149032593 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person perspective looking down from a very low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, predominantly gray and white cat paw visible taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrantly colored rubber dog toy ball, brightly striped with red and yellow, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pile of fluffy white cat litter resembling miniature sand dunes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paw push the ball forward and then scrape the litter from the side to cover it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007652136497199535 + ], + "opposite_logprob_tokens": [ + -4.882652282714844 + ], + "logprob_reward": -0.007652136497199535, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007652136497199535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05420425906777382, + 0.09927263855934143, + 0.24553966522216797, + 0.2039443701505661, + 0.06507928669452667, + 0.046692803502082825, + 0.07690110057592392, + 0.14004302024841309, + 0.03875376656651497, + 0.029569115489721298 + ], + "credit_map": [ + 0.007389918435364962, + 0.004837309010326862, + 0.010320091620087624, + 0.005253667943179607, + 0.007788714487105608, + 0.006417165510356426, + 0.006256277672946453, + 0.005941108334809542, + 0.010191483423113823, + 0.009601465426385403, + 0.0159028060734272, + 0.010883267968893051, + 0.013662033714354038, + 0.01428688783198595, + 0.01471989881247282, + 0.010024777613580227, + 0.012681526131927967, + 0.09457702189683914, + 0.007566663436591625, + 0.009107033722102642, + 0.01366173755377531, + 0.013996889814734459, + 0.012964282184839249, + 0.08098448067903519, + 0.015452466905117035, + 0.016270065680146217, + 0.034788522869348526, + 0.01436591800302267, + 0.04722652956843376, + 0.026197684928774834, + 0.01608310081064701, + 0.03356005251407623, + 0.006218722555786371, + 0.011915420182049274, + 0.00962059386074543, + 0.010274437256157398, + 0.005607008468359709, + 0.003332936903461814, + 0.012017374858260155, + 0.006092782597988844, + 0.00702317152172327, + 0.004476469475775957, + 0.006322245113551617, + 0.008066946640610695, + 0.004471179563552141, + 0.004141396842896938, + 0.00450710067525506, + 0.007684288080781698, + 0.00547985453158617, + 0.006403104867786169, + 0.01571638695895672, + 0.012969822622835636, + 0.007395760156214237, + 0.006374975200742483, + 0.008970913477241993, + 0.013590275309979916, + 0.01080147922039032, + 0.0119323106482625, + 0.01695805974304676, + 0.010140113532543182, + 0.026152629405260086, + 0.030104756355285645, + 0.011000778526067734, + 0.02295287325978279, + 0.0030542989261448383, + 0.002466048812493682, + 0.007443481590598822, + 0.0028235758654773235, + 0.006114979740232229, + 0.0054849558509886265, + 0.004085463937371969, + 0.007280958816409111, + 0.003496543038636446, + 0.002221479080617428, + 0.003070484148338437, + 0.0031032655388116837, + 0.00577167421579361, + 0.00344834104180336, + 0.004419147036969662, + 0.004038176964968443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 7 + ], + "selected_reward": -0.007652136497199535 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person perspective looking down from a very low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, predominantly gray and white cat paw visible taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrantly colored rubber dog toy ball, brightly striped with red and yellow, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pile of fluffy white cat litter resembling miniature sand dunes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paw push the ball forward and then scrape the litter from the side to cover it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011874536285176873 + ], + "opposite_logprob_tokens": [ + -6.751187324523926 + ], + "logprob_reward": -0.0011874536285176873, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011874536285176873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14363738894462585, + 0.10437712073326111, + 0.26678502559661865, + 0.0878957062959671, + 0.08937845379114151, + 0.08885440230369568, + 0.07962428033351898, + 0.055406711995601654, + 0.03869336098432541, + 0.04534749314188957 + ], + "credit_map": [ + 0.014912553131580353, + 0.018807070329785347, + 0.018431510776281357, + 0.026829715818166733, + 0.020035270601511, + 0.01611095853149891, + 0.016084251925349236, + 0.012426081113517284, + 0.01158438716083765, + 0.008341801352798939, + 0.015316635370254517, + 0.011371023021638393, + 0.013190642930567265, + 0.01783830299973488, + 0.013163536787033081, + 0.013570801354944706, + 0.011321192607283592, + 0.11115237325429916, + 0.008152117021381855, + 0.009633985348045826, + 0.0192499328404665, + 0.009618835523724556, + 0.01548390369862318, + 0.08217271417379379, + 0.006091675255447626, + 0.00832679308950901, + 0.018631042912602425, + 0.012116994708776474, + 0.01036478765308857, + 0.009436721913516521, + 0.010531964711844921, + 0.012395733036100864, + 0.008822919800877571, + 0.016147809103131294, + 0.012390177696943283, + 0.013675929047167301, + 0.007472206838428974, + 0.005797703750431538, + 0.015391717664897442, + 0.00967999454587698, + 0.011226089671254158, + 0.009140376932919025, + 0.013095143251121044, + 0.017322327941656113, + 0.008320048451423645, + 0.006989614572376013, + 0.0098988963291049, + 0.012861911207437515, + 0.007379771675914526, + 0.008737429045140743, + 0.010036497376859188, + 0.01599513180553913, + 0.009302155114710331, + 0.008105740882456303, + 0.011706982739269733, + 0.008360576815903187, + 0.00506137078627944, + 0.004969402682036161, + 0.00777441868558526, + 0.005765988957136869, + 0.009682907722890377, + 0.009530366398394108, + 0.005055803805589676, + 0.007566459011286497, + 0.0035022401716560125, + 0.007058404851704836, + 0.006964879110455513, + 0.003884067991748452, + 0.005625266116112471, + 0.0032250091899186373, + 0.0043702819384634495, + 0.004063209984451532, + 0.003588533028960228, + 0.008230527862906456, + 0.006493131630122662, + 0.005625325255095959, + 0.005418980028480291, + 0.0054881032556295395, + 0.005149907432496548, + 0.005352988373488188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.0011874536285176873 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person perspective looking down from a very low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, predominantly gray and white cat paw visible taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrantly colored rubber dog toy ball, brightly striped with red and yellow, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pile of fluffy white cat litter resembling miniature sand dunes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paw push the ball forward and then scrape the litter from the side to cover it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977538712322712 + ], + "opposite_logprob_tokens": [ + -4.634775161743164 + ], + "logprob_reward": -0.00977538712322712, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977538712322712, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0615418367087841, + 0.05262574926018715, + 0.182917520403862, + 0.15191926062107086, + 0.02862156741321087, + 0.03536220267415047, + 0.08429746329784393, + 0.02544102631509304, + 0.16421547532081604, + 0.21305783092975616 + ], + "credit_map": [ + 0.008859977126121521, + 0.005962994880974293, + 0.012671535834670067, + 0.005544838961213827, + 0.008617698214948177, + 0.0068191043101251125, + 0.008464849554002285, + 0.00460084481164813, + 0.004187783692032099, + 0.005125518422573805, + 0.004537316504865885, + 0.003930526319891214, + 0.012220298871397972, + 0.00942389015108347, + 0.006754113826900721, + 0.006446304265409708, + 0.007852348499000072, + 0.07781879603862762, + 0.0036647755187004805, + 0.0060304091311991215, + 0.009318610653281212, + 0.005752261728048325, + 0.009197954088449478, + 0.06328237056732178, + 0.012181634083390236, + 0.013438615016639233, + 0.03527488932013512, + 0.009557757526636124, + 0.028842322528362274, + 0.018549595028162003, + 0.011798755265772343, + 0.02227570302784443, + 0.002573454286903143, + 0.002841873560100794, + 0.005858603399246931, + 0.0024852221831679344, + 0.0038190626073628664, + 0.0023728760425001383, + 0.005056696943938732, + 0.0036137807182967663, + 0.004532040096819401, + 0.003556539537385106, + 0.007011818699538708, + 0.004445993807166815, + 0.0038107868749648333, + 0.002723448211327195, + 0.003966758493334055, + 0.005314820446074009, + 0.0070195249281823635, + 0.007662639487534761, + 0.012854248285293579, + 0.014137815684080124, + 0.011805922724306583, + 0.008015173487365246, + 0.009939974173903465, + 0.012862177565693855, + 0.0028103827498853207, + 0.003467259230092168, + 0.003065722528845072, + 0.003727166447788477, + 0.0034545823000371456, + 0.003263527527451515, + 0.0026936000213027, + 0.002958789700642228, + 0.012130160816013813, + 0.011753268539905548, + 0.04601718857884407, + 0.009087935090065002, + 0.019970819354057312, + 0.015313535928726196, + 0.01108049787580967, + 0.03886208310723305, + 0.023604268208146095, + 0.013907899148762226, + 0.029940759763121605, + 0.012390036135911942, + 0.051674067974090576, + 0.02162325568497181, + 0.017821483314037323, + 0.0420960858464241 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 8 + ], + "selected_reward": -0.00977538712322712 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person perspective looking down from a very low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, predominantly gray and white cat paw visible taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrantly colored rubber dog toy ball, brightly striped with red and yellow, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pile of fluffy white cat litter resembling miniature sand dunes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paw push the ball forward and then scrape the litter from the side to cover it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895881682634354 + ], + "opposite_logprob_tokens": [ + -2.578958749771118 + ], + "logprob_reward": -0.07895881682634354, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895881682634354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11256039887666702, + 0.1467474400997162, + 0.15547896921634674, + 0.12479578703641891, + 0.06193351745605469, + 0.038527797907590866, + 0.11408711969852448, + 0.09818021953105927, + 0.0884774774312973, + 0.059211261570453644 + ], + "credit_map": [ + 0.01114469300955534, + 0.011013194918632507, + 0.015732519328594208, + 0.011285987682640553, + 0.02026582881808281, + 0.015229806303977966, + 0.013878596015274525, + 0.014009780250489712, + 0.010205033235251904, + 0.012326794676482677, + 0.027618365362286568, + 0.016734907403588295, + 0.02027188427746296, + 0.02563132718205452, + 0.016550922766327858, + 0.01740819774568081, + 0.013080907985568047, + 0.05752228945493698, + 0.006817658897489309, + 0.006658199243247509, + 0.013290878385305405, + 0.015314962714910507, + 0.014384726993739605, + 0.02840934693813324, + 0.01039664726704359, + 0.011686733923852444, + 0.030741097405552864, + 0.012660239823162556, + 0.020770039409399033, + 0.011072807013988495, + 0.009131643921136856, + 0.01833658665418625, + 0.006813525687903166, + 0.010051107034087181, + 0.011452223174273968, + 0.00784260593354702, + 0.006271379068493843, + 0.0044047837145626545, + 0.008205719292163849, + 0.006892175879329443, + 0.0035852573346346617, + 0.00439448282122612, + 0.006581394467502832, + 0.00612820265814662, + 0.003995239268988371, + 0.0027585099451243877, + 0.004742258228361607, + 0.006342452485114336, + 0.010702492669224739, + 0.012000140734016895, + 0.01653764769434929, + 0.02162778750061989, + 0.012821954675018787, + 0.01591429114341736, + 0.012510756962001324, + 0.011972044594585896, + 0.006842126604169607, + 0.008729474619030952, + 0.011927229352295399, + 0.009516993537545204, + 0.02007051557302475, + 0.01845417730510235, + 0.007915118709206581, + 0.01472458802163601, + 0.005699438042938709, + 0.006008679512888193, + 0.028931526467204094, + 0.005820159800350666, + 0.00785141158849001, + 0.0050441003404557705, + 0.006502437870949507, + 0.02261972986161709, + 0.006437089294195175, + 0.0076021007262170315, + 0.010786380618810654, + 0.005075447261333466, + 0.01070943009108305, + 0.004024222493171692, + 0.007579143159091473, + 0.0069974507205188274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.07895881682634354 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "textured, frame-by-frame animated shot, captured from a first-person perspective, looks down from a very low angle. A fluffy, predominantly gray and white cat paw takes center stage. This paw, focused yet clumsily, attempts to propel a vibrantly colored rubber dog toy ball—brightly striped with red and yellow—into a pile of fluffy cat litter, composed of fine white particles, resembling miniature sand dunes. Each movement is marked by a distinct pause. The paw first laboriously pushes the ball forward, then, with its pink pads and slightly open claws, persistently and persistently scrapes the litter from the side, attempting to cover the incongruous colored ball. Soft, warm indoor light streams from above and side, illuminating the edges of the cat's fur and casting subtle, yet distinct, shadows on the litter. The entire scene, shrouded in this clumsy effort, is imbued with an air of innocent, endearing cuteness.", + "seed": 1801, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_10/index_2048_seed_1801_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08953320630826056, + "vlm_logprob": -0.08953320630826056, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.08953320630826056, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person perspective looking down from a very low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, predominantly gray and white cat paw visible taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrantly colored rubber dog toy ball, brightly striped with red and yellow, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pile of fluffy white cat litter resembling miniature sand dunes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paw push the ball forward and then scrape the litter from the side to cover it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18232855200767517 + ], + "opposite_logprob_tokens": [ + -1.9323285818099976 + ], + "logprob_reward": -0.18232855200767517, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.18232855200767517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22746723890304565, + 0.05094386264681816, + 0.15478870272636414, + 0.07402782142162323, + 0.04438936710357666, + 0.031076572835445404, + 0.03076036274433136, + 0.15943101048469543, + 0.19369348883628845, + 0.0334215946495533 + ], + "credit_map": [ + 0.030934210866689682, + 0.03987407311797142, + 0.03174076974391937, + 0.021376213058829308, + 0.034236859530210495, + 0.023755479604005814, + 0.020430143922567368, + 0.0251194778829813, + 0.0062616122886538506, + 0.007373567670583725, + 0.0068403929471969604, + 0.00719252647832036, + 0.005209390074014664, + 0.006554942112416029, + 0.005909699480980635, + 0.00560173112899065, + 0.018959106877446175, + 0.014187228865921497, + 0.027442142367362976, + 0.019129909574985504, + 0.01627732627093792, + 0.023733891546726227, + 0.012949485331773758, + 0.02210959978401661, + 0.007848257198929787, + 0.008436107076704502, + 0.009938286617398262, + 0.010441570542752743, + 0.008712658658623695, + 0.009469976648688316, + 0.009561081416904926, + 0.009619880467653275, + 0.006009636912494898, + 0.008114323019981384, + 0.006068982649594545, + 0.0038670687936246395, + 0.003500883001834154, + 0.006571153644472361, + 0.004107982851564884, + 0.006149332504719496, + 0.003689419711008668, + 0.0028918141033500433, + 0.004576513078063726, + 0.00452954089269042, + 0.003696642816066742, + 0.0035456870682537556, + 0.0038701691664755344, + 0.0042767818085849285, + 0.003923484124243259, + 0.002781479386612773, + 0.005520109552890062, + 0.0034790141507983208, + 0.0032390744891017675, + 0.00422848341986537, + 0.00255710375495255, + 0.005031609907746315, + 0.012014024890959263, + 0.01649508997797966, + 0.025268318131566048, + 0.01643233560025692, + 0.01660194620490074, + 0.017590582370758057, + 0.0325075127184391, + 0.022521190345287323, + 0.016150932759046555, + 0.017047598958015442, + 0.03847625479102135, + 0.01977476291358471, + 0.02780899964272976, + 0.022827954962849617, + 0.026439614593982697, + 0.025167375802993774, + 0.0047208345495164394, + 0.00240304390899837, + 0.0033293827436864376, + 0.002897287718951702, + 0.005322965793311596, + 0.00561943044885993, + 0.004619692452251911, + 0.004508954472839832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.18232855200767517 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person perspective looking down from a very low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, predominantly gray and white cat paw visible taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrantly colored rubber dog toy ball, brightly striped with red and yellow, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pile of fluffy white cat litter resembling miniature sand dunes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paw push the ball forward and then scrape the litter from the side to cover it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02638404630124569 + ], + "opposite_logprob_tokens": [ + -3.651384115219116 + ], + "logprob_reward": -0.02638404630124569, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02638404630124569, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2620684504508972, + 0.07325027883052826, + 0.21819135546684265, + 0.06659215688705444, + 0.08906551450490952, + 0.026020873337984085, + 0.04728703945875168, + 0.0750330463051796, + 0.11074157804250717, + 0.03174976259469986 + ], + "credit_map": [ + 0.019821884110569954, + 0.021302923560142517, + 0.04673976078629494, + 0.02604275569319725, + 0.05831831693649292, + 0.030674975365400314, + 0.03271520137786865, + 0.026452617719769478, + 0.006773834116756916, + 0.00999497715383768, + 0.013190479017794132, + 0.007483556400984526, + 0.008060986176133156, + 0.007255829870700836, + 0.008432583883404732, + 0.012058037333190441, + 0.031467437744140625, + 0.02454189397394657, + 0.03594476729631424, + 0.02567523904144764, + 0.022957460954785347, + 0.030115922912955284, + 0.017793362960219383, + 0.02969527058303356, + 0.005754624027758837, + 0.00890089850872755, + 0.010303649120032787, + 0.006897187326103449, + 0.009376143105328083, + 0.00667699845507741, + 0.005780463572591543, + 0.012902195565402508, + 0.007517710328102112, + 0.008581952191889286, + 0.012068445794284344, + 0.0143199497833848, + 0.01102374866604805, + 0.00922224298119545, + 0.01439549122005701, + 0.011935977265238762, + 0.003135081147775054, + 0.002283252775669098, + 0.0043347617611289024, + 0.0033629618119448423, + 0.0022558902855962515, + 0.003200124716386199, + 0.002703613368794322, + 0.00474518584087491, + 0.006258645094931126, + 0.0044935159385204315, + 0.008087653666734695, + 0.005165483336895704, + 0.004543280228972435, + 0.006426149047911167, + 0.003818060737103224, + 0.008494253270328045, + 0.006633723620325327, + 0.006714544724673033, + 0.012825955636799335, + 0.009092487394809723, + 0.007909909822046757, + 0.007707678247243166, + 0.013134296983480453, + 0.011014445684850216, + 0.009488913230597973, + 0.00923882145434618, + 0.02325868420302868, + 0.009083052165806293, + 0.016581431031227112, + 0.014064203947782516, + 0.014313185587525368, + 0.014713287353515625, + 0.003377666464075446, + 0.0027567828074097633, + 0.0033685460221022367, + 0.004099453333765268, + 0.005057736299932003, + 0.005372972227632999, + 0.003709660843014717, + 0.004006947390735149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.02638404630124569 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person perspective looking down from a very low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, predominantly gray and white cat paw visible taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrantly colored rubber dog toy ball, brightly striped with red and yellow, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pile of fluffy white cat litter resembling miniature sand dunes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paw push the ball forward and then scrape the litter from the side to cover it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036256073508411646 + ], + "opposite_logprob_tokens": [ + -5.628625392913818 + ], + "logprob_reward": -0.0036256073508411646, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036256073508411646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1908985674381256, + 0.13566745817661285, + 0.2039092779159546, + 0.07926876842975616, + 0.060881197452545166, + 0.02205709181725979, + 0.021904772147536278, + 0.12806515395641327, + 0.07286307215690613, + 0.08448465913534164 + ], + "credit_map": [ + 0.02066764608025551, + 0.02881374955177307, + 0.03056911751627922, + 0.02238417975604534, + 0.027018027380108833, + 0.012197542004287243, + 0.02915080077946186, + 0.02009749226272106, + 0.012869245372712612, + 0.023055994883179665, + 0.0243449155241251, + 0.015319082885980606, + 0.013740688562393188, + 0.013798287138342857, + 0.01305555272847414, + 0.019483698531985283, + 0.026961607858538628, + 0.01934347301721573, + 0.03297422081232071, + 0.025061974301934242, + 0.022316228598356247, + 0.03236372023820877, + 0.01832466758787632, + 0.02656337432563305, + 0.008982688188552856, + 0.011187209747731686, + 0.011780020780861378, + 0.009653766639530659, + 0.00934552401304245, + 0.010304487310349941, + 0.00897697638720274, + 0.009038099087774754, + 0.0052537452429533005, + 0.007513672579079866, + 0.007895929738879204, + 0.010074859485030174, + 0.008187624625861645, + 0.0064214798621833324, + 0.007543990854173899, + 0.007989892736077309, + 0.0025443551130592823, + 0.0020515872165560722, + 0.003641989780589938, + 0.0022417812142521143, + 0.001642427989281714, + 0.003150245873257518, + 0.0020182004664093256, + 0.004766503348946571, + 0.0023230742663145065, + 0.0019744017627090216, + 0.003862676676362753, + 0.00274583138525486, + 0.0019857874140143394, + 0.0031486370135098696, + 0.0018171574920415878, + 0.00404720613732934, + 0.013366780243813992, + 0.012854734435677528, + 0.019857017323374748, + 0.013664309866726398, + 0.011909299530088902, + 0.015968505293130875, + 0.022151878103613853, + 0.018292630091309547, + 0.008575115352869034, + 0.007861645892262459, + 0.014454697258770466, + 0.005255472846329212, + 0.008043886162340641, + 0.008553224615752697, + 0.00993887335062027, + 0.010180152952671051, + 0.010463491082191467, + 0.006448420695960522, + 0.010128568857908249, + 0.008451702073216438, + 0.011986549943685532, + 0.010498528368771076, + 0.014868476428091526, + 0.011638922616839409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.0036256073508411646 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person perspective looking down from a very low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, predominantly gray and white cat paw visible taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrantly colored rubber dog toy ball, brightly striped with red and yellow, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pile of fluffy white cat litter resembling miniature sand dunes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paw push the ball forward and then scrape the litter from the side to cover it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009777040220797062 + ], + "opposite_logprob_tokens": [ + -4.634777069091797 + ], + "logprob_reward": -0.009777040220797062, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009777040220797062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13572844862937927, + 0.05868383124470711, + 0.04132199287414551, + 0.04397868365049362, + 0.060309264808893204, + 0.05951335281133652, + 0.02488945797085762, + 0.09615232795476913, + 0.059089481830596924, + 0.4203331768512726 + ], + "credit_map": [ + 0.010919807478785515, + 0.01577736623585224, + 0.022430680692195892, + 0.013354489579796791, + 0.027322273701429367, + 0.012324920855462551, + 0.016086798161268234, + 0.017512122169137, + 0.004808257333934307, + 0.0053580026142299175, + 0.009227070957422256, + 0.005380710121244192, + 0.008681211620569229, + 0.005701038055121899, + 0.008234444074332714, + 0.011293101124465466, + 0.0049770730547606945, + 0.003236840944737196, + 0.006777365226298571, + 0.006215985864400864, + 0.003652945626527071, + 0.005887018516659737, + 0.005306718870997429, + 0.005268045701086521, + 0.005539159756153822, + 0.0063066426664590836, + 0.006212176755070686, + 0.005280858837068081, + 0.004699665121734142, + 0.004824097268283367, + 0.004829889163374901, + 0.0062861996702849865, + 0.010665033012628555, + 0.010593196377158165, + 0.007950171828269958, + 0.004756939131766558, + 0.0042593530379235744, + 0.009926184080541134, + 0.003927276004105806, + 0.008231115527451038, + 0.006170515436679125, + 0.006821409799158573, + 0.009034944698214531, + 0.00683166179805994, + 0.0043748486787080765, + 0.009487074799835682, + 0.004937180317938328, + 0.011855718679726124, + 0.0034230765886604786, + 0.0021400710102170706, + 0.004217622801661491, + 0.0033551661763340235, + 0.002477172762155533, + 0.0033801111858338118, + 0.001758463098667562, + 0.004137775860726833, + 0.012168848887085915, + 0.008027536794543266, + 0.015240437351167202, + 0.008159792982041836, + 0.008193768560886383, + 0.01342490129172802, + 0.016878753900527954, + 0.014058299362659454, + 0.0060796416364610195, + 0.005706229247152805, + 0.010624224320054054, + 0.005215678829699755, + 0.007962814532220364, + 0.009368596598505974, + 0.006871299352496862, + 0.007261001039296389, + 0.06345783919095993, + 0.026265623047947884, + 0.048609066754579544, + 0.024716656655073166, + 0.07221098244190216, + 0.06287705898284912, + 0.06218675523996353, + 0.060009200125932693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.009777040220797062 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a first-person perspective looking down from a very low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, predominantly gray and white cat paw visible taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrantly colored rubber dog toy ball, brightly striped with red and yellow, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pile of fluffy white cat litter resembling miniature sand dunes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paw push the ball forward and then scrape the litter from the side to cover it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255507856607437 + ], + "opposite_logprob_tokens": [ + -1.6005507707595825 + ], + "logprob_reward": -0.2255507856607437, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255507856607437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22737081348896027, + 0.08446840196847916, + 0.10209139436483383, + 0.07524055987596512, + 0.11325307190418243, + 0.09712691605091095, + 0.03273489698767662, + 0.09444598853588104, + 0.11094837635755539, + 0.06231959909200668 + ], + "credit_map": [ + 0.024809066206216812, + 0.023989643901586533, + 0.03836328163743019, + 0.026011142879724503, + 0.03666608780622482, + 0.022017737850546837, + 0.025557605549693108, + 0.029956238344311714, + 0.007750248070806265, + 0.011041507124900818, + 0.01375985611230135, + 0.01315747108310461, + 0.007634864654392004, + 0.008707381784915924, + 0.012821890413761139, + 0.009595186449587345, + 0.009199217893183231, + 0.008420757949352264, + 0.020169898867607117, + 0.013373658061027527, + 0.007871532812714577, + 0.01459523942321539, + 0.009839573875069618, + 0.018621506169438362, + 0.006520818453282118, + 0.009028088301420212, + 0.010433346964418888, + 0.009696157649159431, + 0.00879944209009409, + 0.007969316095113754, + 0.009821038693189621, + 0.012972347438335419, + 0.009420901536941528, + 0.008720348589122295, + 0.016896843910217285, + 0.02111908420920372, + 0.014950689859688282, + 0.011001194827258587, + 0.014958552084863186, + 0.016185451298952103, + 0.010901987552642822, + 0.009253455325961113, + 0.015574326738715172, + 0.013896018266677856, + 0.010304942727088928, + 0.012320326641201973, + 0.010120324790477753, + 0.014755524694919586, + 0.006444818805903196, + 0.0030188991222530603, + 0.006193739827722311, + 0.001965619157999754, + 0.0023804570082575083, + 0.005760322790592909, + 0.002098345896229148, + 0.0048726932145655155, + 0.00732495216652751, + 0.00902021024376154, + 0.01753276214003563, + 0.013803594745695591, + 0.011361474171280861, + 0.010638728737831116, + 0.013787336647510529, + 0.010976933874189854, + 0.00934393610805273, + 0.009699905291199684, + 0.022549103945493698, + 0.012183601967990398, + 0.016020143404603004, + 0.013273962773382664, + 0.01374726090580225, + 0.014130457304418087, + 0.011367686092853546, + 0.003529229434207082, + 0.006476426497101784, + 0.004600939340889454, + 0.009672910906374454, + 0.009157107211649418, + 0.008328959345817566, + 0.009186330251395702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.2255507856607437 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_11/index_tensor([1871, 1871])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_11/index_tensor([1871, 1871])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..358a86b847701924853c80cbff900df6f5029de0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_11/index_tensor([1871, 1871])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static medium shot, shrouded in the warm, yellowish hues of old films, transports us to a dimly lit ancient military tent. The sole illumination comes from a flickering oil lamp on a rough wooden table. The dim light casts large, oscillating shadows on the heavy canvas walls, creating a stark contrast between light and dark. Clad in a suit of dark, scratched, and weathered iron armor, Cao Cao stands with his back to the camera, slightly bent over a large parchment military map spread out on the table. Mountains, rivers, and military markers are outlined in ink. His long, tied-up hair is streaked with silver strands, and the shoulders of his armor reflect the faint glow of the oil lamp. He extends his armored hand, his finger hovering for a moment over a city marked on the map. Then, his shoulders visibly slump, and his body rises and falls with a heavy, drawn-out sigh. A puff of white air escapes his lips, quickly dissipating in the cold air. His furrowed brow and tense jawline appear particularly clear in the light and shadow of his profile, and the whole scene is filled with the heavy pressure before making a strategic decision and the lonely atmosphere of a hero in his twilight years.", + "seed": 32873, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_11/index_1871_seed_32873_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03347033807076514, + "vlm_logprob": -0.03347033807076514, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.03347033807076514, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient military tent with heavy canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Cao Cao visible wearing dark, scratched, and weathered iron armor with long hair streaked with silver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large parchment military map with ink outlines of mountains, rivers, and markers spread on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character extend an armored hand to hover a finger over a city on the map before slumping his shoulders and sighing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single flickering oil lamp casting large, oscillating shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00914740189909935 + ], + "opposite_logprob_tokens": [ + -4.8841471672058105 + ], + "logprob_reward": -0.00914740189909935, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00914740189909935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33606046438217163, + 0.09848121553659439, + 0.1400281935930252, + 0.06058196723461151, + 0.05232662707567215, + 0.08561211079359055, + 0.054145682603120804, + 0.03638363629579544, + 0.054875291883945465, + 0.08150485903024673 + ], + "credit_map": [ + 0.038299985229969025, + 0.026729591190814972, + 0.04716353863477707, + 0.03967937454581261, + 0.04877249523997307, + 0.04107218608260155, + 0.062128208577632904, + 0.03221507370471954, + 0.008401851169764996, + 0.007622577715665102, + 0.010556298308074474, + 0.00893414206802845, + 0.022404413670301437, + 0.01237599179148674, + 0.01621115580201149, + 0.011974784545600414, + 0.014205174520611763, + 0.014750493690371513, + 0.014462790451943874, + 0.021222399547696114, + 0.02476455643773079, + 0.019024495035409927, + 0.016438109800219536, + 0.015160187147557735, + 0.010773166082799435, + 0.007394189015030861, + 0.005862159188836813, + 0.004782168660312891, + 0.00615561380982399, + 0.009676973335444927, + 0.008023900911211967, + 0.007913799956440926, + 0.006219842005521059, + 0.007123555522412062, + 0.006990525405853987, + 0.00985852349549532, + 0.00540851941332221, + 0.007595520466566086, + 0.004032899159938097, + 0.00509724160656333, + 0.009749424643814564, + 0.007920688018202782, + 0.0070152743719518185, + 0.006874877959489822, + 0.015577994287014008, + 0.01157431025058031, + 0.016749920323491096, + 0.010149624198675156, + 0.0032371364068239927, + 0.004391157068312168, + 0.005150856915861368, + 0.0053358604200184345, + 0.013316530734300613, + 0.0071623255498707294, + 0.008963536471128464, + 0.006588282063603401, + 0.003777812235057354, + 0.004016935359686613, + 0.00529090641066432, + 0.004475357010960579, + 0.004830921534448862, + 0.0052438597194850445, + 0.004063578322529793, + 0.004684263840317726, + 0.0037816092371940613, + 0.006523118354380131, + 0.006000225432217121, + 0.008689644746482372, + 0.008204839192330837, + 0.007251386530697346, + 0.007470827549695969, + 0.006953639909625053, + 0.006481291726231575, + 0.005323110613971949, + 0.01002774853259325, + 0.007730934303253889, + 0.017912566661834717, + 0.00987192615866661, + 0.015436528250575066, + 0.008720746263861656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.00914740189909935 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient military tent with heavy canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Cao Cao visible wearing dark, scratched, and weathered iron armor with long hair streaked with silver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large parchment military map with ink outlines of mountains, rivers, and markers spread on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character extend an armored hand to hover a finger over a city on the map before slumping his shoulders and sighing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single flickering oil lamp casting large, oscillating shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812495619058609 + ], + "opposite_logprob_tokens": [ + -3.2881250381469727 + ], + "logprob_reward": -0.03812495619058609, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812495619058609, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28995582461357117, + 0.11413861066102982, + 0.1544748842716217, + 0.05383234843611717, + 0.02377500757575035, + 0.019565759226679802, + 0.03572886437177658, + 0.03134099394083023, + 0.12943263351917267, + 0.14775504171848297 + ], + "credit_map": [ + 0.01853129453957081, + 0.01861058548092842, + 0.028658008202910423, + 0.03569357842206955, + 0.05699818953871727, + 0.041077565401792526, + 0.0590866357088089, + 0.03129996359348297, + 0.008497754111886024, + 0.009727489203214645, + 0.011635531671345234, + 0.009412210434675217, + 0.02812490053474903, + 0.01403046678751707, + 0.017182037234306335, + 0.015528236515820026, + 0.01275368221104145, + 0.016857434064149857, + 0.015485912561416626, + 0.03198316693305969, + 0.02497675269842148, + 0.01953505165874958, + 0.01847711391746998, + 0.014405785128474236, + 0.00819806195795536, + 0.004501252435147762, + 0.005289017688483, + 0.0036007787566632032, + 0.0060133193619549274, + 0.009408564306795597, + 0.008127516135573387, + 0.008693844079971313, + 0.002681571524590254, + 0.001982062356546521, + 0.0027844500727951527, + 0.0022219938691705465, + 0.004190513864159584, + 0.0031274084467440844, + 0.003767122980207205, + 0.0030198858585208654, + 0.001932077226229012, + 0.0025048740208148956, + 0.0020355996675789356, + 0.0035757706500589848, + 0.0029853766318410635, + 0.002338949590921402, + 0.002155172871425748, + 0.002037938917055726, + 0.0026589189656078815, + 0.0026598311960697174, + 0.0029292565304785967, + 0.0036651534028351307, + 0.007489609997719526, + 0.004844981711357832, + 0.0068640559911727905, + 0.004617062862962484, + 0.0031673493795096874, + 0.003116070060059428, + 0.0036853416822850704, + 0.0023846051190048456, + 0.004237591288983822, + 0.004919841885566711, + 0.0038653144147247076, + 0.005964880809187889, + 0.009935720823705196, + 0.008025546558201313, + 0.012968757189810276, + 0.00879524927586317, + 0.030430199578404427, + 0.018734022974967957, + 0.021039722487330437, + 0.01950342208147049, + 0.00891623180359602, + 0.008484327234327793, + 0.01701423153281212, + 0.013173887506127357, + 0.035681892186403275, + 0.0193349476903677, + 0.028110727667808533, + 0.01703878864645958 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.03812495619058609 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient military tent with heavy canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Cao Cao visible wearing dark, scratched, and weathered iron armor with long hair streaked with silver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large parchment military map with ink outlines of mountains, rivers, and markers spread on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character extend an armored hand to hover a finger over a city on the map before slumping his shoulders and sighing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single flickering oil lamp casting large, oscillating shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00595568772405386 + ], + "opposite_logprob_tokens": [ + -5.130955696105957 + ], + "logprob_reward": -0.00595568772405386, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00595568772405386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23943908512592316, + 0.11592315137386322, + 0.23120377957820892, + 0.0710335224866867, + 0.08274240046739578, + 0.03663186356425285, + 0.05653997138142586, + 0.04982231184840202, + 0.07250433415174484, + 0.04415950924158096 + ], + "credit_map": [ + 0.02113855630159378, + 0.01802123337984085, + 0.02345818281173706, + 0.03872223570942879, + 0.040256015956401825, + 0.032335612922906876, + 0.041843052953481674, + 0.023664213716983795, + 0.015469552017748356, + 0.010006667114794254, + 0.011919130571186543, + 0.008625401183962822, + 0.024405449628829956, + 0.01576831378042698, + 0.016713593155145645, + 0.01301504671573639, + 0.02456265687942505, + 0.03131073713302612, + 0.025951804593205452, + 0.04863051325082779, + 0.027087081223726273, + 0.030375508591532707, + 0.024394014850258827, + 0.018891476094722748, + 0.008523548021912575, + 0.005783188622444868, + 0.006843335460871458, + 0.004528973717242479, + 0.008317910134792328, + 0.011376282200217247, + 0.014366916380822659, + 0.011293373070657253, + 0.015477343462407589, + 0.009287908673286438, + 0.010777844116091728, + 0.00812675803899765, + 0.006922789849340916, + 0.011543349362909794, + 0.008984805084764957, + 0.011621605604887009, + 0.0039541409350931644, + 0.004219571128487587, + 0.0033399376552551985, + 0.004999089520424604, + 0.007087074685841799, + 0.0036813891492784023, + 0.005805371794849634, + 0.003545293351635337, + 0.0046222023665905, + 0.003958635963499546, + 0.007066456601023674, + 0.009863964281976223, + 0.005888951011002064, + 0.008212476037442684, + 0.008684287779033184, + 0.00824300479143858, + 0.006492819171398878, + 0.006554254796355963, + 0.007176158018410206, + 0.006186057347804308, + 0.004972224123775959, + 0.006925405468791723, + 0.004970690701156855, + 0.006544703617691994, + 0.005594824906438589, + 0.007238590624183416, + 0.006231863051652908, + 0.010738248936831951, + 0.01358287874609232, + 0.00954829528927803, + 0.011364910751581192, + 0.008204721845686436, + 0.0031796463299542665, + 0.0044991434551775455, + 0.005239393096417189, + 0.006516080815345049, + 0.007456901483237743, + 0.005761011503636837, + 0.006956799421459436, + 0.004550532903522253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.00595568772405386 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient military tent with heavy canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Cao Cao visible wearing dark, scratched, and weathered iron armor with long hair streaked with silver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large parchment military map with ink outlines of mountains, rivers, and markers spread on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character extend an armored hand to hover a finger over a city on the map before slumping his shoulders and sighing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single flickering oil lamp casting large, oscillating shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294214427471161 + ], + "opposite_logprob_tokens": [ + -2.2379422187805176 + ], + "logprob_reward": -0.11294214427471161, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294214427471161, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25918179750442505, + 0.11379200220108032, + 0.16503790020942688, + 0.0644519031047821, + 0.03540737181901932, + 0.06699232757091522, + 0.08220288902521133, + 0.05204951763153076, + 0.08143419027328491, + 0.07945006340742111 + ], + "credit_map": [ + 0.019439123570919037, + 0.021745501086115837, + 0.027607588097453117, + 0.038198161870241165, + 0.04527498036623001, + 0.034993112087249756, + 0.0445534810423851, + 0.027369866147637367, + 0.015272092074155807, + 0.011551053263247013, + 0.01161816157400608, + 0.008305834606289864, + 0.02187671698629856, + 0.016176339238882065, + 0.014427428133785725, + 0.01456437911838293, + 0.01988617330789566, + 0.018876757472753525, + 0.01809203438460827, + 0.020895831286907196, + 0.026902372017502785, + 0.025023965165019035, + 0.01754436455667019, + 0.017816394567489624, + 0.006131852976977825, + 0.006802941206842661, + 0.005725860130041838, + 0.005995112471282482, + 0.009754161350429058, + 0.00969476904720068, + 0.009698367677628994, + 0.010648834519088268, + 0.0037642952520400286, + 0.003364161355420947, + 0.003221009159460664, + 0.003065850120037794, + 0.0064720530062913895, + 0.0054372986778616905, + 0.005776364356279373, + 0.004306336399167776, + 0.005030746106058359, + 0.005226177629083395, + 0.006503145210444927, + 0.009511272422969341, + 0.012917276471853256, + 0.0074755046516656876, + 0.012120666913688183, + 0.008207542821764946, + 0.008252919651567936, + 0.008033654652535915, + 0.011830063536763191, + 0.008502861484885216, + 0.014909213408827782, + 0.010387919843196869, + 0.011773286387324333, + 0.008512969128787518, + 0.0052656265906989574, + 0.005301835481077433, + 0.006566260010004044, + 0.004817906301468611, + 0.009509029798209667, + 0.009090524166822433, + 0.00528878765180707, + 0.006209545768797398, + 0.004695261362940073, + 0.0066213468089699745, + 0.006204553879797459, + 0.009836854413151741, + 0.018165696412324905, + 0.011050504632294178, + 0.014330682344734669, + 0.010529283434152603, + 0.005796996410936117, + 0.006420171353965998, + 0.007188958115875721, + 0.006944787222892046, + 0.015627112239599228, + 0.010044286027550697, + 0.014717253856360912, + 0.012710493057966232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.11294214427471161 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient military tent with heavy canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Cao Cao visible wearing dark, scratched, and weathered iron armor with long hair streaked with silver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large parchment military map with ink outlines of mountains, rivers, and markers spread on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character extend an armored hand to hover a finger over a city on the map before slumping his shoulders and sighing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single flickering oil lamp casting large, oscillating shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011815002653747797 + ], + "opposite_logprob_tokens": [ + -6.751181602478027 + ], + "logprob_reward": -0.0011815002653747797, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011815002653747797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16875244677066803, + 0.0689750462770462, + 0.1441413313150406, + 0.0459941029548645, + 0.04927870258688927, + 0.09429710358381271, + 0.07161139696836472, + 0.05637679621577263, + 0.23718877136707306, + 0.06338425725698471 + ], + "credit_map": [ + 0.02766451984643936, + 0.016754262149333954, + 0.017288684844970703, + 0.01906212791800499, + 0.020061159506440163, + 0.024915415793657303, + 0.01618904247879982, + 0.026817236095666885, + 0.0066708787344396114, + 0.00888747163116932, + 0.00996397715061903, + 0.011237052269279957, + 0.00932600162923336, + 0.009002148173749447, + 0.006737074349075556, + 0.007150451187044382, + 0.012831566855311394, + 0.016486456617712975, + 0.016823137179017067, + 0.027340972796082497, + 0.020635120570659637, + 0.020756544545292854, + 0.015423865057528019, + 0.01384367048740387, + 0.006799370516091585, + 0.005410811398178339, + 0.0053093754686415195, + 0.003126234747469425, + 0.003552004462108016, + 0.008567046374082565, + 0.005608291830867529, + 0.007620965130627155, + 0.00650997506454587, + 0.003919405862689018, + 0.004538910463452339, + 0.005061828065663576, + 0.006916137412190437, + 0.00880415365099907, + 0.006280837114900351, + 0.007247455418109894, + 0.01002283301204443, + 0.009318290278315544, + 0.007920579053461552, + 0.006918307859450579, + 0.01673837937414646, + 0.013187224976718426, + 0.01708471216261387, + 0.01310677919536829, + 0.007887264713644981, + 0.005526158958673477, + 0.0098281754180789, + 0.008490004576742649, + 0.013955628499388695, + 0.00803778599947691, + 0.009987575002014637, + 0.007898812182247639, + 0.0068673319183290005, + 0.00767379067838192, + 0.006297331769019365, + 0.007005283609032631, + 0.00740323681384325, + 0.008234268054366112, + 0.004685659892857075, + 0.008209889754652977, + 0.01638641022145748, + 0.027050817385315895, + 0.024729683995246887, + 0.044063396751880646, + 0.03516937047243118, + 0.03836696222424507, + 0.02523965761065483, + 0.02618248388171196, + 0.006765266880393028, + 0.007686768192797899, + 0.008621800690889359, + 0.008674553595483303, + 0.008186389692127705, + 0.009712206199765205, + 0.006688260938972235, + 0.007049014791846275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.0011815002653747797 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, shrouded in the warm, yellowish hues of old films, transports us to a dimly lit ancient military tent. The sole illumination comes from a flickering oil lamp on a rough wooden table. The dim light casts large, oscillating shadows on the heavy canvas walls, creating a stark contrast between light and dark. Clad in a suit of dark, scratched, and weathered iron armor, Cao Cao stands with his back to the camera, slightly bent over a large parchment military map spread out on the table. Mountains, rivers, and military markers are outlined in ink. His long, tied-up hair is streaked with silver strands, and the shoulders of his armor reflect the faint glow of the oil lamp. He extends his armored hand, his finger hovering for a moment over a city marked on the map. Then, his shoulders visibly slump, and his body rises and falls with a heavy, drawn-out sigh. A puff of white air escapes his lips, quickly dissipating in the cold air. His furrowed brow and tense jawline appear particularly clear in the light and shadow of his profile, and the whole scene is filled with the heavy pressure before making a strategic decision and the lonely atmosphere of a hero in his twilight years.", + "seed": 32874, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_11/index_1871_seed_32874_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02022637934423983, + "vlm_logprob": -0.02022637934423983, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.02022637934423983, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient military tent with heavy canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Cao Cao visible wearing dark, scratched, and weathered iron armor with long hair streaked with silver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large parchment military map with ink outlines of mountains, rivers, and markers spread on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character extend an armored hand to hover a finger over a city on the map before slumping his shoulders and sighing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single flickering oil lamp casting large, oscillating shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015737032517790794 + ], + "opposite_logprob_tokens": [ + -4.390737056732178 + ], + "logprob_reward": -0.015737032517790794, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.015737032517790794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1741810292005539, + 0.08793091773986816, + 0.056139927357435226, + 0.08864076435565948, + 0.08277266472578049, + 0.15114851295948029, + 0.13853107392787933, + 0.07497166842222214, + 0.06339576095342636, + 0.08228772133588791 + ], + "credit_map": [ + 0.019190920516848564, + 0.02671756222844124, + 0.024409260600805283, + 0.03626803681254387, + 0.017522739246487617, + 0.017069803550839424, + 0.014874632470309734, + 0.01812804490327835, + 0.010620820336043835, + 0.013036675751209259, + 0.013615088537335396, + 0.015575173310935497, + 0.00904557853937149, + 0.010078255087137222, + 0.007560912054032087, + 0.008398408070206642, + 0.006716816686093807, + 0.00791675690561533, + 0.009562947787344456, + 0.008769514970481396, + 0.005386756267398596, + 0.007930690422654152, + 0.00457402691245079, + 0.005282411817461252, + 0.00742654874920845, + 0.008430066518485546, + 0.01708737015724182, + 0.017299769446253777, + 0.013469638302922249, + 0.011174432933330536, + 0.006686676759272814, + 0.007066246587783098, + 0.00785141997039318, + 0.008940720930695534, + 0.014327391982078552, + 0.01631544716656208, + 0.009788572788238525, + 0.011575319804251194, + 0.006641884334385395, + 0.007331897504627705, + 0.011600545607507229, + 0.015087531879544258, + 0.02449066936969757, + 0.0195122379809618, + 0.028597775846719742, + 0.021965114399790764, + 0.016094405204057693, + 0.01380021870136261, + 0.013495669700205326, + 0.011615230701863766, + 0.021511100232601166, + 0.016711058095097542, + 0.017060857266187668, + 0.03569598123431206, + 0.00983238872140646, + 0.012608782388269901, + 0.006062946282327175, + 0.009057515300810337, + 0.00821808073669672, + 0.008488251827657223, + 0.007081921678036451, + 0.017851583659648895, + 0.0078690554946661, + 0.010342308320105076, + 0.005565699189901352, + 0.009268348105251789, + 0.009467673487961292, + 0.006568050477653742, + 0.0055166189558804035, + 0.013717014342546463, + 0.006264448631554842, + 0.007027902640402317, + 0.008596361614763737, + 0.01683838479220867, + 0.01168227382004261, + 0.009441161528229713, + 0.011620651930570602, + 0.007421079091727734, + 0.007413580548018217, + 0.009274213574826717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.015737032517790794 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient military tent with heavy canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Cao Cao visible wearing dark, scratched, and weathered iron armor with long hair streaked with silver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large parchment military map with ink outlines of mountains, rivers, and markers spread on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character extend an armored hand to hover a finger over a city on the map before slumping his shoulders and sighing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single flickering oil lamp casting large, oscillating shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026373248547315598 + ], + "opposite_logprob_tokens": [ + -3.6513731479644775 + ], + "logprob_reward": -0.026373248547315598, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026373248547315598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19289277493953705, + 0.05977668985724449, + 0.036212313920259476, + 0.027061717584729195, + 0.027105994522571564, + 0.11180128902196884, + 0.11283232271671295, + 0.09724951535463333, + 0.08937553316354752, + 0.24569179117679596 + ], + "credit_map": [ + 0.01868513785302639, + 0.016886331140995026, + 0.04611588642001152, + 0.04452778026461601, + 0.023074541240930557, + 0.013645796105265617, + 0.015296023339033127, + 0.014661268331110477, + 0.0074204085394740105, + 0.005334596615284681, + 0.01401232648640871, + 0.011140352115035057, + 0.006190962623804808, + 0.006552667357027531, + 0.004832394886761904, + 0.0042929803021252155, + 0.0038416695315390825, + 0.0036406628787517548, + 0.007576738949865103, + 0.006320703774690628, + 0.004453039728105068, + 0.004639362450689077, + 0.00288977287709713, + 0.0028503602370619774, + 0.002747348975390196, + 0.0026815852615982294, + 0.003934372682124376, + 0.005034422501921654, + 0.003259735880419612, + 0.004381753038614988, + 0.0025684835854917765, + 0.0024540128652006388, + 0.0024827648885548115, + 0.0025478184688836336, + 0.004649324342608452, + 0.004422828089445829, + 0.004580152221024036, + 0.003127564210444689, + 0.0023981535341590643, + 0.002897389465942979, + 0.009101088158786297, + 0.011700167320668697, + 0.017167815938591957, + 0.016941945999860764, + 0.014324929565191269, + 0.019292067736387253, + 0.011097729206085205, + 0.012175546027719975, + 0.011435562744736671, + 0.009705577045679092, + 0.018871480599045753, + 0.014500120654702187, + 0.014374212361872196, + 0.024669693782925606, + 0.008344563655555248, + 0.010931113734841347, + 0.009128685109317303, + 0.011272651143372059, + 0.013408896513283253, + 0.015491069294512272, + 0.013140494003891945, + 0.02074345387518406, + 0.005866548977792263, + 0.008197713643312454, + 0.007733055390417576, + 0.013836348429322243, + 0.012580006383359432, + 0.009291171096265316, + 0.010197670198976994, + 0.018335679545998573, + 0.007400213275104761, + 0.010001391172409058, + 0.027056576684117317, + 0.04380965977907181, + 0.03622688725590706, + 0.0391215980052948, + 0.033620841801166534, + 0.020956242457032204, + 0.020107917487621307, + 0.024792054668068886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.026373248547315598 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient military tent with heavy canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Cao Cao visible wearing dark, scratched, and weathered iron armor with long hair streaked with silver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large parchment military map with ink outlines of mountains, rivers, and markers spread on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character extend an armored hand to hover a finger over a city on the map before slumping his shoulders and sighing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single flickering oil lamp casting large, oscillating shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003624300705268979 + ], + "opposite_logprob_tokens": [ + -5.628624439239502 + ], + "logprob_reward": -0.003624300705268979, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003624300705268979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10537073761224747, + 0.07168219238519669, + 0.06897000223398209, + 0.06763457506895065, + 0.04458937421441078, + 0.10372599214315414, + 0.07802983373403549, + 0.20751819014549255, + 0.07725394517183304, + 0.1752251386642456 + ], + "credit_map": [ + 0.010713440366089344, + 0.009891441091895103, + 0.020055698230862617, + 0.021354427561163902, + 0.01303275115787983, + 0.011005706153810024, + 0.010192688554525375, + 0.00912458822131157, + 0.008593645878136158, + 0.008909530006349087, + 0.012724194675683975, + 0.01736512966454029, + 0.0054811653681099415, + 0.006679656449705362, + 0.006496633403003216, + 0.005432236008346081, + 0.007158758584409952, + 0.0070932297967374325, + 0.01247984915971756, + 0.014024541713297367, + 0.007034022361040115, + 0.009071256034076214, + 0.006258776411414146, + 0.005849568173289299, + 0.006203293800354004, + 0.005516559351235628, + 0.013516255654394627, + 0.01267883088439703, + 0.01042825449258089, + 0.007680454291403294, + 0.006254851818084717, + 0.005356074310839176, + 0.003464271780103445, + 0.0034990916028618813, + 0.007226110901683569, + 0.01083416398614645, + 0.004745374899357557, + 0.0059400880709290504, + 0.0050246333703398705, + 0.003855644492432475, + 0.008891625329852104, + 0.014210526831448078, + 0.0197296105325222, + 0.01491821464151144, + 0.013518244959414005, + 0.010812987573444843, + 0.010674147866666317, + 0.010970640927553177, + 0.009249521419405937, + 0.008253575302660465, + 0.011054474860429764, + 0.010299396701157093, + 0.007587215397506952, + 0.019374245777726173, + 0.005188359413295984, + 0.007023044396191835, + 0.013559099286794662, + 0.026518456637859344, + 0.02171643264591694, + 0.02361401915550232, + 0.02779492177069187, + 0.04399741813540459, + 0.02400258369743824, + 0.026315251365303993, + 0.005388082470744848, + 0.01247613225132227, + 0.011452625505626202, + 0.009126149117946625, + 0.006837954744696617, + 0.018031857907772064, + 0.005830794107168913, + 0.008110348135232925, + 0.022104166448116302, + 0.0270076896995306, + 0.020679999142885208, + 0.01736496575176716, + 0.028299253433942795, + 0.023725520819425583, + 0.017955433577299118, + 0.018088096752762794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.003624300705268979 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient military tent with heavy canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Cao Cao visible wearing dark, scratched, and weathered iron armor with long hair streaked with silver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large parchment military map with ink outlines of mountains, rivers, and markers spread on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character extend an armored hand to hover a finger over a city on the map before slumping his shoulders and sighing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single flickering oil lamp casting large, oscillating shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865863919258118 + ], + "opposite_logprob_tokens": [ + -3.048658609390259 + ], + "logprob_reward": -0.04865863919258118, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865863919258118, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1802530288696289, + 0.12341783195734024, + 0.04119107499718666, + 0.04770400747656822, + 0.04264490678906441, + 0.1212780773639679, + 0.12361886352300644, + 0.06141019985079765, + 0.15443480014801025, + 0.10404720902442932 + ], + "credit_map": [ + 0.01498339045792818, + 0.013189718127250671, + 0.044333748519420624, + 0.04147623851895332, + 0.02047659642994404, + 0.017213700339198112, + 0.015279638580977917, + 0.013300010934472084, + 0.008850784040987492, + 0.008283417671918869, + 0.026530174538493156, + 0.02685588039457798, + 0.018061302602291107, + 0.012743393890559673, + 0.012841396033763885, + 0.009251485578715801, + 0.004522946197539568, + 0.004478503484278917, + 0.005522390827536583, + 0.007922017946839333, + 0.004889047238975763, + 0.007175121922045946, + 0.0034452134277671576, + 0.003235832555219531, + 0.004354423843324184, + 0.003797275247052312, + 0.009528335183858871, + 0.01149249728769064, + 0.006486825179308653, + 0.0053617400117218494, + 0.003139484440907836, + 0.00354342395439744, + 0.004014838021248579, + 0.003333417698740959, + 0.006327457260340452, + 0.007852672599256039, + 0.005913096014410257, + 0.00651779118925333, + 0.0037707637529820204, + 0.004914869088679552, + 0.009754886850714684, + 0.01127684023231268, + 0.02219906449317932, + 0.018537070602178574, + 0.022012416273355484, + 0.013489389792084694, + 0.01215258240699768, + 0.011855823919177055, + 0.013275829143822193, + 0.011273490265011787, + 0.016266372054815292, + 0.02322942577302456, + 0.017809590324759483, + 0.015569165349006653, + 0.011580266058444977, + 0.014614725485444069, + 0.0057794079184532166, + 0.008405427448451519, + 0.007435991428792477, + 0.0071473438292741776, + 0.006598304491490126, + 0.010730366222560406, + 0.006508296821266413, + 0.008805060759186745, + 0.01081112027168274, + 0.029060961678624153, + 0.021522190421819687, + 0.019249847158789635, + 0.01584562472999096, + 0.025593310594558716, + 0.012241087853908539, + 0.020110659301280975, + 0.009403456933796406, + 0.013414991088211536, + 0.01486181654036045, + 0.02382911555469036, + 0.008593196049332619, + 0.013017283752560616, + 0.012556763365864754, + 0.008370589464902878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.04865863919258118 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient military tent with heavy canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Cao Cao visible wearing dark, scratched, and weathered iron armor with long hair streaked with silver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large parchment military map with ink outlines of mountains, rivers, and markers spread on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character extend an armored hand to hover a finger over a city on the map before slumping his shoulders and sighing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single flickering oil lamp casting large, oscillating shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738675758242607 + ], + "opposite_logprob_tokens": [ + -5.006738662719727 + ], + "logprob_reward": -0.006738675758242607, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738675758242607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13010911643505096, + 0.03575171157717705, + 0.02548101358115673, + 0.04170774295926094, + 0.06511402130126953, + 0.33781465888023376, + 0.07727038115262985, + 0.0752076506614685, + 0.0566377267241478, + 0.1549060046672821 + ], + "credit_map": [ + 0.010577247478067875, + 0.012750571593642235, + 0.01820441335439682, + 0.025516321882605553, + 0.014689470641314983, + 0.01879146695137024, + 0.014802368357777596, + 0.014777245000004768, + 0.003567073028534651, + 0.003652987303212285, + 0.005561928730458021, + 0.007282169535756111, + 0.004484314005821943, + 0.004500634968280792, + 0.003721204586327076, + 0.002981396857649088, + 0.0021190247498452663, + 0.0026650922372937202, + 0.0036167472135275602, + 0.005345073528587818, + 0.0027777529321610928, + 0.0041383556090295315, + 0.0022782147862017155, + 0.002540749963372946, + 0.0042332676239311695, + 0.0036784664262086153, + 0.005617562681436539, + 0.006689105648547411, + 0.005673922598361969, + 0.006838482338935137, + 0.0038926699198782444, + 0.005084260832518339, + 0.005262929480522871, + 0.005018253345042467, + 0.008855310268700123, + 0.010468417778611183, + 0.006865732837468386, + 0.016146205365657806, + 0.004552927333861589, + 0.007944244891405106, + 0.02919889986515045, + 0.036142393946647644, + 0.04724939912557602, + 0.04723051190376282, + 0.052272483706474304, + 0.04781724140048027, + 0.03834173455834389, + 0.03956195339560509, + 0.007149891462177038, + 0.008587056770920753, + 0.011534112505614758, + 0.009381087496876717, + 0.007290007080882788, + 0.016343142837285995, + 0.0071307276375591755, + 0.009854349307715893, + 0.005250571295619011, + 0.006660645827651024, + 0.009476189501583576, + 0.013881989754736423, + 0.00940065085887909, + 0.017594724893569946, + 0.005876974668353796, + 0.007065894082188606, + 0.00474063865840435, + 0.006939108017832041, + 0.007729302626103163, + 0.007216232363134623, + 0.006897760555148125, + 0.009012493304908276, + 0.006340814754366875, + 0.0077613708563148975, + 0.015421607531607151, + 0.025765812024474144, + 0.020882118493318558, + 0.02262643352150917, + 0.015175272710621357, + 0.015947721898555756, + 0.022087469696998596, + 0.016999566927552223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.006738675758242607 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, shrouded in the warm, yellowish hues of old films, transports us to a dimly lit ancient military tent. The sole illumination comes from a flickering oil lamp on a rough wooden table. The dim light casts large, oscillating shadows on the heavy canvas walls, creating a stark contrast between light and dark. Clad in a suit of dark, scratched, and weathered iron armor, Cao Cao stands with his back to the camera, slightly bent over a large parchment military map spread out on the table. Mountains, rivers, and military markers are outlined in ink. His long, tied-up hair is streaked with silver strands, and the shoulders of his armor reflect the faint glow of the oil lamp. He extends his armored hand, his finger hovering for a moment over a city marked on the map. Then, his shoulders visibly slump, and his body rises and falls with a heavy, drawn-out sigh. A puff of white air escapes his lips, quickly dissipating in the cold air. His furrowed brow and tense jawline appear particularly clear in the light and shadow of his profile, and the whole scene is filled with the heavy pressure before making a strategic decision and the lonely atmosphere of a hero in his twilight years.", + "seed": 32875, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_11/index_1871_seed_32875_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03181615807116032, + "vlm_logprob": -0.03181615807116032, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.03181615807116032, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient military tent with heavy canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Cao Cao visible wearing dark, scratched, and weathered iron armor with long hair streaked with silver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large parchment military map with ink outlines of mountains, rivers, and markers spread on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character extend an armored hand to hover a finger over a city on the map before slumping his shoulders and sighing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single flickering oil lamp casting large, oscillating shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007718146312981844 + ], + "opposite_logprob_tokens": [ + -5.507718086242676 + ], + "logprob_reward": -0.007718146312981844, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.007718146312981844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2107815146446228, + 0.11974790692329407, + 0.0539952777326107, + 0.0853843241930008, + 0.047526147216558456, + 0.0898943766951561, + 0.03712362051010132, + 0.06641434878110886, + 0.15683168172836304, + 0.13230079412460327 + ], + "credit_map": [ + 0.032160330563783646, + 0.02058788202702999, + 0.04472291097044945, + 0.038870878517627716, + 0.014813494868576527, + 0.011208665557205677, + 0.025990597903728485, + 0.022426750510931015, + 0.012829442508518696, + 0.011826462112367153, + 0.020309457555413246, + 0.016071274876594543, + 0.011526722460985184, + 0.01338525116443634, + 0.019101358950138092, + 0.014697941951453686, + 0.007272528950124979, + 0.005302706733345985, + 0.00877793412655592, + 0.007070148829370737, + 0.00606208760291338, + 0.0044967420399188995, + 0.008356474339962006, + 0.0066566551104187965, + 0.011717267334461212, + 0.01008597668260336, + 0.017250994220376015, + 0.009129398502409458, + 0.0077559612691402435, + 0.008518077433109283, + 0.011016848497092724, + 0.009909802116453648, + 0.004789524711668491, + 0.004423796199262142, + 0.007248318754136562, + 0.008040179498493671, + 0.004208993632346392, + 0.0036394272465258837, + 0.009845692664384842, + 0.005330218933522701, + 0.015767639502882957, + 0.008430690504610538, + 0.022914644330739975, + 0.012836463749408722, + 0.005277309566736221, + 0.005726406816393137, + 0.008767751045525074, + 0.010173470713198185, + 0.004016927909106016, + 0.0031167548149824142, + 0.007845774292945862, + 0.005543183069676161, + 0.00407544756308198, + 0.0027955009136348963, + 0.004252529237419367, + 0.005477504804730415, + 0.009503265842795372, + 0.007287554442882538, + 0.010427485220134258, + 0.010223877616226673, + 0.007106801960617304, + 0.00423851702362299, + 0.008963463827967644, + 0.008663381449878216, + 0.030675344169139862, + 0.014642543159425259, + 0.0298849456012249, + 0.01934191957116127, + 0.014370463788509369, + 0.014464174397289753, + 0.01287741120904684, + 0.020574884489178658, + 0.012622284702956676, + 0.012789829634130001, + 0.02280569262802601, + 0.017041578888893127, + 0.018996233120560646, + 0.014487591572105885, + 0.016903413459658623, + 0.01665416546165943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.007718146312981844 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient military tent with heavy canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Cao Cao visible wearing dark, scratched, and weathered iron armor with long hair streaked with silver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large parchment military map with ink outlines of mountains, rivers, and markers spread on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character extend an armored hand to hover a finger over a city on the map before slumping his shoulders and sighing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single flickering oil lamp casting large, oscillating shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060834877192974 + ], + "opposite_logprob_tokens": [ + -3.895608425140381 + ], + "logprob_reward": -0.02060834877192974, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060834877192974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2777058780193329, + 0.07154888659715652, + 0.05985230207443237, + 0.07846517860889435, + 0.03374636173248291, + 0.061260730028152466, + 0.04536573961377144, + 0.09330007433891296, + 0.1375148743391037, + 0.1412399709224701 + ], + "credit_map": [ + 0.0423009917140007, + 0.02733769454061985, + 0.05973384156823158, + 0.04474690929055214, + 0.02024390548467636, + 0.015111747197806835, + 0.03662870079278946, + 0.03160204738378525, + 0.008154402486979961, + 0.00861375592648983, + 0.010131233371794224, + 0.008822507224977016, + 0.007179865147918463, + 0.005085458979010582, + 0.014591076411306858, + 0.008970582857728004, + 0.00570069020614028, + 0.006353507284075022, + 0.00864800252020359, + 0.007237258832901716, + 0.008466946892440319, + 0.006254824809730053, + 0.010624428279697895, + 0.00656663766130805, + 0.01251937448978424, + 0.006174473091959953, + 0.015051432885229588, + 0.010064429603517056, + 0.008166533894836903, + 0.0061554270796477795, + 0.01064547710120678, + 0.009688018821179867, + 0.00394693948328495, + 0.0029099988751113415, + 0.0053084599785506725, + 0.0038386345840990543, + 0.004430577624589205, + 0.0025921983178704977, + 0.006444296799600124, + 0.004275251645594835, + 0.00957063864916563, + 0.005808079615235329, + 0.014623232185840607, + 0.007485347334295511, + 0.005819158628582954, + 0.0040823626331985, + 0.006923781707882881, + 0.006948120426386595, + 0.006069737486541271, + 0.004163813311606646, + 0.005015290342271328, + 0.005559039767831564, + 0.008320163935422897, + 0.0037301266565918922, + 0.007216503378003836, + 0.0052910614758729935, + 0.012089251540601254, + 0.008902644738554955, + 0.011399752460420132, + 0.015475659631192684, + 0.01344649214297533, + 0.00583680160343647, + 0.014081122353672981, + 0.012068341486155987, + 0.030416494235396385, + 0.012985575012862682, + 0.025119898840785027, + 0.015546012669801712, + 0.013520893640816212, + 0.010389667935669422, + 0.011268794536590576, + 0.01826752908527851, + 0.013928014785051346, + 0.010959991253912449, + 0.014330469071865082, + 0.01760697364807129, + 0.026637960225343704, + 0.019637588411569595, + 0.02117886021733284, + 0.016960108652710915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.02060834877192974 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient military tent with heavy canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Cao Cao visible wearing dark, scratched, and weathered iron armor with long hair streaked with silver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large parchment military map with ink outlines of mountains, rivers, and markers spread on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character extend an armored hand to hover a finger over a city on the map before slumping his shoulders and sighing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single flickering oil lamp casting large, oscillating shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005261146929115057 + ], + "opposite_logprob_tokens": [ + -5.255260944366455 + ], + "logprob_reward": -0.005261146929115057, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005261146929115057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12206327170133591, + 0.06279290467500687, + 0.06304784119129181, + 0.1161331832408905, + 0.09535594284534454, + 0.050661440938711166, + 0.13944295048713684, + 0.07031600922346115, + 0.17920814454555511, + 0.1009782925248146 + ], + "credit_map": [ + 0.013164343312382698, + 0.011695225723087788, + 0.025113768875598907, + 0.021176353096961975, + 0.012520631775259972, + 0.00970285665243864, + 0.015191368758678436, + 0.013498720712959766, + 0.007642729673534632, + 0.007278752978891134, + 0.010408156551420689, + 0.01601342484354973, + 0.005082623567432165, + 0.0034435142297297716, + 0.006258086767047644, + 0.006665614899247885, + 0.008195953443646431, + 0.006563100032508373, + 0.011543425731360912, + 0.01334451511502266, + 0.004828922916203737, + 0.0037135258316993713, + 0.008024225942790508, + 0.006834172178059816, + 0.017863357439637184, + 0.016356075182557106, + 0.020345645025372505, + 0.023243898525834084, + 0.00977794174104929, + 0.0063657863065600395, + 0.010501394979655743, + 0.011679084971547127, + 0.014553586952388287, + 0.015918858349323273, + 0.01362970657646656, + 0.019081896170973778, + 0.007780933286994696, + 0.005567836109548807, + 0.009385015815496445, + 0.009438109584152699, + 0.008260857313871384, + 0.004881633445620537, + 0.010103344917297363, + 0.005294980015605688, + 0.00471246475353837, + 0.0038388785906136036, + 0.00789437722414732, + 0.005674905143678188, + 0.02560352347791195, + 0.01278926432132721, + 0.027372609823942184, + 0.02202649787068367, + 0.011305325664579868, + 0.01156936027109623, + 0.0131659721955657, + 0.015610406175255775, + 0.00807888526469469, + 0.00541190430521965, + 0.007155636791139841, + 0.009344144724309444, + 0.01046251691877842, + 0.00643681176006794, + 0.014425658620893955, + 0.0090004438534379, + 0.03919174149632454, + 0.017176151275634766, + 0.03405165672302246, + 0.02235918678343296, + 0.01433529146015644, + 0.014976941049098969, + 0.014098827727138996, + 0.023018350824713707, + 0.01046779751777649, + 0.00842199381440878, + 0.010306068696081638, + 0.012765422463417053, + 0.01921556331217289, + 0.012300251983106136, + 0.014975502155721188, + 0.01252569817006588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.005261146929115057 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient military tent with heavy canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Cao Cao visible wearing dark, scratched, and weathered iron armor with long hair streaked with silver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large parchment military map with ink outlines of mountains, rivers, and markers spread on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character extend an armored hand to hover a finger over a city on the map before slumping his shoulders and sighing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single flickering oil lamp casting large, oscillating shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295108497142792 + ], + "opposite_logprob_tokens": [ + -2.2379510402679443 + ], + "logprob_reward": -0.11295108497142792, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295108497142792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18774808943271637, + 0.05043058469891548, + 0.03811495751142502, + 0.04314659163355827, + 0.03825494274497032, + 0.05001552030444145, + 0.041234634816646576, + 0.0868794322013855, + 0.2620646059513092, + 0.2021106630563736 + ], + "credit_map": [ + 0.022007595747709274, + 0.01929185912013054, + 0.035966187715530396, + 0.025898853316903114, + 0.025701789185404778, + 0.015269128605723381, + 0.02523660659790039, + 0.018376069143414497, + 0.00489138625562191, + 0.005094082094728947, + 0.008261427283287048, + 0.007972021587193012, + 0.005146392155438662, + 0.005005148705095053, + 0.008296514861285686, + 0.005763613618910313, + 0.005676558706909418, + 0.003971954341977835, + 0.006011009216308594, + 0.004285195842385292, + 0.005116881802678108, + 0.0037576055619865656, + 0.005188098177313805, + 0.004107653629034758, + 0.007671352941542864, + 0.004110970068722963, + 0.0060053253546357155, + 0.0079987533390522, + 0.004392306320369244, + 0.003089660545811057, + 0.005131041631102562, + 0.004747180268168449, + 0.004042936488986015, + 0.00274568609893322, + 0.008212544955313206, + 0.00605577090755105, + 0.004493124317377806, + 0.0027194861322641373, + 0.005730647128075361, + 0.004254746250808239, + 0.005028523039072752, + 0.003908590413630009, + 0.0077913785353302956, + 0.006021668668836355, + 0.006456858478486538, + 0.005298084579408169, + 0.009578549303114414, + 0.005931870546191931, + 0.004945016000419855, + 0.0036786021664738655, + 0.0071470350958406925, + 0.006547385826706886, + 0.006351614836603403, + 0.0035646220203489065, + 0.004620003513991833, + 0.004380356054753065, + 0.008899876847863197, + 0.006621719803661108, + 0.011723022907972336, + 0.009973720647394657, + 0.013344820588827133, + 0.009176836349070072, + 0.01609840616583824, + 0.011041035875678062, + 0.052480172365903854, + 0.024691151455044746, + 0.046368688344955444, + 0.03362344577908516, + 0.028504429385066032, + 0.0205110851675272, + 0.02045213058590889, + 0.035433489829301834, + 0.01880459301173687, + 0.01540865283459425, + 0.020319484174251556, + 0.025196539238095284, + 0.038314878940582275, + 0.027481885626912117, + 0.031098604202270508, + 0.02548602595925331 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.11295108497142792 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient military tent with heavy canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Cao Cao visible wearing dark, scratched, and weathered iron armor with long hair streaked with silver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large parchment military map with ink outlines of mountains, rivers, and markers spread on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character extend an armored hand to hover a finger over a city on the map before slumping his shoulders and sighing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single flickering oil lamp casting large, oscillating shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012542063370347023 + ], + "opposite_logprob_tokens": [ + -4.387542247772217 + ], + "logprob_reward": -0.012542063370347023, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012542063370347023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27067986130714417, + 0.0465112142264843, + 0.047386229038238525, + 0.02121265046298504, + 0.029499005526304245, + 0.03057127632200718, + 0.050566259771585464, + 0.08881025016307831, + 0.28998249769210815, + 0.12478069216012955 + ], + "credit_map": [ + 0.032635290175676346, + 0.029555361717939377, + 0.05442852899432182, + 0.0429513119161129, + 0.020297467708587646, + 0.02907704748213291, + 0.029960354790091515, + 0.031774528324604034, + 0.0053108022548258305, + 0.00448672566562891, + 0.006180643569678068, + 0.004967684391885996, + 0.005195645149797201, + 0.004497861955314875, + 0.010121875442564487, + 0.005749978590756655, + 0.004792487248778343, + 0.005013750866055489, + 0.00753318564966321, + 0.007988249883055687, + 0.005423014983534813, + 0.003374712774530053, + 0.007383326534181833, + 0.005877501331269741, + 0.003895385190844536, + 0.0027036007959395647, + 0.0032533532939851284, + 0.003040354698896408, + 0.0017468815203756094, + 0.0015070255612954497, + 0.002717593451961875, + 0.0023484581615775824, + 0.003841381287202239, + 0.0030238532926887274, + 0.00454747537150979, + 0.0031783650629222393, + 0.0035601744893938303, + 0.003352635307237506, + 0.004560847766697407, + 0.0034342731814831495, + 0.0046858759596943855, + 0.0030765081755816936, + 0.006240282207727432, + 0.0036741802468895912, + 0.003448860254138708, + 0.002033692318946123, + 0.0035054085310548544, + 0.003906469792127609, + 0.004928583279252052, + 0.0036236755549907684, + 0.006638315040618181, + 0.0077576953917741776, + 0.008339821361005306, + 0.0049064490012824535, + 0.008049560710787773, + 0.006322165485471487, + 0.006746513303369284, + 0.00717560388147831, + 0.012475758790969849, + 0.011703470721840858, + 0.014282150194048882, + 0.009860583581030369, + 0.015323583036661148, + 0.011242596432566643, + 0.042683135718107224, + 0.027887506410479546, + 0.056576814502477646, + 0.04640543833374977, + 0.02924378588795662, + 0.02388579584658146, + 0.02459164708852768, + 0.0387083999812603, + 0.01323460228741169, + 0.015175662003457546, + 0.02472865581512451, + 0.02416762337088585, + 0.00994846411049366, + 0.011341024190187454, + 0.012901204638183117, + 0.013283465057611465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.012542063370347023 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, shrouded in the warm, yellowish hues of old films, transports us to a dimly lit ancient military tent. The sole illumination comes from a flickering oil lamp on a rough wooden table. The dim light casts large, oscillating shadows on the heavy canvas walls, creating a stark contrast between light and dark. Clad in a suit of dark, scratched, and weathered iron armor, Cao Cao stands with his back to the camera, slightly bent over a large parchment military map spread out on the table. Mountains, rivers, and military markers are outlined in ink. His long, tied-up hair is streaked with silver strands, and the shoulders of his armor reflect the faint glow of the oil lamp. He extends his armored hand, his finger hovering for a moment over a city marked on the map. Then, his shoulders visibly slump, and his body rises and falls with a heavy, drawn-out sigh. A puff of white air escapes his lips, quickly dissipating in the cold air. His furrowed brow and tense jawline appear particularly clear in the light and shadow of his profile, and the whole scene is filled with the heavy pressure before making a strategic decision and the lonely atmosphere of a hero in his twilight years.", + "seed": 32876, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_11/index_1871_seed_32876_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.021688776649534704, + "vlm_logprob": -0.021688776649534704, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.021688776649534704, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient military tent with heavy canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Cao Cao visible wearing dark, scratched, and weathered iron armor with long hair streaked with silver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large parchment military map with ink outlines of mountains, rivers, and markers spread on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character extend an armored hand to hover a finger over a city on the map before slumping his shoulders and sighing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single flickering oil lamp casting large, oscillating shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007662546820938587 + ], + "opposite_logprob_tokens": [ + -5.132662773132324 + ], + "logprob_reward": -0.007662546820938587, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.007662546820938587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1561579704284668, + 0.06678900122642517, + 0.2062595635652542, + 0.06983719766139984, + 0.13881906867027283, + 0.03716810792684555, + 0.03959554061293602, + 0.08096134662628174, + 0.09537307918071747, + 0.10903912782669067 + ], + "credit_map": [ + 0.022116754204034805, + 0.018397459760308266, + 0.019715463742613792, + 0.01853467896580696, + 0.024885522201657295, + 0.020186468958854675, + 0.014734547585248947, + 0.01758706569671631, + 0.007203174289315939, + 0.005669800564646721, + 0.005364810582250357, + 0.005405775737017393, + 0.009692481718957424, + 0.009679357521235943, + 0.008305936120450497, + 0.015467661432921886, + 0.03767839074134827, + 0.017968859523534775, + 0.01401455793529749, + 0.024592071771621704, + 0.020962342619895935, + 0.02838079258799553, + 0.02687334455549717, + 0.03578920662403107, + 0.014431647025048733, + 0.005641371011734009, + 0.006433263421058655, + 0.007175184320658445, + 0.011245147325098515, + 0.009795927442610264, + 0.006812775507569313, + 0.008301875554025173, + 0.02237074077129364, + 0.01068408228456974, + 0.01232857909053564, + 0.01480732299387455, + 0.017677268013358116, + 0.026968713849782944, + 0.013274049386382103, + 0.02070830948650837, + 0.004873904399573803, + 0.004172138869762421, + 0.0030905036255717278, + 0.0033075418323278427, + 0.0056368387304246426, + 0.0054196203127503395, + 0.0064778090454638, + 0.004189751110970974, + 0.004473814275115728, + 0.004178423434495926, + 0.0039467578753829, + 0.004191401414573193, + 0.00471530994400382, + 0.008483145385980606, + 0.0033882400020956993, + 0.006218448281288147, + 0.015177886001765728, + 0.007379206828773022, + 0.008562766015529633, + 0.007445690222084522, + 0.008292462676763535, + 0.011210750788450241, + 0.012302204966545105, + 0.010590380057692528, + 0.008936298079788685, + 0.00505896145477891, + 0.00908730085939169, + 0.005921044386923313, + 0.02715659700334072, + 0.01386399194598198, + 0.008310742676258087, + 0.017038146033883095, + 0.015434006229043007, + 0.009661479853093624, + 0.008965072222054005, + 0.013166728429496288, + 0.021063953638076782, + 0.016572140157222748, + 0.01497839204967022, + 0.009197355248034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.007662546820938587 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient military tent with heavy canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Cao Cao visible wearing dark, scratched, and weathered iron armor with long hair streaked with silver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large parchment military map with ink outlines of mountains, rivers, and markers spread on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character extend an armored hand to hover a finger over a city on the map before slumping his shoulders and sighing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single flickering oil lamp casting large, oscillating shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023311134427785873 + ], + "opposite_logprob_tokens": [ + -3.773311138153076 + ], + "logprob_reward": -0.023311134427785873, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023311134427785873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2556036114692688, + 0.08641023188829422, + 0.0652080550789833, + 0.06529685109853745, + 0.1525607407093048, + 0.09598654508590698, + 0.04205887019634247, + 0.05573004484176636, + 0.10977104306221008, + 0.07137397676706314 + ], + "credit_map": [ + 0.02767288312315941, + 0.030833454802632332, + 0.027048997581005096, + 0.028225328773260117, + 0.04311228170990944, + 0.030400902032852173, + 0.029997555539011955, + 0.03831222280859947, + 0.01471997145563364, + 0.009039696305990219, + 0.0072519006207585335, + 0.00968275684863329, + 0.01317355316132307, + 0.01105500664561987, + 0.009909648448228836, + 0.011577701196074486, + 0.011574742384254932, + 0.004067142494022846, + 0.005678901448845863, + 0.009290939196944237, + 0.006789357867091894, + 0.008573610335588455, + 0.006967772729694843, + 0.012265587225556374, + 0.012047563679516315, + 0.004769646096974611, + 0.006065275054425001, + 0.00846913829445839, + 0.008391238749027252, + 0.010197230614721775, + 0.0071587818674743176, + 0.00819797720760107, + 0.026876840740442276, + 0.010378658771514893, + 0.010024002753198147, + 0.01265473011881113, + 0.018552128225564957, + 0.023233775049448013, + 0.012877085246145725, + 0.03796352073550224, + 0.017271675169467926, + 0.007668665610253811, + 0.008925646543502808, + 0.007838374003767967, + 0.015392896719276905, + 0.01289009302854538, + 0.013807937502861023, + 0.012191258370876312, + 0.005987269803881645, + 0.004205205012112856, + 0.006644352804869413, + 0.0042227511294186115, + 0.0039284722879529, + 0.005206504371017218, + 0.006283869501203299, + 0.005580444820225239, + 0.008637013845145702, + 0.006087017711251974, + 0.00749646220356226, + 0.005261311307549477, + 0.007449531927704811, + 0.007863419130444527, + 0.006516606546938419, + 0.006418684031814337, + 0.007116205524653196, + 0.006358770187944174, + 0.010291560553014278, + 0.007403085008263588, + 0.032793689519166946, + 0.020384671166539192, + 0.010837296955287457, + 0.014585757628083229, + 0.00873766653239727, + 0.004384840838611126, + 0.006245318800210953, + 0.006062929518520832, + 0.020215872675180435, + 0.011001795530319214, + 0.008593142032623291, + 0.006132403388619423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.023311134427785873 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient military tent with heavy canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Cao Cao visible wearing dark, scratched, and weathered iron armor with long hair streaked with silver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large parchment military map with ink outlines of mountains, rivers, and markers spread on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character extend an armored hand to hover a finger over a city on the map before slumping his shoulders and sighing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single flickering oil lamp casting large, oscillating shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022025867365300655 + ], + "opposite_logprob_tokens": [ + -6.12720251083374 + ], + "logprob_reward": -0.0022025867365300655, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022025867365300655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1709437221288681, + 0.06956620514392853, + 0.15970644354820251, + 0.1443033069372177, + 0.07164868712425232, + 0.023552197962999344, + 0.07608162611722946, + 0.09724684804677963, + 0.09468097984790802, + 0.09226994961500168 + ], + "credit_map": [ + 0.018863044679164886, + 0.01715843938291073, + 0.03212785720825195, + 0.027810852974653244, + 0.027460232377052307, + 0.019090665504336357, + 0.010286634787917137, + 0.01814599707722664, + 0.007606848143041134, + 0.010477609932422638, + 0.006224532146006823, + 0.009175824001431465, + 0.007723755203187466, + 0.00841510109603405, + 0.007017035502940416, + 0.012925504706799984, + 0.025462646037340164, + 0.011358603835105896, + 0.010790841653943062, + 0.02228756807744503, + 0.016445342451334, + 0.02360250987112522, + 0.017377221956849098, + 0.03238170966506004, + 0.023447489365935326, + 0.011462043970823288, + 0.013022704981267452, + 0.013322973623871803, + 0.02720395103096962, + 0.02032356522977352, + 0.014899552799761295, + 0.02062102220952511, + 0.015475318767130375, + 0.00587461655959487, + 0.009157338179647923, + 0.007461518980562687, + 0.008910649456083775, + 0.010008208453655243, + 0.005824336316436529, + 0.008936703205108643, + 0.003908910788595676, + 0.0026330226100981236, + 0.0030278610065579414, + 0.0024835076183080673, + 0.003330641658976674, + 0.0031521874479949474, + 0.002289906609803438, + 0.0027261609211564064, + 0.013154980726540089, + 0.007320469245314598, + 0.009822281077504158, + 0.01026260107755661, + 0.008030906319618225, + 0.009061881341040134, + 0.008536693640053272, + 0.009891812689602375, + 0.01286114938557148, + 0.0081082284450531, + 0.009979335591197014, + 0.0076746209524571896, + 0.014390148222446442, + 0.015733642503619194, + 0.015387127175927162, + 0.013112596236169338, + 0.014255012385547161, + 0.011118472553789616, + 0.010954363271594048, + 0.008784780278801918, + 0.014054689556360245, + 0.014551880769431591, + 0.00810164026916027, + 0.012860139831900597, + 0.020515207201242447, + 0.007645444478839636, + 0.008910699747502804, + 0.009253299795091152, + 0.012388359755277634, + 0.014480850659310818, + 0.011272573843598366, + 0.007803511805832386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.0022025867365300655 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient military tent with heavy canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Cao Cao visible wearing dark, scratched, and weathered iron armor with long hair streaked with silver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large parchment military map with ink outlines of mountains, rivers, and markers spread on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character extend an armored hand to hover a finger over a city on the map before slumping his shoulders and sighing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single flickering oil lamp casting large, oscillating shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001370191574097 + ], + "opposite_logprob_tokens": [ + -2.6950137615203857 + ], + "logprob_reward": -0.07001370191574097, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001370191574097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13872280716896057, + 0.1192268654704094, + 0.07999075204133987, + 0.04530046507716179, + 0.06941671669483185, + 0.030294042080640793, + 0.153972789645195, + 0.13462284207344055, + 0.16870516538619995, + 0.05974755063652992 + ], + "credit_map": [ + 0.01640845462679863, + 0.01847176067531109, + 0.019939862191677094, + 0.01957826316356659, + 0.018873650580644608, + 0.016408532857894897, + 0.010536322370171547, + 0.018505945801734924, + 0.02316378429532051, + 0.011983790434896946, + 0.014282887801527977, + 0.010618339292705059, + 0.016162527725100517, + 0.016235757619142532, + 0.01034528948366642, + 0.016434473916888237, + 0.01108445692807436, + 0.005020169075578451, + 0.00756674911826849, + 0.01583619974553585, + 0.007259579841047525, + 0.010398220270872116, + 0.01082579419016838, + 0.011999581940472126, + 0.008180790580809116, + 0.004035536199808121, + 0.00447873305529356, + 0.0048102387227118015, + 0.0064302776008844376, + 0.00618616072461009, + 0.005434761755168438, + 0.005743962712585926, + 0.009266612119972706, + 0.006483058910816908, + 0.007152251433581114, + 0.0096952635794878, + 0.010097192600369453, + 0.00989279244095087, + 0.006680008489638567, + 0.010149531997740269, + 0.005222085863351822, + 0.002656554104760289, + 0.0037086582742631435, + 0.0025452265981584787, + 0.00495346263051033, + 0.004211022052913904, + 0.0027965670451521873, + 0.004200462251901627, + 0.03579065203666687, + 0.01560207549482584, + 0.017636779695749283, + 0.01340287271887064, + 0.015737788751721382, + 0.020205402746796608, + 0.014267885126173496, + 0.021329332143068314, + 0.03549080714583397, + 0.013828262686729431, + 0.016801433637738228, + 0.012650402262806892, + 0.013236332684755325, + 0.014704041182994843, + 0.011848489753901958, + 0.016063060611486435, + 0.010131444782018661, + 0.006720779463648796, + 0.01601342111825943, + 0.0066427066922187805, + 0.0608239471912384, + 0.02821291610598564, + 0.009815450757741928, + 0.03034449741244316, + 0.008623307570815086, + 0.0038436080794781446, + 0.005059308372437954, + 0.006429471541196108, + 0.008212340995669365, + 0.007898330688476562, + 0.011860077269375324, + 0.007821100763976574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.07001370191574097 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient military tent with heavy canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Cao Cao visible wearing dark, scratched, and weathered iron armor with long hair streaked with silver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large parchment military map with ink outlines of mountains, rivers, and markers spread on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character extend an armored hand to hover a finger over a city on the map before slumping his shoulders and sighing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single flickering oil lamp casting large, oscillating shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005253913346678019 + ], + "opposite_logprob_tokens": [ + -5.255253791809082 + ], + "logprob_reward": -0.005253913346678019, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005253913346678019, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17237696051597595, + 0.0982990711927414, + 0.055913735181093216, + 0.07525131106376648, + 0.15402138233184814, + 0.04702029749751091, + 0.06584780663251877, + 0.11792875826358795, + 0.07740574330091476, + 0.13593487441539764 + ], + "credit_map": [ + 0.027385903522372246, + 0.024134360253810883, + 0.017826037481427193, + 0.022980360314249992, + 0.022829866036772728, + 0.02151891402900219, + 0.015011638402938843, + 0.020689886063337326, + 0.015355344861745834, + 0.011751004494726658, + 0.010471387766301632, + 0.01043534092605114, + 0.014922910369932652, + 0.013242673128843307, + 0.006942005828022957, + 0.01517841499298811, + 0.007843051105737686, + 0.004802926443517208, + 0.004080034326761961, + 0.005738161038607359, + 0.006426765583455563, + 0.007026147563010454, + 0.01213985588401556, + 0.007856801152229309, + 0.009100020863115788, + 0.004804295487701893, + 0.0061012692749500275, + 0.006021943408995867, + 0.014868284575641155, + 0.010916718281805515, + 0.009643670171499252, + 0.013795116916298866, + 0.021298043429851532, + 0.012783564627170563, + 0.011231090873479843, + 0.018412942066788673, + 0.020962100476026535, + 0.028068212792277336, + 0.013723942451179028, + 0.027541499584913254, + 0.007275938987731934, + 0.004363930318504572, + 0.004480396863073111, + 0.005146475043147802, + 0.006269143894314766, + 0.007158445660024881, + 0.007184552028775215, + 0.0051414212211966515, + 0.007815866731107235, + 0.008038618601858616, + 0.005758540704846382, + 0.005321052391082048, + 0.008126901462674141, + 0.011996755376458168, + 0.008439038880169392, + 0.010351046919822693, + 0.01231563463807106, + 0.010773838497698307, + 0.012549974955618382, + 0.010817181318998337, + 0.015095661394298077, + 0.017378555610775948, + 0.022294295951724052, + 0.01670362986624241, + 0.005176799837499857, + 0.0033410617616027594, + 0.003971964120864868, + 0.003666189033538103, + 0.026403438299894333, + 0.011156019754707813, + 0.006014890503138304, + 0.01767539046704769, + 0.018419519066810608, + 0.008268050849437714, + 0.011213715188205242, + 0.010369374416768551, + 0.0330386757850647, + 0.016110055148601532, + 0.0251332838088274, + 0.013382203876972198 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.005253913346678019 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_12/index_tensor([6934, 6934])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_12/index_tensor([6934, 6934])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8b7739c63af27a8891b37d1bcbf4a2fd579acc43 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_12/index_tensor([6934, 6934])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle tracking shot glides swiftly above the ground, keeping pace with a small, sleek quadcopter drone as it navigates a vast, sun-drenched field of wildflowers. Constructed of matte black carbon fiber, the drone is equipped with four translucent propellers that blur into near-invisibility as it spins, and a soft cyan LED glows from its front. It gallops merrily, agilely executing sharp, banked turns, effortlessly darting through dense fields of red poppies, white daisies, and purple lavender, all swaying in the gentle afternoon breeze. As the camera follows, the drone momentarily halts its forward momentum, delicately hovering just inches above a vibrant yellow flower, tilting its chassis downward like a playful bow before quickly rising and resuming its playful dance. The entire scene is bathed in the warm, golden afternoon light, casting long, soft shadows that make the delicate petals appear almost translucent, creating a tranquil and wondrous atmosphere that blends technological elegance with the beauty of nature.", + "seed": 26130, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_12/index_6934_seed_26130_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.040018240362405776, + "vlm_logprob": -0.040018240362405776, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.040018240362405776, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, sleek quadcopter drone constructed of matte black carbon fiber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drone have four translucent propellers and a soft cyan LED glowing from its front?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a vast, sun-drenched field containing red poppies, white daisies, and purple lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone execute sharp, banked turns and momentarily hover just inches above a vibrant yellow flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle tracking shot that glides swiftly above the ground following the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06720158457756042 + ], + "opposite_logprob_tokens": [ + -2.817201614379883 + ], + "logprob_reward": -0.06720158457756042, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06720158457756042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05780230090022087, + 0.05741433426737785, + 0.052477508783340454, + 0.043214939534664154, + 0.18382006883621216, + 0.05203444883227348, + 0.03054063767194748, + 0.0331442728638649, + 0.026127347722649574, + 0.4634241759777069 + ], + "credit_map": [ + 0.004034900572150946, + 0.004772675223648548, + 0.01045913714915514, + 0.007699074689298868, + 0.006096578668802977, + 0.007671259809285402, + 0.006909755989909172, + 0.010158922523260117, + 0.005689470563083887, + 0.005402994342148304, + 0.007160444278270006, + 0.005387197248637676, + 0.006452229805290699, + 0.008114269934594631, + 0.008098683319985867, + 0.011109055019915104, + 0.005172256380319595, + 0.006874303333461285, + 0.006992213893681765, + 0.007515777368098497, + 0.0059692212380468845, + 0.005608151201158762, + 0.006691778544336557, + 0.007653810083866119, + 0.0036075126845389605, + 0.002519693225622177, + 0.011294958181679249, + 0.0054657114669680595, + 0.004114422015845776, + 0.005407233256846666, + 0.004572349600493908, + 0.006233059801161289, + 0.017891965806484222, + 0.015046710148453712, + 0.018045328557491302, + 0.017614103853702545, + 0.03273797035217285, + 0.025420716032385826, + 0.027164384722709656, + 0.029898906126618385, + 0.005867717321962118, + 0.004521308932453394, + 0.006386975292116404, + 0.010498656891286373, + 0.00400632806122303, + 0.007861536927521229, + 0.006279442925006151, + 0.006612484809011221, + 0.0019458199385553598, + 0.004357776138931513, + 0.003737520892173052, + 0.004471985157579184, + 0.003283493220806122, + 0.004811229184269905, + 0.003711770288646221, + 0.004221045877784491, + 0.0038031423464417458, + 0.00399006437510252, + 0.004324267152696848, + 0.004590880591422319, + 0.0015812446363270283, + 0.003838386619463563, + 0.004271504003554583, + 0.006744785234332085, + 0.0018080845475196838, + 0.0017647029599174857, + 0.0029252241365611553, + 0.0026793708093464375, + 0.0037595087196677923, + 0.005370568484067917, + 0.0021170820109546185, + 0.00570280896499753, + 0.020302528515458107, + 0.012985875830054283, + 0.021220510825514793, + 0.026972264051437378, + 0.03220793232321739, + 0.14552229642868042, + 0.07205413281917572, + 0.13215865194797516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.06720158457756042 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, sleek quadcopter drone constructed of matte black carbon fiber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drone have four translucent propellers and a soft cyan LED glowing from its front?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a vast, sun-drenched field containing red poppies, white daisies, and purple lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone execute sharp, banked turns and momentarily hover just inches above a vibrant yellow flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle tracking shot that glides swiftly above the ground following the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029813479632139206 + ], + "opposite_logprob_tokens": [ + -3.529813528060913 + ], + "logprob_reward": -0.029813479632139206, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029813479632139206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055431194603443146, + 0.05389207974076271, + 0.06523709744215012, + 0.03623972833156586, + 0.3714514374732971, + 0.03103751689195633, + 0.03127043694257736, + 0.031802207231521606, + 0.021628983318805695, + 0.3020092844963074 + ], + "credit_map": [ + 0.004345311783254147, + 0.0051263258792459965, + 0.007731984835118055, + 0.00831695832312107, + 0.006311631761491299, + 0.00724199740216136, + 0.00779410544782877, + 0.008562884293496609, + 0.003852487774565816, + 0.0049826898612082005, + 0.007694334723055363, + 0.005293098744004965, + 0.007209476549178362, + 0.008059957064688206, + 0.01132818590849638, + 0.005471853539347649, + 0.0071060992777347565, + 0.00726709421724081, + 0.00993409100919962, + 0.010091307573020458, + 0.006497477181255817, + 0.006722899619489908, + 0.008364537730813026, + 0.00925359595566988, + 0.003018078627064824, + 0.0036342323292046785, + 0.005442887544631958, + 0.0058721331879496574, + 0.00335069396533072, + 0.004900621250271797, + 0.005040960852056742, + 0.004980124998837709, + 0.03543343022465706, + 0.027628218755126, + 0.04907223582267761, + 0.04573654755949974, + 0.0510292686522007, + 0.051259882748126984, + 0.05080198124051094, + 0.06048990413546562, + 0.0031167673878371716, + 0.0021663920488208532, + 0.0028101475909352303, + 0.007057298440486193, + 0.0027379218954592943, + 0.004608028568327427, + 0.0033528816420584917, + 0.0051880828104913235, + 0.0034292624332010746, + 0.0034611239098012447, + 0.0034207168500870466, + 0.005877932067960501, + 0.003457374405115843, + 0.003723674453794956, + 0.004240776877850294, + 0.0036595789715647697, + 0.002927056048065424, + 0.004819852765649557, + 0.00321780308149755, + 0.004944389220327139, + 0.002799359383061528, + 0.004208316095173359, + 0.004123827908188105, + 0.004761602729558945, + 0.002241440350189805, + 0.0016211092006415129, + 0.0046619814820587635, + 0.0028261435218155384, + 0.002166305435821414, + 0.0027283048257231712, + 0.002513470593839884, + 0.0028702293056994677, + 0.01301229652017355, + 0.008904941380023956, + 0.015233518555760384, + 0.017331896349787712, + 0.019685693085193634, + 0.09562002122402191, + 0.04538785666227341, + 0.08683308959007263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 2 + ], + "selected_reward": -0.029813479632139206 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, sleek quadcopter drone constructed of matte black carbon fiber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drone have four translucent propellers and a soft cyan LED glowing from its front?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a vast, sun-drenched field containing red poppies, white daisies, and purple lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone execute sharp, banked turns and momentarily hover just inches above a vibrant yellow flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle tracking shot that glides swiftly above the ground following the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304787144064903 + ], + "opposite_logprob_tokens": [ + -3.1680479049682617 + ], + "logprob_reward": -0.04304787144064903, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304787144064903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1082330271601677, + 0.04291713237762451, + 0.08985453844070435, + 0.11294208467006683, + 0.07193140685558319, + 0.19293013215065002, + 0.04025755077600479, + 0.16175298392772675, + 0.07057765871286392, + 0.1086035966873169 + ], + "credit_map": [ + 0.007588409353047609, + 0.008539825677871704, + 0.013701695017516613, + 0.011713466607034206, + 0.010321133770048618, + 0.015956049785017967, + 0.01117470022290945, + 0.02923772670328617, + 0.004932916257530451, + 0.0030541655141860247, + 0.005664929747581482, + 0.0030318947974592447, + 0.0037575270980596542, + 0.006659443024545908, + 0.004826952237635851, + 0.010989299975335598, + 0.010048145428299904, + 0.009937996044754982, + 0.015419498085975647, + 0.010664217174053192, + 0.008904216811060905, + 0.010253782384097576, + 0.010020425543189049, + 0.014606243930757046, + 0.009563755244016647, + 0.00626905495300889, + 0.01831550896167755, + 0.010331453755497932, + 0.009840388782322407, + 0.015024684369564056, + 0.02091715857386589, + 0.022680066525936127, + 0.006226579193025827, + 0.005799939855933189, + 0.007626810111105442, + 0.006398810539394617, + 0.012169444002211094, + 0.00954482052475214, + 0.011239545419812202, + 0.012925444170832634, + 0.013977470807731152, + 0.009575903415679932, + 0.030172819271683693, + 0.06473109126091003, + 0.00655865203589201, + 0.017654718831181526, + 0.021953504532575607, + 0.028305936604738235, + 0.006832464132457972, + 0.004434911068528891, + 0.005110176280140877, + 0.0054899221286177635, + 0.003327458631247282, + 0.0048476336523890495, + 0.004473840817809105, + 0.005741136148571968, + 0.01681164652109146, + 0.019399015232920647, + 0.02035423181951046, + 0.0404793918132782, + 0.010540783405303955, + 0.019708460196852684, + 0.013820365071296692, + 0.020639073103666306, + 0.010962116532027721, + 0.004242707975208759, + 0.012869270518422127, + 0.007218545768409967, + 0.006941435858607292, + 0.01294642873108387, + 0.005342187359929085, + 0.010054947808384895, + 0.004681319929659367, + 0.0028816526755690575, + 0.0058665163815021515, + 0.005702896974980831, + 0.008611262775957584, + 0.03418901190161705, + 0.01536064874380827, + 0.03131027892231941 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 3 + ], + "selected_reward": -0.04304787144064903 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, sleek quadcopter drone constructed of matte black carbon fiber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drone have four translucent propellers and a soft cyan LED glowing from its front?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a vast, sun-drenched field containing red poppies, white daisies, and purple lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone execute sharp, banked turns and momentarily hover just inches above a vibrant yellow flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle tracking shot that glides swiftly above the ground following the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026344802230596542 + ], + "opposite_logprob_tokens": [ + -3.6513447761535645 + ], + "logprob_reward": -0.026344802230596542, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026344802230596542, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.041222717612981796, + 0.022740621119737625, + 0.09702546149492264, + 0.06172535941004753, + 0.3829162120819092, + 0.052699096500873566, + 0.05248020216822624, + 0.09970752149820328, + 0.03481760621070862, + 0.15466517210006714 + ], + "credit_map": [ + 0.002829526783898473, + 0.0037435288541018963, + 0.005616825073957443, + 0.005801267456263304, + 0.0037431290838867426, + 0.006540802773088217, + 0.0049553788267076015, + 0.007992257364094257, + 0.002724917372688651, + 0.002632182789966464, + 0.002955720294266939, + 0.0021500643342733383, + 0.002648895373567939, + 0.003100595436990261, + 0.0028938946779817343, + 0.0036343506071716547, + 0.009959094226360321, + 0.013944843783974648, + 0.012913025915622711, + 0.01321040652692318, + 0.010994217358529568, + 0.01030668057501316, + 0.012879575602710247, + 0.012817620299756527, + 0.005978097673505545, + 0.0049081784673035145, + 0.01341602485626936, + 0.0067248037084937096, + 0.00657371710985899, + 0.007128927391022444, + 0.011139088310301304, + 0.0058565204963088036, + 0.04210544750094414, + 0.0301306564360857, + 0.03957025706768036, + 0.03899035230278969, + 0.06630939245223999, + 0.054984599351882935, + 0.05259658396244049, + 0.058228954672813416, + 0.003924258518964052, + 0.004323857370764017, + 0.0081448620185256, + 0.007685024756938219, + 0.005041569471359253, + 0.00693444674834609, + 0.0089652631431818, + 0.007679816335439682, + 0.007856158539652824, + 0.005366719793528318, + 0.005674479994922876, + 0.007882310077548027, + 0.004999498836696148, + 0.0076207066886126995, + 0.006642259657382965, + 0.006438067182898521, + 0.011165918782353401, + 0.012927035801112652, + 0.012689423747360706, + 0.02136886492371559, + 0.007157115265727043, + 0.013538064435124397, + 0.007513773627579212, + 0.013347319327294827, + 0.002951574046164751, + 0.001909501850605011, + 0.0067352247424423695, + 0.00443933205679059, + 0.0028927805833518505, + 0.005748639814555645, + 0.003116386942565441, + 0.0070241657085716724, + 0.0071686143055558205, + 0.003624219913035631, + 0.0073529998771846294, + 0.007402798160910606, + 0.010515504516661167, + 0.05116837099194527, + 0.020626889541745186, + 0.04680577665567398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 7 + ], + "selected_reward": -0.026344802230596542 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, sleek quadcopter drone constructed of matte black carbon fiber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drone have four translucent propellers and a soft cyan LED glowing from its front?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a vast, sun-drenched field containing red poppies, white daisies, and purple lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone execute sharp, banked turns and momentarily hover just inches above a vibrant yellow flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle tracking shot that glides swiftly above the ground following the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368346393108368 + ], + "opposite_logprob_tokens": [ + -3.4086835384368896 + ], + "logprob_reward": -0.03368346393108368, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368346393108368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12331871688365936, + 0.0381399430334568, + 0.05759476497769356, + 0.04311266541481018, + 0.24803295731544495, + 0.07939690351486206, + 0.05405895784497261, + 0.03949686139822006, + 0.050048958510160446, + 0.2667992115020752 + ], + "credit_map": [ + 0.009630117565393448, + 0.011403698474168777, + 0.019017444923520088, + 0.01541513204574585, + 0.012234780006110668, + 0.0237114354968071, + 0.015417158603668213, + 0.01648896001279354, + 0.003812280250713229, + 0.004028151743113995, + 0.006310680415481329, + 0.0039009018801152706, + 0.004015219397842884, + 0.00572932930663228, + 0.00573720782995224, + 0.004606171511113644, + 0.005917160306125879, + 0.005369921214878559, + 0.010210759937763214, + 0.009058322757482529, + 0.0051302481442689896, + 0.0070249405689537525, + 0.0062749446369707584, + 0.008608470670878887, + 0.004169365391135216, + 0.0037209538277238607, + 0.006619666237384081, + 0.004371311515569687, + 0.005289525259286165, + 0.005178981460630894, + 0.008798841387033463, + 0.004964022897183895, + 0.02564282715320587, + 0.021284690126776695, + 0.02636036090552807, + 0.026620972901582718, + 0.0435180589556694, + 0.0349181666970253, + 0.034443199634552, + 0.03524469956755638, + 0.009204944595694542, + 0.005581534467637539, + 0.007294046692550182, + 0.01481532771140337, + 0.005634296219795942, + 0.00789781752973795, + 0.009407284669578075, + 0.019561655819416046, + 0.005913795903325081, + 0.005274837836623192, + 0.006911398842930794, + 0.008272799663245678, + 0.0053368061780929565, + 0.006434725131839514, + 0.007844094187021255, + 0.008070504292845726, + 0.005575911141932011, + 0.005015158094465733, + 0.004491476342082024, + 0.006666960194706917, + 0.0022682086564600468, + 0.004797694273293018, + 0.0034015709534287453, + 0.0072798822075128555, + 0.006626366637647152, + 0.00285654840990901, + 0.007106258999556303, + 0.0068707833997905254, + 0.004290369804948568, + 0.008375879377126694, + 0.004899668041616678, + 0.009023087099194527, + 0.011113228276371956, + 0.006761335302144289, + 0.008646722882986069, + 0.012467971071600914, + 0.016514604911208153, + 0.0936724841594696, + 0.03165806829929352, + 0.08596482127904892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.03368346393108368 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot glides swiftly above the ground, keeping pace with a small, sleek quadcopter drone as it navigates a vast, sun-drenched field of wildflowers. Constructed of matte black carbon fiber, the drone is equipped with four translucent propellers that blur into near-invisibility as it spins, and a soft cyan LED glows from its front. It gallops merrily, agilely executing sharp, banked turns, effortlessly darting through dense fields of red poppies, white daisies, and purple lavender, all swaying in the gentle afternoon breeze. As the camera follows, the drone momentarily halts its forward momentum, delicately hovering just inches above a vibrant yellow flower, tilting its chassis downward like a playful bow before quickly rising and resuming its playful dance. The entire scene is bathed in the warm, golden afternoon light, casting long, soft shadows that make the delicate petals appear almost translucent, creating a tranquil and wondrous atmosphere that blends technological elegance with the beauty of nature.", + "seed": 26131, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_12/index_6934_seed_26131_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.10975462533533573, + "vlm_logprob": -0.10975462533533573, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.10975462533533573, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, sleek quadcopter drone constructed of matte black carbon fiber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drone have four translucent propellers and a soft cyan LED glowing from its front?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a vast, sun-drenched field containing red poppies, white daisies, and purple lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone execute sharp, banked turns and momentarily hover just inches above a vibrant yellow flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle tracking shot that glides swiftly above the ground following the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.39261066913604736 + ], + "opposite_logprob_tokens": [ + -1.1426106691360474 + ], + "logprob_reward": -0.39261066913604736, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.39261066913604736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15248191356658936, + 0.06428752094507217, + 0.05456901341676712, + 0.179163858294487, + 0.049338437616825104, + 0.07541628926992416, + 0.09137392044067383, + 0.14474545419216156, + 0.03438262641429901, + 0.15424098074436188 + ], + "credit_map": [ + 0.017991788685321808, + 0.02152487076818943, + 0.02257140539586544, + 0.014082139357924461, + 0.019538184627890587, + 0.020114708691835403, + 0.020532771944999695, + 0.016126062721014023, + 0.0051518515683710575, + 0.011281423270702362, + 0.005374769680202007, + 0.006927276961505413, + 0.012470018118619919, + 0.007096668239682913, + 0.010787981562316418, + 0.005197532474994659, + 0.005739070009440184, + 0.004675382748246193, + 0.007325372658669949, + 0.007241727784276009, + 0.00608841422945261, + 0.010902361944317818, + 0.005312229041010141, + 0.007284453138709068, + 0.04994383454322815, + 0.01933862455189228, + 0.017925001680850983, + 0.014110037125647068, + 0.011538111604750156, + 0.040457550436258316, + 0.010678054764866829, + 0.015172631479799747, + 0.005226490553468466, + 0.006235235370695591, + 0.008742991834878922, + 0.006807718425989151, + 0.003993717487901449, + 0.00561728747561574, + 0.006231874693185091, + 0.006483121775090694, + 0.007124303840100765, + 0.007991242222487926, + 0.011932452209293842, + 0.009542031213641167, + 0.008929390460252762, + 0.009077673777937889, + 0.009950664825737476, + 0.010868524201214314, + 0.00979678425937891, + 0.00860772468149662, + 0.012522461824119091, + 0.00959418062120676, + 0.021581590175628662, + 0.01730642281472683, + 0.005826635751873255, + 0.006138123571872711, + 0.0150664197281003, + 0.021149979904294014, + 0.021773304790258408, + 0.016804656013846397, + 0.015125676058232784, + 0.014849775470793247, + 0.01816979981958866, + 0.021805843338370323, + 0.003711419180035591, + 0.003656571265310049, + 0.004189843311905861, + 0.006221243180334568, + 0.0033353366889059544, + 0.003976480104029179, + 0.004790648352354765, + 0.0045010861940681934, + 0.020458918064832687, + 0.020902691408991814, + 0.02413388155400753, + 0.028216872364282608, + 0.009766793809831142, + 0.013789009302854538, + 0.013286427594721317, + 0.023686382919549942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.39261066913604736 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, sleek quadcopter drone constructed of matte black carbon fiber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drone have four translucent propellers and a soft cyan LED glowing from its front?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a vast, sun-drenched field containing red poppies, white daisies, and purple lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone execute sharp, banked turns and momentarily hover just inches above a vibrant yellow flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle tracking shot that glides swiftly above the ground following the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043080300092697144 + ], + "opposite_logprob_tokens": [ + -3.1680803298950195 + ], + "logprob_reward": -0.043080300092697144, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043080300092697144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10821115970611572, + 0.050898585468530655, + 0.09457124769687653, + 0.2373867779970169, + 0.0696115493774414, + 0.09556279331445694, + 0.06533600389957428, + 0.14770330488681793, + 0.06513214111328125, + 0.06558646261692047 + ], + "credit_map": [ + 0.01359994150698185, + 0.014830688945949078, + 0.021027114242315292, + 0.011268365196883678, + 0.009403456002473831, + 0.01603350229561329, + 0.011784417554736137, + 0.01026366651058197, + 0.006807039957493544, + 0.005653164815157652, + 0.008063340559601784, + 0.005006034858524799, + 0.004867876414209604, + 0.0068317558616399765, + 0.007724660448729992, + 0.005944712553173304, + 0.01135567482560873, + 0.008352261036634445, + 0.021238671615719795, + 0.007883618585765362, + 0.013079206459224224, + 0.014633572660386562, + 0.010590046644210815, + 0.0074381958693265915, + 0.06011208891868591, + 0.023423930630087852, + 0.027738502249121666, + 0.023770691826939583, + 0.014562182128429413, + 0.04938889294862747, + 0.017236363142728806, + 0.021154118701815605, + 0.006804102566093206, + 0.008774029091000557, + 0.011324884369969368, + 0.01802028715610504, + 0.005328557454049587, + 0.006536329630762339, + 0.006471707485616207, + 0.006351653952151537, + 0.007642908021807671, + 0.00919833593070507, + 0.011486643925309181, + 0.01396153587847948, + 0.012006768025457859, + 0.0173958707600832, + 0.011441731825470924, + 0.012429002672433853, + 0.007461311295628548, + 0.006444217637181282, + 0.011256993748247623, + 0.01034360472112894, + 0.007986812852323055, + 0.00944919977337122, + 0.0057937684468925, + 0.0066000898368656635, + 0.015095103532075882, + 0.022038238123059273, + 0.021918218582868576, + 0.014755136333405972, + 0.02077626809477806, + 0.014232027344405651, + 0.016935231164097786, + 0.021953081712126732, + 0.010153301060199738, + 0.007399602327495813, + 0.009473406709730625, + 0.009405909106135368, + 0.005777561105787754, + 0.008593549020588398, + 0.006596736144274473, + 0.007732073310762644, + 0.005618480499833822, + 0.00737369479611516, + 0.011017338372766972, + 0.009100683964788914, + 0.006044122390449047, + 0.008457818999886513, + 0.004853405524045229, + 0.013120915740728378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.043080300092697144 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, sleek quadcopter drone constructed of matte black carbon fiber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drone have four translucent propellers and a soft cyan LED glowing from its front?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a vast, sun-drenched field containing red poppies, white daisies, and purple lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone execute sharp, banked turns and momentarily hover just inches above a vibrant yellow flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle tracking shot that glides swiftly above the ground following the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634120173752308 + ], + "opposite_logprob_tokens": [ + -3.651341199874878 + ], + "logprob_reward": -0.02634120173752308, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634120173752308, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13104531168937683, + 0.10672655701637268, + 0.1417984515428543, + 0.1486942172050476, + 0.11632996052503586, + 0.05684538930654526, + 0.07636500895023346, + 0.09049726277589798, + 0.027707835659384727, + 0.10398995876312256 + ], + "credit_map": [ + 0.017948785796761513, + 0.013531658798456192, + 0.030369332060217857, + 0.013536067679524422, + 0.009932352229952812, + 0.01589009165763855, + 0.012256670743227005, + 0.017580369487404823, + 0.014629613608121872, + 0.014143440872430801, + 0.01675846241414547, + 0.009270510636270046, + 0.011653234250843525, + 0.014283237047493458, + 0.014706247486174107, + 0.011281822808086872, + 0.017474085092544556, + 0.011655761860311031, + 0.025868907570838928, + 0.018537987023591995, + 0.012889743782579899, + 0.027816060930490494, + 0.010868312790989876, + 0.016687611117959023, + 0.021070638671517372, + 0.0234907828271389, + 0.021808765828609467, + 0.02028411068022251, + 0.011439147405326366, + 0.018010081723332405, + 0.013207259587943554, + 0.019383445382118225, + 0.012053025886416435, + 0.01684071496129036, + 0.012267923913896084, + 0.03468538075685501, + 0.008312195539474487, + 0.01386451069265604, + 0.008179805241525173, + 0.010126424953341484, + 0.0048595271073281765, + 0.006645857822149992, + 0.01151096262037754, + 0.004882689565420151, + 0.008109183982014656, + 0.011179724708199501, + 0.00485687842592597, + 0.004800571594387293, + 0.008897542022168636, + 0.010464786551892757, + 0.012450204230844975, + 0.0061906180344522, + 0.011849823407828808, + 0.016267573460936546, + 0.004600476939231157, + 0.005643991753458977, + 0.009730659425258636, + 0.01193212065845728, + 0.013798579573631287, + 0.011051397770643234, + 0.007429527584463358, + 0.009141347371041775, + 0.01275921892374754, + 0.01465442031621933, + 0.004536912776529789, + 0.0037520411424338818, + 0.003651741426438093, + 0.004809923004359007, + 0.0021286781411617994, + 0.002955419709905982, + 0.0024648953694850206, + 0.0034082273487001657, + 0.015232586301863194, + 0.012020239606499672, + 0.0145144397392869, + 0.024062655866146088, + 0.006510161329060793, + 0.008751139044761658, + 0.010165253654122353, + 0.012733490206301212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.02634120173752308 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, sleek quadcopter drone constructed of matte black carbon fiber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drone have four translucent propellers and a soft cyan LED glowing from its front?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a vast, sun-drenched field containing red poppies, white daisies, and purple lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone execute sharp, banked turns and momentarily hover just inches above a vibrant yellow flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle tracking shot that glides swiftly above the ground following the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865409806370735 + ], + "opposite_logprob_tokens": [ + -3.048654079437256 + ], + "logprob_reward": -0.04865409806370735, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865409806370735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06832265108823776, + 0.10133901238441467, + 0.15010392665863037, + 0.21343405544757843, + 0.051035258919000626, + 0.06101229414343834, + 0.12381661683320999, + 0.13115382194519043, + 0.054593008011579514, + 0.04518930986523628 + ], + "credit_map": [ + 0.00950737576931715, + 0.007388918194919825, + 0.01118822954595089, + 0.008980452083051205, + 0.006438989657908678, + 0.009294869378209114, + 0.006541773211210966, + 0.008982042782008648, + 0.007405308540910482, + 0.01996118202805519, + 0.008988702669739723, + 0.008496729657053947, + 0.02157435193657875, + 0.01187371276319027, + 0.015939414501190186, + 0.007099601905792952, + 0.0187405776232481, + 0.013944142498075962, + 0.02327757515013218, + 0.02141527831554413, + 0.01829473488032818, + 0.020089903846383095, + 0.015381802804768085, + 0.018959905952215195, + 0.061540618538856506, + 0.022537989541888237, + 0.024152621626853943, + 0.016150105744600296, + 0.011685545556247234, + 0.0488537959754467, + 0.01186379138380289, + 0.01664959453046322, + 0.004259583540260792, + 0.006412157788872719, + 0.00789324939250946, + 0.009667308069765568, + 0.004457322880625725, + 0.0042845262214541435, + 0.0066164713352918625, + 0.00744463549926877, + 0.00649725878611207, + 0.008849686942994595, + 0.012446294538676739, + 0.006893528625369072, + 0.0059072901494801044, + 0.006676000542938709, + 0.006465207785367966, + 0.007277026306837797, + 0.008416270837187767, + 0.010991061106324196, + 0.01815618760883808, + 0.021166305989027023, + 0.0205087848007679, + 0.013351145200431347, + 0.014228614047169685, + 0.01699824444949627, + 0.01572417840361595, + 0.015497230924665928, + 0.018921907991170883, + 0.012926267459988594, + 0.014685575850307941, + 0.01963869482278824, + 0.014908021315932274, + 0.01885194703936577, + 0.00685908691957593, + 0.0064646098762750626, + 0.006925556808710098, + 0.010157766751945019, + 0.004958970006555319, + 0.0069574154913425446, + 0.006016852334141731, + 0.006252752151340246, + 0.004913865588605404, + 0.005270110908895731, + 0.0075774770230054855, + 0.008372988551855087, + 0.0038380154874175787, + 0.005913210567086935, + 0.0037808765191584826, + 0.005522759631276131 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 7 + ], + "selected_reward": -0.04865409806370735 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, sleek quadcopter drone constructed of matte black carbon fiber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drone have four translucent propellers and a soft cyan LED glowing from its front?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a vast, sun-drenched field containing red poppies, white daisies, and purple lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone execute sharp, banked turns and momentarily hover just inches above a vibrant yellow flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle tracking shot that glides swiftly above the ground following the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808685764670372 + ], + "opposite_logprob_tokens": [ + -3.2880868911743164 + ], + "logprob_reward": -0.03808685764670372, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808685764670372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17937009036540985, + 0.072509765625, + 0.07142765074968338, + 0.22455163300037384, + 0.05059102550148964, + 0.05469249561429024, + 0.1463889628648758, + 0.06033977493643761, + 0.09450116753578186, + 0.04562744125723839 + ], + "credit_map": [ + 0.020062237977981567, + 0.025093087926506996, + 0.025539357215166092, + 0.017148049548268318, + 0.022999674081802368, + 0.025232871994376183, + 0.025575904175639153, + 0.01771889626979828, + 0.007579309865832329, + 0.010499437339603901, + 0.008028044365346432, + 0.007496175356209278, + 0.013210096396505833, + 0.0086000245064497, + 0.011793882586061954, + 0.005302794743329287, + 0.009602553211152554, + 0.006647004280239344, + 0.008312511257827282, + 0.00722285732626915, + 0.008093639276921749, + 0.014607883058488369, + 0.008092761039733887, + 0.008848440833389759, + 0.06210941821336746, + 0.023914294317364693, + 0.025710875168442726, + 0.019480962306261063, + 0.012322884984314442, + 0.0486561618745327, + 0.01397594902664423, + 0.01838107220828533, + 0.004879684653133154, + 0.007958969101309776, + 0.006320473272353411, + 0.00865235272794962, + 0.0046579656191170216, + 0.007093839813023806, + 0.005048966035246849, + 0.005978775210678577, + 0.005118597764521837, + 0.006705719046294689, + 0.009984157048165798, + 0.006618304643779993, + 0.006253190338611603, + 0.006367172114551067, + 0.007352661807090044, + 0.0062926942482590675, + 0.009185523726046085, + 0.011365558952093124, + 0.022158948704600334, + 0.022780586034059525, + 0.0222612414509058, + 0.01968071237206459, + 0.01980452984571457, + 0.019151858985424042, + 0.007020059507340193, + 0.011208921670913696, + 0.009483292698860168, + 0.007447214797139168, + 0.0047068470157682896, + 0.006253975443542004, + 0.007706831209361553, + 0.006512628868222237, + 0.010322246700525284, + 0.010913177393376827, + 0.0113895945250988, + 0.015660282224416733, + 0.009611770510673523, + 0.01256521511822939, + 0.010915317572653294, + 0.013123567216098309, + 0.005827806890010834, + 0.006212275940924883, + 0.005134536884725094, + 0.00713602127507329, + 0.0046740020625293255, + 0.007087957579642534, + 0.004613020457327366, + 0.004941822495311499 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.03808685764670372 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot glides swiftly above the ground, keeping pace with a small, sleek quadcopter drone as it navigates a vast, sun-drenched field of wildflowers. Constructed of matte black carbon fiber, the drone is equipped with four translucent propellers that blur into near-invisibility as it spins, and a soft cyan LED glows from its front. It gallops merrily, agilely executing sharp, banked turns, effortlessly darting through dense fields of red poppies, white daisies, and purple lavender, all swaying in the gentle afternoon breeze. As the camera follows, the drone momentarily halts its forward momentum, delicately hovering just inches above a vibrant yellow flower, tilting its chassis downward like a playful bow before quickly rising and resuming its playful dance. The entire scene is bathed in the warm, golden afternoon light, casting long, soft shadows that make the delicate petals appear almost translucent, creating a tranquil and wondrous atmosphere that blends technological elegance with the beauty of nature.", + "seed": 26132, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_12/index_6934_seed_26132_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.12060571312904358, + "vlm_logprob": -0.12060571312904358, + "vlm_yes_no_margin": 2.5, + "vlm_token_credit": -0.12060571312904358, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, sleek quadcopter drone constructed of matte black carbon fiber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drone have four translucent propellers and a soft cyan LED glowing from its front?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a vast, sun-drenched field containing red poppies, white daisies, and purple lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone execute sharp, banked turns and momentarily hover just inches above a vibrant yellow flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle tracking shot that glides swiftly above the ground following the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.39427855610847473 + ], + "opposite_logprob_tokens": [ + -1.1442785263061523 + ], + "logprob_reward": -0.39427855610847473, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.39427855610847473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.033222656697034836, + 0.1585124135017395, + 0.4643547832965851, + 0.03363547846674919, + 0.026731308549642563, + 0.021041564643383026, + 0.048451147973537445, + 0.05473947152495384, + 0.13809072971343994, + 0.02122044377028942 + ], + "credit_map": [ + 0.0029685976915061474, + 0.00405182084068656, + 0.004403240978717804, + 0.005153776612132788, + 0.0035255583934485912, + 0.004580153152346611, + 0.004255453124642372, + 0.004284054972231388, + 0.03751380741596222, + 0.02591104246675968, + 0.01154528558254242, + 0.007731694262474775, + 0.01736636273562908, + 0.032136984169483185, + 0.010275417007505894, + 0.01603182591497898, + 0.04724327474832535, + 0.07977265119552612, + 0.05010267347097397, + 0.03353438898921013, + 0.0411682091653347, + 0.07141198962926865, + 0.0508231520652771, + 0.09029844403266907, + 0.0038445049431174994, + 0.004381717648357153, + 0.0030718643683940172, + 0.0036776172928512096, + 0.004959463141858578, + 0.0029680386651307344, + 0.007042319048196077, + 0.00368995382450521, + 0.002062389627099037, + 0.004504619166254997, + 0.0026161461137235165, + 0.003524787025526166, + 0.003544223727658391, + 0.003848109859973192, + 0.0031566093675792217, + 0.0034744220320135355, + 0.001692493329755962, + 0.0025929678231477737, + 0.002817918546497822, + 0.0020988991018384695, + 0.003483316395431757, + 0.0028145327232778072, + 0.0030027718748897314, + 0.0025386661291122437, + 0.006480890326201916, + 0.008305954746901989, + 0.003901960328221321, + 0.003578657517209649, + 0.008463895879685879, + 0.007999557070434093, + 0.004961626138538122, + 0.004758602473884821, + 0.00791717879474163, + 0.006105801090598106, + 0.005023237317800522, + 0.009893370792269707, + 0.004724438302218914, + 0.0070989797823131084, + 0.006224555894732475, + 0.0077519044280052185, + 0.015098512172698975, + 0.015698229894042015, + 0.019510090351104736, + 0.015463996678590775, + 0.014020751230418682, + 0.017795000225305557, + 0.022106437012553215, + 0.018397711217403412, + 0.0026405882090330124, + 0.003997539170086384, + 0.0025423213373869658, + 0.002677010139450431, + 0.0023667998611927032, + 0.0024894156958907843, + 0.002372808987274766, + 0.0021339592058211565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 8 + ], + "selected_reward": -0.39427855610847473 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, sleek quadcopter drone constructed of matte black carbon fiber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drone have four translucent propellers and a soft cyan LED glowing from its front?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a vast, sun-drenched field containing red poppies, white daisies, and purple lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone execute sharp, banked turns and momentarily hover just inches above a vibrant yellow flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle tracking shot that glides swiftly above the ground following the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062083687633275986 + ], + "opposite_logprob_tokens": [ + -2.8120837211608887 + ], + "logprob_reward": -0.062083687633275986, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062083687633275986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08847849071025848, + 0.19837170839309692, + 0.3804895281791687, + 0.04277007281780243, + 0.026308447122573853, + 0.030882004648447037, + 0.05942493677139282, + 0.06206071749329567, + 0.06317032873630524, + 0.04804377630352974 + ], + "credit_map": [ + 0.008495528250932693, + 0.011804169043898582, + 0.012492758221924305, + 0.014730027876794338, + 0.009593896567821503, + 0.012089948169887066, + 0.008965402841567993, + 0.010306763462722301, + 0.041609227657318115, + 0.016421478241682053, + 0.010160768404603004, + 0.005428133998066187, + 0.02356017380952835, + 0.07650236040353775, + 0.00929199531674385, + 0.015397574752569199, + 0.038960374891757965, + 0.06709751486778259, + 0.03935369849205017, + 0.02603122964501381, + 0.033989936113357544, + 0.06204061955213547, + 0.040915362536907196, + 0.07210079580545425, + 0.0039544859901070595, + 0.005472226999700069, + 0.004972555208951235, + 0.003679976100102067, + 0.005236949305981398, + 0.006402608472853899, + 0.006079945247620344, + 0.006971330381929874, + 0.0028590725269168615, + 0.0031714397482573986, + 0.0029248057398945093, + 0.0029714342672377825, + 0.0030636172741651535, + 0.0038949456065893173, + 0.0038346494548022747, + 0.003588484600186348, + 0.002709053922444582, + 0.0038991346955299377, + 0.002988913096487522, + 0.004900104831904173, + 0.0032397143077105284, + 0.0052499654702842236, + 0.00394412549212575, + 0.003950995393097401, + 0.0085980836302042, + 0.009276662021875381, + 0.004148602485656738, + 0.004504724871367216, + 0.009380512870848179, + 0.009841879829764366, + 0.006166899111121893, + 0.007507577072829008, + 0.010461525991559029, + 0.010081855580210686, + 0.0061257812194526196, + 0.008154409937560558, + 0.006577633321285248, + 0.0076791830360889435, + 0.0067485724575817585, + 0.006231761537492275, + 0.008074408397078514, + 0.00877206027507782, + 0.006797901354730129, + 0.0073095508851110935, + 0.010501131415367126, + 0.008160548284649849, + 0.007156308740377426, + 0.006398429628461599, + 0.004895160440355539, + 0.007703007198870182, + 0.0066745164804160595, + 0.007705475203692913, + 0.006057608872652054, + 0.004198435693979263, + 0.004793139640241861, + 0.00601643742993474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.062083687633275986 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, sleek quadcopter drone constructed of matte black carbon fiber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drone have four translucent propellers and a soft cyan LED glowing from its front?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a vast, sun-drenched field containing red poppies, white daisies, and purple lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone execute sharp, banked turns and momentarily hover just inches above a vibrant yellow flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle tracking shot that glides swiftly above the ground following the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865046218037605 + ], + "opposite_logprob_tokens": [ + -3.0486505031585693 + ], + "logprob_reward": -0.04865046218037605, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865046218037605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0711473822593689, + 0.22939907014369965, + 0.47260603308677673, + 0.06321951746940613, + 0.02114313282072544, + 0.017680225893855095, + 0.008637807331979275, + 0.03571322187781334, + 0.06748919934034348, + 0.012964511290192604 + ], + "credit_map": [ + 0.0062132542952895164, + 0.010385056026279926, + 0.008999415673315525, + 0.011184466071426868, + 0.008121159859001637, + 0.009622128680348396, + 0.008275661617517471, + 0.00834623258560896, + 0.05338120087981224, + 0.019421368837356567, + 0.014606954529881477, + 0.008407660759985447, + 0.026876425370573997, + 0.07465730607509613, + 0.010979301296174526, + 0.021068841218948364, + 0.04932784289121628, + 0.08198557049036026, + 0.051233142614364624, + 0.03422250971198082, + 0.04110714793205261, + 0.07088553160429001, + 0.05156408250331879, + 0.09228016436100006, + 0.007815363816916943, + 0.01064428873360157, + 0.007737590931355953, + 0.005229496862739325, + 0.0062670293264091015, + 0.009097552858293056, + 0.0056740157306194305, + 0.010754170827567577, + 0.00233846390619874, + 0.003394119208678603, + 0.003621127223595977, + 0.002845257055014372, + 0.0025409820955246687, + 0.001832001842558384, + 0.0017432718304917216, + 0.0028279090765863657, + 0.00198655528947711, + 0.0019822895992547274, + 0.00230593653395772, + 0.0032437886111438274, + 0.001328534446656704, + 0.0022137307096272707, + 0.0020393168088048697, + 0.002580073429271579, + 0.0010161844547837973, + 0.0010876775486394763, + 0.0008281737100332975, + 0.0007466917741112411, + 0.0009792017517611384, + 0.0010171466274186969, + 0.001342667150311172, + 0.001620063791051507, + 0.0043646832928061485, + 0.004185078199952841, + 0.00523866293951869, + 0.007177427876740694, + 0.0024688434787094593, + 0.002286203671246767, + 0.005331871565431356, + 0.004660448059439659, + 0.007774646859616041, + 0.005893083289265633, + 0.008165564388036728, + 0.01315043494105339, + 0.009255792014300823, + 0.005154453217983246, + 0.011298448778688908, + 0.00679676653817296, + 0.0015562146436423063, + 0.0012171493144705892, + 0.002038035076111555, + 0.0020493504125624895, + 0.0011391721200197935, + 0.0010463332291692495, + 0.0020699896849691868, + 0.0018482652958482504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.04865046218037605 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, sleek quadcopter drone constructed of matte black carbon fiber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drone have four translucent propellers and a soft cyan LED glowing from its front?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a vast, sun-drenched field containing red poppies, white daisies, and purple lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone execute sharp, banked turns and momentarily hover just inches above a vibrant yellow flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle tracking shot that glides swiftly above the ground following the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496479198336601 + ], + "opposite_logprob_tokens": [ + -2.929964780807495 + ], + "logprob_reward": -0.05496479198336601, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496479198336601, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09568105638027191, + 0.3339807689189911, + 0.08154267072677612, + 0.0460486002266407, + 0.09156788885593414, + 0.0623670294880867, + 0.044994790107011795, + 0.049992918968200684, + 0.12126054614782333, + 0.07256367802619934 + ], + "credit_map": [ + 0.0074174911715090275, + 0.012739073485136032, + 0.011808451265096664, + 0.013325382955372334, + 0.00859047845005989, + 0.008329499512910843, + 0.014067189767956734, + 0.019403500482439995, + 0.06814160197973251, + 0.11213670670986176, + 0.021610921248793602, + 0.023169949650764465, + 0.029730824753642082, + 0.03161410242319107, + 0.019293664023280144, + 0.0282830111682415, + 0.008629662916064262, + 0.013571912422776222, + 0.008535584434866905, + 0.006113642361015081, + 0.008365251123905182, + 0.013699590228497982, + 0.008545697666704655, + 0.014081336557865143, + 0.004981937352567911, + 0.006712777074426413, + 0.005805585999041796, + 0.004180310294032097, + 0.005439641885459423, + 0.0075113167986273766, + 0.0060357931070029736, + 0.005381242372095585, + 0.008238517679274082, + 0.01850459724664688, + 0.00793027225881815, + 0.013744854368269444, + 0.012054722756147385, + 0.010114363394677639, + 0.011129004880785942, + 0.009851563721895218, + 0.005294394213706255, + 0.008058498613536358, + 0.007590755354613066, + 0.0069710128009319305, + 0.009009161032736301, + 0.00970346387475729, + 0.007874767296016216, + 0.007864980027079582, + 0.0039453282952308655, + 0.0058291698805987835, + 0.004245060961693525, + 0.005068651400506496, + 0.006903265602886677, + 0.006626268848776817, + 0.005776027217507362, + 0.0066010174341499805, + 0.007823915220797062, + 0.00716985622420907, + 0.005625663325190544, + 0.005507453344762325, + 0.0069986432790756226, + 0.006818484049290419, + 0.005020396783947945, + 0.00502850953489542, + 0.012002848088741302, + 0.009604057297110558, + 0.024572988972067833, + 0.026885347440838814, + 0.007487728260457516, + 0.00715916883200407, + 0.020742688328027725, + 0.012805725447833538, + 0.00798525009304285, + 0.010257240384817123, + 0.011698351241648197, + 0.010398770682513714, + 0.008312148042023182, + 0.006824920419603586, + 0.007482581771910191, + 0.009604421444237232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.05496479198336601 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, sleek quadcopter drone constructed of matte black carbon fiber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drone have four translucent propellers and a soft cyan LED glowing from its front?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a vast, sun-drenched field containing red poppies, white daisies, and purple lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone execute sharp, banked turns and momentarily hover just inches above a vibrant yellow flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle tracking shot that glides swiftly above the ground following the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305106773972511 + ], + "opposite_logprob_tokens": [ + -3.16805100440979 + ], + "logprob_reward": -0.04305106773972511, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305106773972511, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050098489969968796, + 0.19668813049793243, + 0.4385904371738434, + 0.03939357399940491, + 0.043642643839120865, + 0.05634188652038574, + 0.05259879305958748, + 0.029413312673568726, + 0.05395505577325821, + 0.03927769139409065 + ], + "credit_map": [ + 0.004256338346749544, + 0.004628273192793131, + 0.007209423463791609, + 0.008840188384056091, + 0.005396557040512562, + 0.0042708078399300575, + 0.008022365160286427, + 0.007474539801478386, + 0.04068975895643234, + 0.05822933465242386, + 0.014142914675176144, + 0.016244636848568916, + 0.019695663824677467, + 0.013016478158533573, + 0.013543671928346157, + 0.0211256742477417, + 0.04393249750137329, + 0.07264391332864761, + 0.048412542790174484, + 0.03373207524418831, + 0.03935583680868149, + 0.06655821949243546, + 0.04878334328532219, + 0.08517203480005264, + 0.003794765565544367, + 0.004842029884457588, + 0.0047273957170546055, + 0.0045917765237390995, + 0.004430611152201891, + 0.0050429049879312515, + 0.005653253756463528, + 0.006310837808996439, + 0.0036349031142890453, + 0.00718251708894968, + 0.003475036472082138, + 0.007387036457657814, + 0.005133522674441338, + 0.005422674119472504, + 0.005739312153309584, + 0.005667640827596188, + 0.004771902225911617, + 0.0068938094191253185, + 0.0073885549791157246, + 0.005785416811704636, + 0.009320412762463093, + 0.007246201857924461, + 0.008262558840215206, + 0.006673030089586973, + 0.005628976505249739, + 0.007798224221915007, + 0.005855323746800423, + 0.006532951723784208, + 0.007896491326391697, + 0.005987316835671663, + 0.005790065508335829, + 0.007109441328793764, + 0.005658898502588272, + 0.006074385717511177, + 0.002727426588535309, + 0.003120896639302373, + 0.0035816715098917484, + 0.004120595287531614, + 0.0021073638927191496, + 0.0020220738369971514, + 0.005070029757916927, + 0.004601213149726391, + 0.01102772168815136, + 0.0093140359967947, + 0.003835960291326046, + 0.0043856422416865826, + 0.009833191521465778, + 0.005887260194867849, + 0.0037009709049016237, + 0.005018309690058231, + 0.006187941879034042, + 0.005988783203065395, + 0.003064907854422927, + 0.0024701303336769342, + 0.0062103671953082085, + 0.006636282429099083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 5 + ], + "selected_reward": -0.04305106773972511 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot glides swiftly above the ground, keeping pace with a small, sleek quadcopter drone as it navigates a vast, sun-drenched field of wildflowers. Constructed of matte black carbon fiber, the drone is equipped with four translucent propellers that blur into near-invisibility as it spins, and a soft cyan LED glows from its front. It gallops merrily, agilely executing sharp, banked turns, effortlessly darting through dense fields of red poppies, white daisies, and purple lavender, all swaying in the gentle afternoon breeze. As the camera follows, the drone momentarily halts its forward momentum, delicately hovering just inches above a vibrant yellow flower, tilting its chassis downward like a playful bow before quickly rising and resuming its playful dance. The entire scene is bathed in the warm, golden afternoon light, casting long, soft shadows that make the delicate petals appear almost translucent, creating a tranquil and wondrous atmosphere that blends technological elegance with the beauty of nature.", + "seed": 26133, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_12/index_6934_seed_26133_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.049820948764681815, + "vlm_logprob": -0.049820948764681815, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.049820948764681815, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, sleek quadcopter drone constructed of matte black carbon fiber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drone have four translucent propellers and a soft cyan LED glowing from its front?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a vast, sun-drenched field containing red poppies, white daisies, and purple lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone execute sharp, banked turns and momentarily hover just inches above a vibrant yellow flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle tracking shot that glides swiftly above the ground following the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1342633068561554 + ], + "opposite_logprob_tokens": [ + -2.134263277053833 + ], + "logprob_reward": -0.1342633068561554, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1342633068561554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4802091717720032, + 0.08336915820837021, + 0.038535092025995255, + 0.05271275341510773, + 0.10245136171579361, + 0.034857530146837234, + 0.018526187166571617, + 0.03989284113049507, + 0.04130897298455238, + 0.10813693702220917 + ], + "credit_map": [ + 0.04472621530294418, + 0.036045026034116745, + 0.07208360731601715, + 0.03974534571170807, + 0.048083145171403885, + 0.05842891335487366, + 0.07195492833852768, + 0.1091419905424118, + 0.006271780468523502, + 0.010667338036000729, + 0.008810628205537796, + 0.013386368751525879, + 0.011006919667124748, + 0.010310020297765732, + 0.008926939219236374, + 0.013989158906042576, + 0.003095290157943964, + 0.004707366228103638, + 0.004397762939333916, + 0.0031677649822086096, + 0.008771248161792755, + 0.006941585801541805, + 0.0030274996533989906, + 0.004426577594131231, + 0.004875045269727707, + 0.0047359773889184, + 0.0063039870001375675, + 0.008317526429891586, + 0.004511021077632904, + 0.0066450913436710835, + 0.010450314730405807, + 0.006873792503029108, + 0.007938576862215996, + 0.013638654723763466, + 0.007460706401616335, + 0.021894557401537895, + 0.011192267760634422, + 0.01056672167032957, + 0.019314521923661232, + 0.010445361025631428, + 0.003114780643954873, + 0.004362029489129782, + 0.003661337075755, + 0.004899842198938131, + 0.004717825911939144, + 0.004137793090194464, + 0.0048563494347035885, + 0.005107571836560965, + 0.0027800973039120436, + 0.0021886301692575216, + 0.0031290503684431314, + 0.00276814098469913, + 0.0016533101443201303, + 0.0016842534532770514, + 0.002075567375868559, + 0.0022471400443464518, + 0.003477361286059022, + 0.00516550475731492, + 0.0049228365533053875, + 0.005110862199217081, + 0.007232150062918663, + 0.004504427779465914, + 0.004206478595733643, + 0.005273220129311085, + 0.004653938114643097, + 0.004104358609765768, + 0.004358384758234024, + 0.005139574874192476, + 0.004890949930995703, + 0.005998394917696714, + 0.004563930444419384, + 0.007599443197250366, + 0.012949824333190918, + 0.010287029668688774, + 0.01898612640798092, + 0.01061083935201168, + 0.016212861984968185, + 0.011969885788857937, + 0.01226856093853712, + 0.014851810410618782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.1342633068561554 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, sleek quadcopter drone constructed of matte black carbon fiber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drone have four translucent propellers and a soft cyan LED glowing from its front?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a vast, sun-drenched field containing red poppies, white daisies, and purple lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone execute sharp, banked turns and momentarily hover just inches above a vibrant yellow flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle tracking shot that glides swiftly above the ground following the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0263676755130291 + ], + "opposite_logprob_tokens": [ + -3.651367664337158 + ], + "logprob_reward": -0.0263676755130291, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0263676755130291, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16929179430007935, + 0.24891215562820435, + 0.033256419003009796, + 0.06490281969308853, + 0.13219165802001953, + 0.06571144610643387, + 0.06903109699487686, + 0.05048705264925957, + 0.040872491896152496, + 0.12534300982952118 + ], + "credit_map": [ + 0.018183454871177673, + 0.01317776832729578, + 0.03346877545118332, + 0.01597636379301548, + 0.02241731621325016, + 0.025791538879275322, + 0.02111399546265602, + 0.019162587821483612, + 0.021860415115952492, + 0.03635687008500099, + 0.032917220145463943, + 0.027730759233236313, + 0.03226426988840103, + 0.02844500169157982, + 0.031897060573101044, + 0.037440571933984756, + 0.0034742809366434813, + 0.003909071907401085, + 0.004072864539921284, + 0.0033815146889537573, + 0.003995461389422417, + 0.004914626944810152, + 0.00391622819006443, + 0.005592373665422201, + 0.005619203206151724, + 0.005437211599200964, + 0.009095595218241215, + 0.009155362844467163, + 0.00664642034098506, + 0.009133925661444664, + 0.012075887992978096, + 0.007739215157926083, + 0.012391000054776669, + 0.012840605340898037, + 0.01566227711737156, + 0.02204188145697117, + 0.012656047008931637, + 0.014246813952922821, + 0.02154216542840004, + 0.02081087790429592, + 0.005818068981170654, + 0.007918916642665863, + 0.008342702873051167, + 0.009144794195890427, + 0.009126189164817333, + 0.008023313246667385, + 0.00856572575867176, + 0.008771747350692749, + 0.012675256468355656, + 0.007269192021340132, + 0.008982811123132706, + 0.007157220039516687, + 0.005354126915335655, + 0.006826015189290047, + 0.005953057203441858, + 0.014813428744673729, + 0.005056778900325298, + 0.005360559560358524, + 0.009994164109230042, + 0.0062841554172337055, + 0.005755456630140543, + 0.004321217071264982, + 0.0056023309007287025, + 0.008112391456961632, + 0.003653406398370862, + 0.003605443285778165, + 0.006215041037648916, + 0.0059580509550869465, + 0.0046084788627922535, + 0.005458668805658817, + 0.005812551826238632, + 0.005560855846852064, + 0.009295773692429066, + 0.013753808103501797, + 0.017286233603954315, + 0.015285960398614407, + 0.016514943912625313, + 0.014129837043583393, + 0.01687988080084324, + 0.02219657599925995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.0263676755130291 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, sleek quadcopter drone constructed of matte black carbon fiber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drone have four translucent propellers and a soft cyan LED glowing from its front?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a vast, sun-drenched field containing red poppies, white daisies, and purple lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone execute sharp, banked turns and momentarily hover just inches above a vibrant yellow flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle tracking shot that glides swiftly above the ground following the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029799016192555428 + ], + "opposite_logprob_tokens": [ + -3.529798984527588 + ], + "logprob_reward": -0.029799016192555428, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029799016192555428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7356540560722351, + 0.05560886114835739, + 0.04776351526379585, + 0.01523007545620203, + 0.08617651462554932, + 0.021893417462706566, + 0.006284525152295828, + 0.008436837233603, + 0.005210352595895529, + 0.01774188131093979 + ], + "credit_map": [ + 0.06718774139881134, + 0.0532805509865284, + 0.1114698126912117, + 0.05687374994158745, + 0.07629188150167465, + 0.08581813424825668, + 0.11998622119426727, + 0.1647459715604782, + 0.0029139406979084015, + 0.010662417858839035, + 0.005296663846820593, + 0.005501894745975733, + 0.007480337284505367, + 0.010481627658009529, + 0.0056692082434892654, + 0.007602768484503031, + 0.004221154842525721, + 0.005566063802689314, + 0.00662538455799222, + 0.004096754360944033, + 0.006265508010983467, + 0.0077231829054653645, + 0.006519902031868696, + 0.006745560560375452, + 0.001276131602935493, + 0.0009813953656703234, + 0.0021673927549272776, + 0.0019849154632538557, + 0.0012490570079535246, + 0.0015577555168420076, + 0.0028967440593987703, + 0.0031166840344667435, + 0.007950704544782639, + 0.00756828673183918, + 0.01552229467779398, + 0.017261572182178497, + 0.005080059170722961, + 0.008453859016299248, + 0.011151866056025028, + 0.013187871314585209, + 0.0023117982782423496, + 0.002369860652834177, + 0.0029114303179085255, + 0.0036329785361886024, + 0.0020767590031027794, + 0.0020186523906886578, + 0.0025302560534328222, + 0.004041682928800583, + 0.0008431964088231325, + 0.0007289659697562456, + 0.001137985265813768, + 0.0009283024119213223, + 0.000703567115124315, + 0.0005782480584457517, + 0.0007554469048045576, + 0.0006088129593990743, + 0.0007600795361213386, + 0.0009977614972740412, + 0.0011384389363229275, + 0.0011532014468684793, + 0.001088909339159727, + 0.0008297794847749174, + 0.0011889784364029765, + 0.0012796880910173059, + 0.0007554320036433637, + 0.0004798730369657278, + 0.0007983636460267007, + 0.0005348905688151717, + 0.0003577604948077351, + 0.0007625414291396737, + 0.000494971580337733, + 0.0010265198070555925, + 0.0012144945794716477, + 0.0017391202272847295, + 0.0033573261462152004, + 0.0016088567208498716, + 0.0013035627780482173, + 0.0020561558194458485, + 0.0018350797472521663, + 0.004627285059541464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.029799016192555428 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, sleek quadcopter drone constructed of matte black carbon fiber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drone have four translucent propellers and a soft cyan LED glowing from its front?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a vast, sun-drenched field containing red poppies, white daisies, and purple lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone execute sharp, banked turns and momentarily hover just inches above a vibrant yellow flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle tracking shot that glides swiftly above the ground following the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038101546466350555 + ], + "opposite_logprob_tokens": [ + -3.2881014347076416 + ], + "logprob_reward": -0.038101546466350555, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038101546466350555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6611246466636658, + 0.05158976465463638, + 0.01912223920226097, + 0.02994738146662712, + 0.10164535045623779, + 0.030125662684440613, + 0.02761182188987732, + 0.0320097915828228, + 0.021440545096993446, + 0.025382760912179947 + ], + "credit_map": [ + 0.06093931943178177, + 0.05063585191965103, + 0.10255040973424911, + 0.0558365061879158, + 0.07060615718364716, + 0.07718007266521454, + 0.09711329638957977, + 0.14626306295394897, + 0.003682004753500223, + 0.009250744245946407, + 0.0060186986811459064, + 0.005646523088216782, + 0.0049815708771348, + 0.006067239213734865, + 0.008608104661107063, + 0.00733488705009222, + 0.0018342059338465333, + 0.0019748189952224493, + 0.002708673244342208, + 0.0018932819366455078, + 0.0022073457948863506, + 0.0027027097530663013, + 0.0024638937320560217, + 0.0033373108599334955, + 0.002751610241830349, + 0.002407294465228915, + 0.004603226203471422, + 0.0032842159271240234, + 0.0030574838165193796, + 0.0034398315474390984, + 0.004834621213376522, + 0.005569098982959986, + 0.008138365112245083, + 0.01522101741284132, + 0.011126827448606491, + 0.017309442162513733, + 0.010285292752087116, + 0.010261580348014832, + 0.012951762415468693, + 0.016351081430912018, + 0.003162401495501399, + 0.0029223167803138494, + 0.004135897383093834, + 0.003967897035181522, + 0.004678762052208185, + 0.003540441393852234, + 0.004290435463190079, + 0.003427513176575303, + 0.002728472463786602, + 0.0031236079521477222, + 0.005302353296428919, + 0.003970042336732149, + 0.002873716875910759, + 0.0029320595785975456, + 0.0034026685170829296, + 0.0032789024990051985, + 0.0033578791189938784, + 0.004533275496214628, + 0.005682799499481916, + 0.0041198330000042915, + 0.0028829574584960938, + 0.0025604425463825464, + 0.004218017216771841, + 0.004654590971767902, + 0.0020549704786390066, + 0.0016238903626799583, + 0.003655312117189169, + 0.0021974737755954266, + 0.002064452273771167, + 0.0031215467024594545, + 0.002661977196112275, + 0.004060924518853426, + 0.0027592885307967663, + 0.002481479197740555, + 0.004109655506908894, + 0.0031941020861268044, + 0.0025140619836747646, + 0.0030476623214781284, + 0.003488679649308324, + 0.003787832800298929 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.038101546466350555 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, sleek quadcopter drone constructed of matte black carbon fiber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drone have four translucent propellers and a soft cyan LED glowing from its front?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a vast, sun-drenched field containing red poppies, white daisies, and purple lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone execute sharp, banked turns and momentarily hover just inches above a vibrant yellow flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle tracking shot that glides swiftly above the ground following the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020573198795318604 + ], + "opposite_logprob_tokens": [ + -3.895573139190674 + ], + "logprob_reward": -0.020573198795318604, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020573198795318604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7869466543197632, + 0.036112748086452484, + 0.015036724507808685, + 0.017487626522779465, + 0.04332713410258293, + 0.017931533977389336, + 0.012674177065491676, + 0.039761368185281754, + 0.0107142748311162, + 0.020007720217108727 + ], + "credit_map": [ + 0.07093683630228043, + 0.060627587139606476, + 0.11107916384935379, + 0.06389357149600983, + 0.083790622651577, + 0.09486778825521469, + 0.12171723693609238, + 0.1800338476896286, + 0.0031432549003511667, + 0.004150012973695993, + 0.004985657054930925, + 0.004123684484511614, + 0.003834577277302742, + 0.005115853622555733, + 0.0057038008235394955, + 0.00505590345710516, + 0.0011166566982865334, + 0.0014906899305060506, + 0.0018594582797959447, + 0.001104892697185278, + 0.0028933719731867313, + 0.002543141134083271, + 0.0016632588813081384, + 0.0023652545642107725, + 0.0014110840857028961, + 0.0013458470348268747, + 0.002441809279844165, + 0.0022965462412685156, + 0.001641424954868853, + 0.0021470317151397467, + 0.003285395447164774, + 0.002918486250564456, + 0.004070811439305544, + 0.007935934700071812, + 0.004743773955851793, + 0.005499027203768492, + 0.004819819238036871, + 0.0050713433884084225, + 0.006613776553422213, + 0.004572648089379072, + 0.0015212050639092922, + 0.0017151107313111424, + 0.0025912155397236347, + 0.003394264495000243, + 0.0024448451586067677, + 0.0021011342760175467, + 0.0024760339874774218, + 0.0016877255402505398, + 0.0018399386899545789, + 0.0017399381613358855, + 0.0026601410936564207, + 0.0013480414636433125, + 0.0012943812180310488, + 0.0010276453103870153, + 0.0014347985852509737, + 0.0013292927760630846, + 0.0038942887913435698, + 0.004368486814200878, + 0.0075553846545517445, + 0.004889634903520346, + 0.003535610856488347, + 0.003592117689549923, + 0.005418777000159025, + 0.00650706747546792, + 0.0012298465007916093, + 0.001173906959593296, + 0.001452607219107449, + 0.0012446284526959062, + 0.0013294249074533582, + 0.0014798693591728806, + 0.0012867770856246352, + 0.0015172152779996395, + 0.001903614611364901, + 0.0018686284311115742, + 0.003500960301607847, + 0.0023238020949065685, + 0.00181505735963583, + 0.002679239958524704, + 0.0029121385887265205, + 0.003004278987646103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.020573198795318604 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_13/index_tensor([6493, 6493])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_13/index_tensor([6493, 6493])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2d7e5887b7f75eba220d10d6f83adf626af41d48 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_13/index_tensor([6493, 6493])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, steady drone pans gracefully around a majestic ancient pagoda, perched atop a mist-shrouded mountain. The setting sun sinks into the distant mountains, painting the sky a warm gradient of gold and deep orange. Its low-angled light sweeps across the pagoda like a layer of flowing honey. The multi-story pagoda is constructed of weathered, bluish-gray brick. The rough texture, minute cracks, and scattered wisps of verdant moss in the sidelight give it a sculptural, three-dimensional quality. The camera pans across the delicate eaves and brackets. Ancient bronze bells hang from the eaves, still in the breeze, gleaming with the last rays of the setting sun. As the camera pans, the tower's long shadow slowly shifts across the hillside, permeating the entire scene with a solemn, tranquil, and timeless atmosphere, expressed in warm, deep tones.", + "seed": 27062, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_13/index_6493_seed_27062_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.027754477248527108, + "vlm_logprob": -0.027754477248527108, + "vlm_yes_no_margin": 4.775, + "vlm_token_credit": -0.027754477248527108, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a multi-story ancient pagoda visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pagoda constructed of weathered bluish-gray brick with visible cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ancient bronze bells hanging from the pagoda's eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a warm gradient of gold and deep orange due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan around the pagoda?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010182439582422376 + ], + "opposite_logprob_tokens": [ + -8.126018524169922 + ], + "logprob_reward": -0.0010182439582422376, + "yes_no_margin_reward": 8.125, + "token_credit_reward": -0.0010182439582422376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23316000401973724, + 0.0516822449862957, + 0.08579486608505249, + 0.0664360374212265, + 0.055481959134340286, + 0.12482001632452011, + 0.041877973824739456, + 0.19139961898326874, + 0.06788446009159088, + 0.0814628154039383 + ], + "credit_map": [ + 0.0134741747751832, + 0.06800232827663422, + 0.0228168573230505, + 0.029316099360585213, + 0.054397184401750565, + 0.010587310418486595, + 0.021516993641853333, + 0.013049056753516197, + 0.005529964342713356, + 0.008596931584179401, + 0.007882291451096535, + 0.006143800914287567, + 0.007309960201382637, + 0.004186902195215225, + 0.006249555852264166, + 0.005782837513834238, + 0.0076939756982028484, + 0.010483050718903542, + 0.00912779476493597, + 0.018523411825299263, + 0.016940182074904442, + 0.005553757306188345, + 0.009173769503831863, + 0.008298917673528194, + 0.010490593500435352, + 0.007322721183300018, + 0.006814263295382261, + 0.012238623574376106, + 0.00635142857208848, + 0.0038095617201179266, + 0.010766208171844482, + 0.008642631582915783, + 0.004794423934072256, + 0.005971373990178108, + 0.007707845885306597, + 0.008962152525782585, + 0.005885193590074778, + 0.004653047304600477, + 0.006830967962741852, + 0.01067694928497076, + 0.007849231362342834, + 0.010216210037469864, + 0.014063346199691296, + 0.01832049898803234, + 0.017646698281168938, + 0.018311141058802605, + 0.028947656974196434, + 0.009465230628848076, + 0.00384520017541945, + 0.004207897000014782, + 0.0062028467655181885, + 0.005432432517409325, + 0.004638325423002243, + 0.008543179370462894, + 0.005479707848280668, + 0.003528383793309331, + 0.023995444178581238, + 0.013465884141623974, + 0.05356660112738609, + 0.02414170280098915, + 0.02074926719069481, + 0.012264368124306202, + 0.02527889609336853, + 0.01793743669986725, + 0.005390514619648457, + 0.0060933781787753105, + 0.014949444681406021, + 0.009076325222849846, + 0.00859910063445568, + 0.00550894346088171, + 0.011191909201443195, + 0.007074839901179075, + 0.008307648822665215, + 0.014115402474999428, + 0.01088822539895773, + 0.00842492375522852, + 0.01590091735124588, + 0.004421570338308811, + 0.013345824554562569, + 0.006058293394744396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.0010182439582422376 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a multi-story ancient pagoda visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pagoda constructed of weathered bluish-gray brick with visible cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ancient bronze bells hanging from the pagoda's eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a warm gradient of gold and deep orange due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan around the pagoda?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012552423402667046 + ], + "opposite_logprob_tokens": [ + -4.387552261352539 + ], + "logprob_reward": -0.012552423402667046, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012552423402667046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08602677285671234, + 0.02781962789595127, + 0.2513432204723358, + 0.20718304812908173, + 0.07239288836717606, + 0.13113655149936676, + 0.030054597184062004, + 0.12760575115680695, + 0.027916796505451202, + 0.038520727306604385 + ], + "credit_map": [ + 0.006391130853444338, + 0.02104809507727623, + 0.014122601598501205, + 0.01052421797066927, + 0.012515134178102016, + 0.004499569535255432, + 0.00925328116863966, + 0.0076727368868887424, + 0.002384249120950699, + 0.005445902701467276, + 0.00288832513615489, + 0.004415417090058327, + 0.0032829949632287025, + 0.0017921943217515945, + 0.004453218542039394, + 0.0031573278829455376, + 0.021188627928495407, + 0.02889455296099186, + 0.02194250375032425, + 0.043367382138967514, + 0.07605469226837158, + 0.017415644600987434, + 0.024654071778059006, + 0.017825735732913017, + 0.009693032130599022, + 0.016932476311922073, + 0.011022794991731644, + 0.047621507197618484, + 0.07773012667894363, + 0.010439741425216198, + 0.019874760881066322, + 0.013868610374629498, + 0.00666831387206912, + 0.008287923410534859, + 0.008330762386322021, + 0.011144955642521381, + 0.010257983580231667, + 0.006103672087192535, + 0.01190666388720274, + 0.009692616760730743, + 0.009500902146100998, + 0.00883143674582243, + 0.011961787939071655, + 0.02021610178053379, + 0.013300325721502304, + 0.014524323865771294, + 0.044481758028268814, + 0.008319906890392303, + 0.002035028301179409, + 0.004095933400094509, + 0.0041610379703342915, + 0.003943804185837507, + 0.004993967246264219, + 0.0039091333746910095, + 0.004424003418534994, + 0.0024916904512792826, + 0.013881408609449863, + 0.012886750511825085, + 0.013746984302997589, + 0.019276686012744904, + 0.015146131627261639, + 0.014816557057201862, + 0.018394874408841133, + 0.01945636048913002, + 0.0022760562133044004, + 0.0031072660349309444, + 0.003891098778694868, + 0.004882977809756994, + 0.00450907601043582, + 0.0017986613092944026, + 0.005232424009591341, + 0.0022192352917045355, + 0.0036494513042271137, + 0.005888250190764666, + 0.004856589715927839, + 0.004450634587556124, + 0.006561376620084047, + 0.0022264893632382154, + 0.007034509442746639, + 0.003853423520922661 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 5 + ], + "selected_reward": -0.012552423402667046 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a multi-story ancient pagoda visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pagoda constructed of weathered bluish-gray brick with visible cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ancient bronze bells hanging from the pagoda's eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a warm gradient of gold and deep orange due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan around the pagoda?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897523045539856 + ], + "opposite_logprob_tokens": [ + -2.578975200653076 + ], + "logprob_reward": -0.07897523045539856, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897523045539856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1632261872291565, + 0.027840835973620415, + 0.08512777835130692, + 0.13937756419181824, + 0.0550796240568161, + 0.09817899018526077, + 0.053733207285404205, + 0.19410070776939392, + 0.10030350089073181, + 0.08303157240152359 + ], + "credit_map": [ + 0.009650210849940777, + 0.024345921352505684, + 0.031685762107372284, + 0.01440314669162035, + 0.03948250040411949, + 0.009805424138903618, + 0.01566818170249462, + 0.018185047432780266, + 0.0020803085062652826, + 0.005616968963295221, + 0.0030406650621443987, + 0.004650226328521967, + 0.0034192288294434547, + 0.0017530402401462197, + 0.004057429730892181, + 0.0032229695934802294, + 0.007942604832351208, + 0.008774494752287865, + 0.0152021124958992, + 0.012288651429116726, + 0.014830389991402626, + 0.0062125008553266525, + 0.009935954585671425, + 0.009941071271896362, + 0.022908952087163925, + 0.010295749641954899, + 0.040378935635089874, + 0.020263703539967537, + 0.017152002081274986, + 0.007873806171119213, + 0.010953259654343128, + 0.00955115631222725, + 0.0052212863229215145, + 0.006733051501214504, + 0.005555275361984968, + 0.009846110828220844, + 0.009034279733896255, + 0.004255546256899834, + 0.008488606661558151, + 0.005945470649749041, + 0.009505968540906906, + 0.010681528598070145, + 0.00897263828665018, + 0.01604500412940979, + 0.012523530051112175, + 0.012723084539175034, + 0.019871177151799202, + 0.00785606075078249, + 0.004137097857892513, + 0.008729381486773491, + 0.005502756219357252, + 0.006721164099872112, + 0.009526937268674374, + 0.0053391666151583195, + 0.00830803718417883, + 0.005468669347465038, + 0.02595999650657177, + 0.015250494703650475, + 0.045619189739227295, + 0.023185614496469498, + 0.01894206926226616, + 0.02141943946480751, + 0.021588625386357307, + 0.0221352931112051, + 0.008819655515253544, + 0.012523874640464783, + 0.013560780324041843, + 0.012895261868834496, + 0.014960630796849728, + 0.0077755325473845005, + 0.019742833450436592, + 0.010024940595030785, + 0.008095628581941128, + 0.01067486684769392, + 0.013679315336048603, + 0.013078421354293823, + 0.013115597888827324, + 0.0048403069376945496, + 0.011093618348240852, + 0.00845381896942854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.07897523045539856 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a multi-story ancient pagoda visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pagoda constructed of weathered bluish-gray brick with visible cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ancient bronze bells hanging from the pagoda's eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a warm gradient of gold and deep orange due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan around the pagoda?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031891223043203354 + ], + "opposite_logprob_tokens": [ + -5.7531890869140625 + ], + "logprob_reward": -0.0031891223043203354, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031891223043203354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27920979261398315, + 0.07038228213787079, + 0.07168816030025482, + 0.05511137843132019, + 0.05380037799477577, + 0.1112191304564476, + 0.1476762294769287, + 0.0880977064371109, + 0.09598485380411148, + 0.026830120012164116 + ], + "credit_map": [ + 0.014705446548759937, + 0.04535432532429695, + 0.04547690972685814, + 0.02260667271912098, + 0.08158648014068604, + 0.017982950434088707, + 0.024242518469691277, + 0.027254460379481316, + 0.006512025836855173, + 0.015254831872880459, + 0.008950048126280308, + 0.008302665315568447, + 0.009386681951582432, + 0.005147191695868969, + 0.008093987591564655, + 0.008734844624996185, + 0.008237238973379135, + 0.0060848668217659, + 0.016339482739567757, + 0.012073483318090439, + 0.007498836610466242, + 0.004892003256827593, + 0.007872845977544785, + 0.0086893979460001, + 0.0031434488482773304, + 0.005545184947550297, + 0.005317156668752432, + 0.007071414031088352, + 0.01931486651301384, + 0.00355628808028996, + 0.006372098810970783, + 0.00479091377928853, + 0.00735356193035841, + 0.0036180326715111732, + 0.0074288626201450825, + 0.010004724375903606, + 0.0038848738186061382, + 0.00611014012247324, + 0.009080679155886173, + 0.006319504696875811, + 0.010873074643313885, + 0.0076067568734288216, + 0.017602797597646713, + 0.014757419936358929, + 0.008396695367991924, + 0.023663971573114395, + 0.019249016419053078, + 0.009069390594959259, + 0.014884022995829582, + 0.01195103581994772, + 0.02115628309547901, + 0.012729418464004993, + 0.01188669539988041, + 0.041957832872867584, + 0.02207525447010994, + 0.011035672388970852, + 0.01357910968363285, + 0.006583711132407188, + 0.023443594574928284, + 0.011472785845398903, + 0.008935874328017235, + 0.0047938344068825245, + 0.011078284122049809, + 0.008210508152842522, + 0.009584281593561172, + 0.008679126389324665, + 0.02455330453813076, + 0.010656386613845825, + 0.007538831327110529, + 0.009195614606142044, + 0.011516892351210117, + 0.01426040381193161, + 0.002476912224665284, + 0.0025743122678250074, + 0.003104201052337885, + 0.003687439486384392, + 0.0051486617885529995, + 0.0020462688989937305, + 0.005245602689683437, + 0.0025467192754149437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.0031891223043203354 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a multi-story ancient pagoda visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pagoda constructed of weathered bluish-gray brick with visible cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ancient bronze bells hanging from the pagoda's eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a warm gradient of gold and deep orange due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan around the pagoda?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04303736612200737 + ], + "opposite_logprob_tokens": [ + -3.1680374145507812 + ], + "logprob_reward": -0.04303736612200737, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04303736612200737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09214000403881073, + 0.024915926158428192, + 0.10948269069194794, + 0.10644617676734924, + 0.03920796513557434, + 0.27616846561431885, + 0.09366663545370102, + 0.19383202493190765, + 0.02549608051776886, + 0.03864400461316109 + ], + "credit_map": [ + 0.0057186647318303585, + 0.030015433207154274, + 0.006336305756121874, + 0.015015006996691227, + 0.010959663428366184, + 0.004282387439161539, + 0.012150895781815052, + 0.007661648094654083, + 0.0022011359687894583, + 0.0037341017741709948, + 0.003303838660940528, + 0.00400895718485117, + 0.0028880045283585787, + 0.0018185328226536512, + 0.004435298964381218, + 0.0025260564871132374, + 0.008763426914811134, + 0.010773380286991596, + 0.020362064242362976, + 0.012820056639611721, + 0.03149901703000069, + 0.007014204747974873, + 0.011071163229644299, + 0.0071793850511312485, + 0.009928218089044094, + 0.009769280441105366, + 0.02452899143099785, + 0.01361481286585331, + 0.013400241732597351, + 0.005709637887775898, + 0.014446447603404522, + 0.01504854578524828, + 0.002584842499345541, + 0.0029416207689791918, + 0.0054439278319478035, + 0.007312728092074394, + 0.003381693037226796, + 0.0034616009797900915, + 0.005013396497815847, + 0.009068155661225319, + 0.02844061516225338, + 0.022777706384658813, + 0.02238348312675953, + 0.04604734480381012, + 0.04151017218828201, + 0.03435433283448219, + 0.058943089097738266, + 0.021711722016334534, + 0.008837553672492504, + 0.00931370910257101, + 0.015434157103300095, + 0.012034469284117222, + 0.013257598504424095, + 0.015751035884022713, + 0.01198987103998661, + 0.0070482403971254826, + 0.033467311412096024, + 0.01562988944351673, + 0.048632699996232986, + 0.022797154262661934, + 0.01701710745692253, + 0.015974964946508408, + 0.024553800001740456, + 0.015759091824293137, + 0.0033132892567664385, + 0.0020720376633107662, + 0.006072002463042736, + 0.0030218088068068027, + 0.002735246205702424, + 0.002082440536469221, + 0.0038957269862294197, + 0.002303528133779764, + 0.004600523505359888, + 0.007194398436695337, + 0.004491157364100218, + 0.003676386782899499, + 0.007241464219987392, + 0.0023075377102941275, + 0.0059696706011891365, + 0.0031628648284822702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 2 + ], + "selected_reward": -0.04303736612200737 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, steady drone pans gracefully around a majestic ancient pagoda, perched atop a mist-shrouded mountain. The setting sun sinks into the distant mountains, painting the sky a warm gradient of gold and deep orange. Its low-angled light sweeps across the pagoda like a layer of flowing honey. The multi-story pagoda is constructed of weathered, bluish-gray brick. The rough texture, minute cracks, and scattered wisps of verdant moss in the sidelight give it a sculptural, three-dimensional quality. The camera pans across the delicate eaves and brackets. Ancient bronze bells hang from the eaves, still in the breeze, gleaming with the last rays of the setting sun. As the camera pans, the tower's long shadow slowly shifts across the hillside, permeating the entire scene with a solemn, tranquil, and timeless atmosphere, expressed in warm, deep tones.", + "seed": 27063, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_13/index_6493_seed_27063_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.049783791927620766, + "vlm_logprob": -0.049783791927620766, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.049783791927620766, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a multi-story ancient pagoda visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pagoda constructed of weathered bluish-gray brick with visible cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ancient bronze bells hanging from the pagoda's eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a warm gradient of gold and deep orange due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan around the pagoda?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011120808776468039 + ], + "opposite_logprob_tokens": [ + -8.12611198425293 + ], + "logprob_reward": -0.0011120808776468039, + "yes_no_margin_reward": 8.125, + "token_credit_reward": -0.0011120808776468039, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08961507678031921, + 0.05491304025053978, + 0.02493610978126526, + 0.13331696391105652, + 0.08383043855428696, + 0.07801873236894608, + 0.05049910768866539, + 0.04874340817332268, + 0.2814483046531677, + 0.1546788364648819 + ], + "credit_map": [ + 0.007700610905885696, + 0.011032667942345142, + 0.008724975399672985, + 0.012885402888059616, + 0.008723023347556591, + 0.012939509004354477, + 0.01110717561095953, + 0.016501720994710922, + 0.005036241840571165, + 0.007596136536449194, + 0.006838346365839243, + 0.006368720903992653, + 0.007461779285222292, + 0.007174669299274683, + 0.007004621438682079, + 0.007432527374476194, + 0.0020571795757859945, + 0.0025510508567094803, + 0.0032991997431963682, + 0.0031812316738069057, + 0.0039468808099627495, + 0.0033266597893089056, + 0.003389605088159442, + 0.003184304339811206, + 0.011289321817457676, + 0.015780150890350342, + 0.02276706136763096, + 0.01689177379012108, + 0.013919902965426445, + 0.014638141728937626, + 0.018807481974363327, + 0.019223136827349663, + 0.005066705867648125, + 0.021681025624275208, + 0.012699226848781109, + 0.011059978976845741, + 0.007605113089084625, + 0.006454353220760822, + 0.012076269835233688, + 0.0071877664886415005, + 0.005745237227529287, + 0.01020453218370676, + 0.01041504554450512, + 0.01323804073035717, + 0.010583946481347084, + 0.004964274819940329, + 0.01257962454110384, + 0.010288036428391933, + 0.0029731078539043665, + 0.005671042948961258, + 0.005922685842961073, + 0.009808935225009918, + 0.006414173636585474, + 0.004056449513882399, + 0.008822135627269745, + 0.006830581929534674, + 0.0033437085803598166, + 0.007727390620857477, + 0.00562827056273818, + 0.006617852486670017, + 0.006015989929437637, + 0.002601152053102851, + 0.008667889051139355, + 0.008141162805259228, + 0.03169890120625496, + 0.037275612354278564, + 0.046577222645282745, + 0.045913975685834885, + 0.026674777269363403, + 0.01725728064775467, + 0.03386877104640007, + 0.04218176007270813, + 0.012780239805579185, + 0.015931159257888794, + 0.019011663272976875, + 0.014235582202672958, + 0.022682661190629005, + 0.010911097750067711, + 0.0360933281481266, + 0.023033106699585915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.0011120808776468039 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a multi-story ancient pagoda visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pagoda constructed of weathered bluish-gray brick with visible cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ancient bronze bells hanging from the pagoda's eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a warm gradient of gold and deep orange due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan around the pagoda?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980896644294262 + ], + "opposite_logprob_tokens": [ + -3.52980899810791 + ], + "logprob_reward": -0.02980896644294262, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980896644294262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0696527436375618, + 0.04983905330300331, + 0.025344576686620712, + 0.3006440997123718, + 0.10814238339662552, + 0.0634579211473465, + 0.06755373626947403, + 0.08560509979724884, + 0.07833386957645416, + 0.1514265388250351 + ], + "credit_map": [ + 0.004793841857463121, + 0.00628878315910697, + 0.004441543482244015, + 0.008358342573046684, + 0.005573416594415903, + 0.008038101717829704, + 0.02138669788837433, + 0.010772015899419785, + 0.004595541395246983, + 0.006472117267549038, + 0.006170568522065878, + 0.006409484893083572, + 0.00587001908570528, + 0.0068456330336630344, + 0.006154046859592199, + 0.007321644574403763, + 0.001867608749307692, + 0.0031368827912956476, + 0.003679684130474925, + 0.003323030425235629, + 0.0031178449280560017, + 0.0045459880493581295, + 0.0022269575856626034, + 0.003446579212322831, + 0.02163408137857914, + 0.030082685872912407, + 0.058984145522117615, + 0.034204985946416855, + 0.03618650510907173, + 0.03187358006834984, + 0.05139627680182457, + 0.03628183901309967, + 0.007979070767760277, + 0.011338397860527039, + 0.016607681289315224, + 0.016829000785946846, + 0.011438933201134205, + 0.012364493682980537, + 0.021696440875530243, + 0.009888367727398872, + 0.003928363788872957, + 0.013614698313176632, + 0.008129730820655823, + 0.010844309814274311, + 0.006978674326092005, + 0.0045186118222773075, + 0.009294875897467136, + 0.006148659624159336, + 0.0033453635405749083, + 0.010204804129898548, + 0.008616240695118904, + 0.01139982882887125, + 0.00685314042493701, + 0.006571958772838116, + 0.01019961852580309, + 0.010362773202359676, + 0.006871243007481098, + 0.010545163415372372, + 0.013531719334423542, + 0.010039576329290867, + 0.011324804276227951, + 0.005074609071016312, + 0.015083357691764832, + 0.013134630396962166, + 0.005951345898211002, + 0.010404662229120731, + 0.01038484275341034, + 0.011046944186091423, + 0.009518599137663841, + 0.004715993534773588, + 0.012864071875810623, + 0.013447407633066177, + 0.012534263543784618, + 0.01522785797715187, + 0.02254605107009411, + 0.026283886283636093, + 0.01897270232439041, + 0.006537586450576782, + 0.02726053074002266, + 0.022063659504055977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 4 + ], + "selected_reward": -0.02980896644294262 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a multi-story ancient pagoda visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pagoda constructed of weathered bluish-gray brick with visible cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ancient bronze bells hanging from the pagoda's eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a warm gradient of gold and deep orange due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan around the pagoda?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292894184589386 + ], + "opposite_logprob_tokens": [ + -2.237928867340088 + ], + "logprob_reward": -0.11292894184589386, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292894184589386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06783099472522736, + 0.04667813703417778, + 0.0918172150850296, + 0.3124825358390808, + 0.09145995229482651, + 0.037590060383081436, + 0.13620156049728394, + 0.06656496971845627, + 0.08400901407003403, + 0.06536559760570526 + ], + "credit_map": [ + 0.0055784787982702255, + 0.007977712899446487, + 0.005511752795428038, + 0.009480227716267109, + 0.00634978199377656, + 0.004284091293811798, + 0.018398135900497437, + 0.010250807739794254, + 0.004214457701891661, + 0.00802040658891201, + 0.003605668665841222, + 0.006893405225127935, + 0.004794331267476082, + 0.006142409052699804, + 0.005930623505264521, + 0.0070768315345048904, + 0.006791231222450733, + 0.01207897998392582, + 0.01143592782318592, + 0.01815227046608925, + 0.010598886758089066, + 0.010760686360299587, + 0.010976179502904415, + 0.011023047380149364, + 0.020739365369081497, + 0.033443011343479156, + 0.04670393466949463, + 0.036592718213796616, + 0.03518782556056976, + 0.03158402442932129, + 0.05493757873773575, + 0.05329408869147301, + 0.0063083600252866745, + 0.012891091406345367, + 0.014605140313506126, + 0.011565782129764557, + 0.008177799172699451, + 0.014669687487185001, + 0.014751661568880081, + 0.008490427397191525, + 0.0030361050739884377, + 0.005028536543250084, + 0.005236611235886812, + 0.0057641747407615185, + 0.005425555631518364, + 0.0032267351634800434, + 0.005497212056070566, + 0.0043751331977546215, + 0.007085456047207117, + 0.027991201728582382, + 0.016026394441723824, + 0.03247177228331566, + 0.015692953020334244, + 0.006383249536156654, + 0.014499470591545105, + 0.016051067039370537, + 0.004576174542307854, + 0.007776539772748947, + 0.007674524560570717, + 0.0070444936864078045, + 0.009010184556245804, + 0.004588757641613483, + 0.010852154344320297, + 0.015042145736515522, + 0.007109237834811211, + 0.009925402700901031, + 0.014077511616051197, + 0.013371901586651802, + 0.008198232389986515, + 0.00730849290266633, + 0.01241868082433939, + 0.011599553748965263, + 0.005120486952364445, + 0.007603355683386326, + 0.0073426892049610615, + 0.007775024510920048, + 0.0096905417740345, + 0.003780155209824443, + 0.013217175379395485, + 0.010836167261004448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.11292894184589386 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a multi-story ancient pagoda visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pagoda constructed of weathered bluish-gray brick with visible cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ancient bronze bells hanging from the pagoda's eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a warm gradient of gold and deep orange due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan around the pagoda?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016066256910562515 + ], + "opposite_logprob_tokens": [ + -4.141066074371338 + ], + "logprob_reward": -0.016066256910562515, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016066256910562515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15881609916687012, + 0.026166100054979324, + 0.03445880115032196, + 0.1266573816537857, + 0.17876432836055756, + 0.022157302126288414, + 0.04321944713592529, + 0.05147353187203407, + 0.27044355869293213, + 0.08784342557191849 + ], + "credit_map": [ + 0.0156286358833313, + 0.010139006190001965, + 0.007937660440802574, + 0.021318485960364342, + 0.01175462082028389, + 0.024203207343816757, + 0.04381335526704788, + 0.024021128192543983, + 0.0021212249994277954, + 0.0035943398252129555, + 0.0030826248694211245, + 0.0031135878525674343, + 0.0028567304834723473, + 0.0038431279826909304, + 0.0038682008162140846, + 0.003686262294650078, + 0.0030401928815990686, + 0.0030771582387387753, + 0.005563030950725079, + 0.003593892091885209, + 0.004185202997177839, + 0.005293914116919041, + 0.003947334364056587, + 0.005758074577897787, + 0.009259859099984169, + 0.014846563339233398, + 0.02043277956545353, + 0.015507214702665806, + 0.013530031777918339, + 0.01252745185047388, + 0.021826334297657013, + 0.018727155402302742, + 0.01226628478616476, + 0.017164597287774086, + 0.028214238584041595, + 0.022184697911143303, + 0.01769176498055458, + 0.03367968648672104, + 0.026308799162507057, + 0.02125425823032856, + 0.0016477213939651847, + 0.0038009348791092634, + 0.0028705024160444736, + 0.004487022291868925, + 0.0024706609547138214, + 0.001395510509610176, + 0.0028608739376068115, + 0.002624075161293149, + 0.0029385986272245646, + 0.0051344651728868484, + 0.00645874161273241, + 0.005697713233530521, + 0.006232398096472025, + 0.002925811568275094, + 0.00907055102288723, + 0.004761169198900461, + 0.00369248166680336, + 0.00819388497620821, + 0.006040041800588369, + 0.005984350107610226, + 0.008140355348587036, + 0.0032469069119542837, + 0.0073870145715773106, + 0.008788494393229485, + 0.03025086782872677, + 0.03447451442480087, + 0.043145183473825455, + 0.04178204759955406, + 0.02583736553788185, + 0.016835937276482582, + 0.038408804684877396, + 0.03970884904265404, + 0.006634332239627838, + 0.007680064998567104, + 0.012864552438259125, + 0.016009261831641197, + 0.011276274919509888, + 0.006401511374861002, + 0.011822606436908245, + 0.015154821798205376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.016066256910562515 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a multi-story ancient pagoda visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pagoda constructed of weathered bluish-gray brick with visible cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ancient bronze bells hanging from the pagoda's eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a warm gradient of gold and deep orange due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan around the pagoda?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08900271356105804 + ], + "opposite_logprob_tokens": [ + -2.4640026092529297 + ], + "logprob_reward": -0.08900271356105804, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08900271356105804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04480898007750511, + 0.016721082851290703, + 0.012745337560772896, + 0.07368618249893188, + 0.12333473563194275, + 0.03956836834549904, + 0.04640207812190056, + 0.09523800760507584, + 0.044468704611063004, + 0.5030264854431152 + ], + "credit_map": [ + 0.0035505632404237986, + 0.004328921902924776, + 0.0041792611591517925, + 0.006308775395154953, + 0.006627986207604408, + 0.003359251655638218, + 0.007488508243113756, + 0.00896571297198534, + 0.001442023436538875, + 0.00221101101487875, + 0.0022927208337932825, + 0.0020365293603390455, + 0.0022644586861133575, + 0.0019043778302147985, + 0.0018728590803220868, + 0.0026971029583364725, + 0.0010298194829374552, + 0.00199028174392879, + 0.0014553568325936794, + 0.001579812029376626, + 0.0016924672527238727, + 0.001803276943974197, + 0.0017512984341010451, + 0.0014430253067985177, + 0.004946876782923937, + 0.0074824378825724125, + 0.013227162882685661, + 0.007970541715621948, + 0.008416294120252132, + 0.009412962011992931, + 0.01372404396533966, + 0.008505864068865776, + 0.008766783401370049, + 0.011337034404277802, + 0.02144113928079605, + 0.014056426472961903, + 0.012943093664944172, + 0.024488095194101334, + 0.015229376032948494, + 0.015072792768478394, + 0.002612926298752427, + 0.007269422523677349, + 0.005086557473987341, + 0.008339131250977516, + 0.0046480935998260975, + 0.002210349775850773, + 0.005145540926605463, + 0.004256343934684992, + 0.002673768438398838, + 0.008863527327775955, + 0.005631573032587767, + 0.012364082969725132, + 0.004324665293097496, + 0.00241082813590765, + 0.005136771127581596, + 0.004996859468519688, + 0.00614116620272398, + 0.016292739659547806, + 0.010505156591534615, + 0.011514594778418541, + 0.014458179473876953, + 0.0056261466816067696, + 0.012118400074541569, + 0.018581625074148178, + 0.003166286740452051, + 0.006107131484895945, + 0.0049884021282196045, + 0.0048258607275784016, + 0.005443761590868235, + 0.002340163802728057, + 0.008227578364312649, + 0.009369518607854843, + 0.03495122119784355, + 0.04457820579409599, + 0.08780895918607712, + 0.11565238982439041, + 0.05766670033335686, + 0.024701837450265884, + 0.062381695955991745, + 0.07528550177812576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 7 + ], + "selected_reward": -0.08900271356105804 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, steady drone pans gracefully around a majestic ancient pagoda, perched atop a mist-shrouded mountain. The setting sun sinks into the distant mountains, painting the sky a warm gradient of gold and deep orange. Its low-angled light sweeps across the pagoda like a layer of flowing honey. The multi-story pagoda is constructed of weathered, bluish-gray brick. The rough texture, minute cracks, and scattered wisps of verdant moss in the sidelight give it a sculptural, three-dimensional quality. The camera pans across the delicate eaves and brackets. Ancient bronze bells hang from the eaves, still in the breeze, gleaming with the last rays of the setting sun. As the camera pans, the tower's long shadow slowly shifts across the hillside, permeating the entire scene with a solemn, tranquil, and timeless atmosphere, expressed in warm, deep tones.", + "seed": 27064, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_13/index_6493_seed_27064_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0390496319392696, + "vlm_logprob": -0.0390496319392696, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.0390496319392696, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a multi-story ancient pagoda visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pagoda constructed of weathered bluish-gray brick with visible cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ancient bronze bells hanging from the pagoda's eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a warm gradient of gold and deep orange due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan around the pagoda?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0016767027555033565 + ], + "opposite_logprob_tokens": [ + -7.376676559448242 + ], + "logprob_reward": -0.0016767027555033565, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0016767027555033565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1388636976480484, + 0.063722625374794, + 0.033759381622076035, + 0.14883659780025482, + 0.10078651458024979, + 0.07438933104276657, + 0.08869384974241257, + 0.054350122809410095, + 0.03215070813894272, + 0.2644471824169159 + ], + "credit_map": [ + 0.03172536566853523, + 0.020693814381957054, + 0.020639240741729736, + 0.02033512480556965, + 0.011701484210789204, + 0.009746093302965164, + 0.014662878587841988, + 0.00935970526188612, + 0.007827697321772575, + 0.008553539402782917, + 0.012409292161464691, + 0.010264151729643345, + 0.003974941559135914, + 0.006763308774679899, + 0.006791172083467245, + 0.007138523738831282, + 0.0033068463671952486, + 0.006022448185831308, + 0.005914164241403341, + 0.006997282616794109, + 0.0024090842343866825, + 0.0025110398419201374, + 0.004011761862784624, + 0.0025867533404380083, + 0.017115062102675438, + 0.022189293056726456, + 0.02420475147664547, + 0.022128932178020477, + 0.013595428317785263, + 0.013199565932154655, + 0.02322644181549549, + 0.013177124783396721, + 0.0077287303283810616, + 0.014212751761078835, + 0.016377873718738556, + 0.013127489946782589, + 0.010785805061459541, + 0.008566412143409252, + 0.02086804062128067, + 0.009119406342506409, + 0.005694158840924501, + 0.013031944632530212, + 0.008464791812002659, + 0.009739061817526817, + 0.007774613797664642, + 0.007909266278147697, + 0.013967164792120457, + 0.007808328606188297, + 0.006721860263496637, + 0.013407533057034016, + 0.012651195749640465, + 0.01762493886053562, + 0.008503169752657413, + 0.006797348149120808, + 0.012612015940248966, + 0.010375777259469032, + 0.0023917772341519594, + 0.00873751100152731, + 0.003930565435439348, + 0.017336837947368622, + 0.005737971514463425, + 0.0036521318834275007, + 0.0064026303589344025, + 0.0061606974340975285, + 0.0015975793357938528, + 0.007388606201857328, + 0.0027435484807938337, + 0.005732551217079163, + 0.005300420336425304, + 0.0027882817666977644, + 0.003998434636741877, + 0.00260128709487617, + 0.01445977296680212, + 0.023574959486722946, + 0.019962022081017494, + 0.07113942503929138, + 0.029462188482284546, + 0.011114900931715965, + 0.081425741314888, + 0.013308187015354633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.0016767027555033565 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a multi-story ancient pagoda visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pagoda constructed of weathered bluish-gray brick with visible cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ancient bronze bells hanging from the pagoda's eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a warm gradient of gold and deep orange due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan around the pagoda?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043075960129499435 + ], + "opposite_logprob_tokens": [ + -3.1680760383605957 + ], + "logprob_reward": -0.043075960129499435, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043075960129499435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05656472221016884, + 0.025586415082216263, + 0.016605976969003677, + 0.02775803953409195, + 0.07103602588176727, + 0.04490678384900093, + 0.07725315541028976, + 0.08162225782871246, + 0.02868565171957016, + 0.5699809789657593 + ], + "credit_map": [ + 0.011291775852441788, + 0.00817912071943283, + 0.008322539739310741, + 0.008800780400633812, + 0.0048959809355437756, + 0.003969657700508833, + 0.006736702751368284, + 0.004368164110928774, + 0.0035281621385365725, + 0.0047697871923446655, + 0.003929710481315851, + 0.003995920531451702, + 0.0017879699589684606, + 0.0022294761147350073, + 0.002988356165587902, + 0.002357033547013998, + 0.0016772417584434152, + 0.0016135663026943803, + 0.00257103075273335, + 0.002134121023118496, + 0.001287052291445434, + 0.0025279077235609293, + 0.0018923081224784255, + 0.0029027496930211782, + 0.0031876328866928816, + 0.004884210880845785, + 0.003912997432053089, + 0.004168704152107239, + 0.002248492557555437, + 0.0028988819103688, + 0.003967869095504284, + 0.00248925038613379, + 0.0065229167230427265, + 0.011296057142317295, + 0.014029675163328648, + 0.009795241057872772, + 0.009019024670124054, + 0.005447953939437866, + 0.009900080971419811, + 0.005025078542530537, + 0.0035816559102386236, + 0.007365914527326822, + 0.003885566024109721, + 0.006290691439062357, + 0.004951030947268009, + 0.007147739175707102, + 0.005973968654870987, + 0.005710214376449585, + 0.003262522630393505, + 0.009500307962298393, + 0.007575587369501591, + 0.024171970784664154, + 0.0037158550694584846, + 0.005643888376653194, + 0.008050397969782352, + 0.015332621522247791, + 0.0035211192443966866, + 0.016569199040532112, + 0.0059249806217849255, + 0.025968588888645172, + 0.007393589708954096, + 0.004127722233533859, + 0.01087188720703125, + 0.007245171815156937, + 0.0011672921245917678, + 0.0044515542685985565, + 0.0027416327502578497, + 0.0062883044593036175, + 0.004272854421287775, + 0.0033596917055547237, + 0.0046951486729085445, + 0.001709172036498785, + 0.021918881684541702, + 0.0326959528028965, + 0.029793590307235718, + 0.1915120929479599, + 0.03414786234498024, + 0.014074196107685566, + 0.22063970565795898, + 0.025198670104146004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.043075960129499435 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a multi-story ancient pagoda visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pagoda constructed of weathered bluish-gray brick with visible cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ancient bronze bells hanging from the pagoda's eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a warm gradient of gold and deep orange due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan around the pagoda?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899109274148941 + ], + "opposite_logprob_tokens": [ + -2.578991174697876 + ], + "logprob_reward": -0.07899109274148941, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899109274148941, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1440337747335434, + 0.03777559846639633, + 0.04418834298849106, + 0.03321346268057823, + 0.041971661150455475, + 0.05130229517817497, + 0.0566612109541893, + 0.39025673270225525, + 0.021852312609553337, + 0.17874467372894287 + ], + "credit_map": [ + 0.03326192498207092, + 0.02097221650183201, + 0.020932979881763458, + 0.022276153787970543, + 0.013925003819167614, + 0.01002370286732912, + 0.01427441369742155, + 0.00836737360805273, + 0.006643913220614195, + 0.006782956887036562, + 0.006197985261678696, + 0.005591952707618475, + 0.0034480225294828415, + 0.0028377023991197348, + 0.0030163235496729612, + 0.0032567426096647978, + 0.004165520425885916, + 0.006443137302994728, + 0.005269497167319059, + 0.005628617946058512, + 0.005404279101639986, + 0.004756598267704248, + 0.006401189602911472, + 0.0061195045709609985, + 0.0031583011150360107, + 0.005027926526963711, + 0.0052155680023133755, + 0.00408698245882988, + 0.002691822126507759, + 0.0030949139036238194, + 0.006848966237157583, + 0.00308898463845253, + 0.0036946602631360292, + 0.00918684620410204, + 0.005165057722479105, + 0.006374530028551817, + 0.005322814919054508, + 0.0037164995446801186, + 0.005247786175459623, + 0.0032634648960083723, + 0.005652110558003187, + 0.0045747095718979836, + 0.00414690189063549, + 0.0053700851276516914, + 0.005629466846585274, + 0.01071231346577406, + 0.005347760394215584, + 0.009868951514363289, + 0.002905620029196143, + 0.007986370474100113, + 0.004958902485668659, + 0.014848793856799603, + 0.004249266814440489, + 0.003785829758271575, + 0.009080990217626095, + 0.008845435455441475, + 0.01826702058315277, + 0.060587868094444275, + 0.030245058238506317, + 0.13847887516021729, + 0.03308409824967384, + 0.021122734993696213, + 0.043382495641708374, + 0.04508854076266289, + 0.001227030879817903, + 0.005065028555691242, + 0.0016913708532229066, + 0.0037795165553689003, + 0.0038681034930050373, + 0.0015834198566153646, + 0.0028704022988677025, + 0.0017674412811174989, + 0.008803322911262512, + 0.01581532321870327, + 0.011929688975214958, + 0.05729978159070015, + 0.016782931983470917, + 0.009763729758560658, + 0.04900136962532997, + 0.009348522871732712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.07899109274148941 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a multi-story ancient pagoda visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pagoda constructed of weathered bluish-gray brick with visible cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ancient bronze bells hanging from the pagoda's eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a warm gradient of gold and deep orange due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan around the pagoda?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015112657565623522 + ], + "opposite_logprob_tokens": [ + -6.501511096954346 + ], + "logprob_reward": -0.0015112657565623522, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015112657565623522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11852764338254929, + 0.03921724110841751, + 0.021994484588503838, + 0.01922672614455223, + 0.044557057321071625, + 0.04106896370649338, + 0.033347297459840775, + 0.49008846282958984, + 0.04322371259331703, + 0.14874836802482605 + ], + "credit_map": [ + 0.016219060868024826, + 0.014931763522326946, + 0.016214989125728607, + 0.018763016909360886, + 0.008248847909271717, + 0.010804293677210808, + 0.017353611066937447, + 0.015992077067494392, + 0.006439079996198416, + 0.005772956181317568, + 0.006057380698621273, + 0.006537790410220623, + 0.0023905301932245493, + 0.0039151213131845, + 0.004488178063184023, + 0.0036162096075713634, + 0.002566794166341424, + 0.002815261483192444, + 0.0027444835286587477, + 0.003055618377402425, + 0.0016485188389196992, + 0.0024396188091486692, + 0.0033729965798556805, + 0.0033511959481984377, + 0.0011968244798481464, + 0.0023819073103368282, + 0.002665799343958497, + 0.0021436959505081177, + 0.001929454505443573, + 0.002806649310514331, + 0.0035308850929141045, + 0.0025715120136737823, + 0.003150834934785962, + 0.006492098793387413, + 0.00729768630117178, + 0.006410717964172363, + 0.004819490015506744, + 0.004403699655085802, + 0.0077601829543709755, + 0.004222352523356676, + 0.0038993130438029766, + 0.00571893947198987, + 0.004208595957607031, + 0.004432945512235165, + 0.004226303193718195, + 0.007432779297232628, + 0.00540403975173831, + 0.005746051669120789, + 0.0029706682544201612, + 0.004681550897657871, + 0.004354721866548061, + 0.006071436684578657, + 0.003151511773467064, + 0.004738567862659693, + 0.003392835846170783, + 0.003986007533967495, + 0.023194225504994392, + 0.0746321976184845, + 0.0379398949444294, + 0.17945711314678192, + 0.03826114907860756, + 0.02625073492527008, + 0.053445275872945786, + 0.056907862424850464, + 0.0020970581099390984, + 0.006280583795160055, + 0.004741516429930925, + 0.006759298034012318, + 0.0049558645114302635, + 0.004843646194785833, + 0.010095389559864998, + 0.0034503634087741375, + 0.005468518007546663, + 0.011130502447485924, + 0.007429803721606731, + 0.04900475591421127, + 0.012014497071504593, + 0.006148661486804485, + 0.04980817809700966, + 0.007743465714156628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.0015112657565623522 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a multi-story ancient pagoda visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pagoda constructed of weathered bluish-gray brick with visible cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ancient bronze bells hanging from the pagoda's eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a warm gradient of gold and deep orange due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan around the pagoda?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999313831329346 + ], + "opposite_logprob_tokens": [ + -2.694993019104004 + ], + "logprob_reward": -0.06999313831329346, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999313831329346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04547477513551712, + 0.033813346177339554, + 0.015760326758027077, + 0.022674985229969025, + 0.07620858401060104, + 0.13294468820095062, + 0.0851798951625824, + 0.5130571126937866, + 0.022700117900967598, + 0.05218623951077461 + ], + "credit_map": [ + 0.008774938993155956, + 0.007062830962240696, + 0.005828422494232655, + 0.005934266373515129, + 0.0041666459292173386, + 0.0038478602655231953, + 0.005754085257649422, + 0.004105721600353718, + 0.004575301427394152, + 0.006475522648543119, + 0.005604187957942486, + 0.005684680305421352, + 0.002689099172130227, + 0.003074203385040164, + 0.0031082131899893284, + 0.0026021329686045647, + 0.0018761873943731189, + 0.002490046201273799, + 0.002578780287876725, + 0.0023025842383503914, + 0.0010982987005263567, + 0.0018436794634908438, + 0.0015425954479724169, + 0.0020281539764255285, + 0.002281778259202838, + 0.0032090304885059595, + 0.0035573290660977364, + 0.003361606504768133, + 0.0020625784527510405, + 0.0027023032307624817, + 0.002661783015355468, + 0.002838574117049575, + 0.007607914507389069, + 0.01027578953653574, + 0.008962615393102169, + 0.010334539227187634, + 0.008520047180354595, + 0.00885326974093914, + 0.012320670299232006, + 0.009333721362054348, + 0.011185693554580212, + 0.018689027056097984, + 0.01597302220761776, + 0.01687106303870678, + 0.014557004906237125, + 0.020871547982096672, + 0.01459845993667841, + 0.02019885927438736, + 0.006617553066462278, + 0.014600270427763462, + 0.012257272377610207, + 0.016873568296432495, + 0.00795071106404066, + 0.007011497858911753, + 0.013092976994812489, + 0.006776034366339445, + 0.024544836953282356, + 0.07810009270906448, + 0.03965698927640915, + 0.18906672298908234, + 0.03897346556186676, + 0.02767195925116539, + 0.05467719957232475, + 0.06036576256155968, + 0.0013887089444324374, + 0.0036652525886893272, + 0.0021752319298684597, + 0.0033974754624068737, + 0.0033039634581655264, + 0.0015820859698578715, + 0.005067236255854368, + 0.0021201602648943663, + 0.0030015739612281322, + 0.007168395910412073, + 0.003677245695143938, + 0.013855982571840286, + 0.007399844937026501, + 0.004932086914777756, + 0.00893767736852169, + 0.0032134284265339375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.06999313831329346 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, steady drone pans gracefully around a majestic ancient pagoda, perched atop a mist-shrouded mountain. The setting sun sinks into the distant mountains, painting the sky a warm gradient of gold and deep orange. Its low-angled light sweeps across the pagoda like a layer of flowing honey. The multi-story pagoda is constructed of weathered, bluish-gray brick. The rough texture, minute cracks, and scattered wisps of verdant moss in the sidelight give it a sculptural, three-dimensional quality. The camera pans across the delicate eaves and brackets. Ancient bronze bells hang from the eaves, still in the breeze, gleaming with the last rays of the setting sun. As the camera pans, the tower's long shadow slowly shifts across the hillside, permeating the entire scene with a solemn, tranquil, and timeless atmosphere, expressed in warm, deep tones.", + "seed": 27065, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_13/index_6493_seed_27065_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02493413332849741, + "vlm_logprob": -0.02493413332849741, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.02493413332849741, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a multi-story ancient pagoda visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pagoda constructed of weathered bluish-gray brick with visible cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ancient bronze bells hanging from the pagoda's eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a warm gradient of gold and deep orange due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan around the pagoda?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0033791130408644676 + ], + "opposite_logprob_tokens": [ + -6.7533793449401855 + ], + "logprob_reward": -0.0033791130408644676, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0033791130408644676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10235513001680374, + 0.035496603697538376, + 0.03481833264231682, + 0.1456688791513443, + 0.03960442915558815, + 0.08349357545375824, + 0.092894546687603, + 0.05074552446603775, + 0.2886617183685303, + 0.12626130878925323 + ], + "credit_map": [ + 0.010325245559215546, + 0.013849093578755856, + 0.009625380858778954, + 0.009997794404625893, + 0.014149031601846218, + 0.011119963601231575, + 0.010400054045021534, + 0.022888554260134697, + 0.0051793185994029045, + 0.0048972065560519695, + 0.004211691673845053, + 0.0037804609164595604, + 0.0047929794527590275, + 0.004671666771173477, + 0.004159314092248678, + 0.003803962841629982, + 0.003446295391768217, + 0.00481446972116828, + 0.004486240446567535, + 0.005893412046134472, + 0.003676243359223008, + 0.00441553583368659, + 0.0035843471996486187, + 0.004501787945628166, + 0.008263222873210907, + 0.006676363293081522, + 0.006372099742293358, + 0.006783138029277325, + 0.0065704891458153725, + 0.007017697207629681, + 0.05247990041971207, + 0.05150594189763069, + 0.0033278558403253555, + 0.004750221036374569, + 0.009803793393075466, + 0.006859068758785725, + 0.004427183419466019, + 0.0038105170242488384, + 0.004069420974701643, + 0.002556367078796029, + 0.005786509718745947, + 0.006126128137111664, + 0.00893240887671709, + 0.025771891698241234, + 0.00696529308333993, + 0.004319481085985899, + 0.02020270936191082, + 0.005389146041125059, + 0.008987932465970516, + 0.009652754291892052, + 0.011555952951312065, + 0.019136305898427963, + 0.009427913464605808, + 0.010155984200537205, + 0.016311997547745705, + 0.007665707729756832, + 0.0041352808475494385, + 0.0032760000322014093, + 0.008370527997612953, + 0.009561014361679554, + 0.0030757703352719545, + 0.002882681554183364, + 0.010434040799736977, + 0.009010206907987595, + 0.03421468660235405, + 0.01979944109916687, + 0.019887365400791168, + 0.022976431995630264, + 0.018874641507864, + 0.016987990587949753, + 0.07452508807182312, + 0.08139606565237045, + 0.013984453864395618, + 0.019237179309129715, + 0.011138426139950752, + 0.018696637824177742, + 0.02031688578426838, + 0.008242081850767136, + 0.016848942264914513, + 0.017796695232391357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 9 + ], + "selected_reward": -0.0033791130408644676 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a multi-story ancient pagoda visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pagoda constructed of weathered bluish-gray brick with visible cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ancient bronze bells hanging from the pagoda's eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a warm gradient of gold and deep orange due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan around the pagoda?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029820652678608894 + ], + "opposite_logprob_tokens": [ + -3.529820680618286 + ], + "logprob_reward": -0.029820652678608894, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029820652678608894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07304904609918594, + 0.033464837819337845, + 0.0138050252571702, + 0.09565381705760956, + 0.043863169848918915, + 0.03178863227367401, + 0.0786464661359787, + 0.08960910141468048, + 0.44849395751953125, + 0.09162601828575134 + ], + "credit_map": [ + 0.006786245387047529, + 0.008800224401056767, + 0.00761971902102232, + 0.00752477440983057, + 0.009181304834783077, + 0.006543373689055443, + 0.010285644792020321, + 0.016307760030031204, + 0.0036384230479598045, + 0.006568077020347118, + 0.004052847623825073, + 0.00583166629076004, + 0.0029616367537528276, + 0.003232806921005249, + 0.003909860737621784, + 0.0032695166300982237, + 0.0011843787506222725, + 0.0013012500712648034, + 0.0014908972661942244, + 0.003049806458875537, + 0.0009360768599435687, + 0.00171493215020746, + 0.0020653263200074434, + 0.002062358194962144, + 0.005774514749646187, + 0.003622216871008277, + 0.003751203650608659, + 0.003958198707550764, + 0.0036260129418224096, + 0.004162033554166555, + 0.03604155406355858, + 0.03471808508038521, + 0.0036085296887904406, + 0.005310975015163422, + 0.01059721503406763, + 0.008018097840249538, + 0.0049973102286458015, + 0.003772255266085267, + 0.0046025533229112625, + 0.0029562332201749086, + 0.0019870528485625982, + 0.0025251165498048067, + 0.006611319724470377, + 0.007047382649034262, + 0.0031695931684225798, + 0.0021768042352050543, + 0.005651723127812147, + 0.0026196397375315428, + 0.005078459158539772, + 0.006781621370464563, + 0.00897875428199768, + 0.018763182684779167, + 0.006456951145082712, + 0.00893381703644991, + 0.013664818368852139, + 0.009988862089812756, + 0.006572364829480648, + 0.005715136881917715, + 0.01646646484732628, + 0.01809532381594181, + 0.004186918959021568, + 0.006490716245025396, + 0.015475562773644924, + 0.016606615856289864, + 0.060366589576005936, + 0.02478954754769802, + 0.024406805634498596, + 0.027983782812952995, + 0.0258961021900177, + 0.023568112403154373, + 0.12239383161067963, + 0.1390891969203949, + 0.00879832822829485, + 0.009849537163972855, + 0.010025841183960438, + 0.01175599079579115, + 0.008516359142959118, + 0.006468188017606735, + 0.019751209765672684, + 0.016460562124848366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 9 + ], + "selected_reward": -0.029820652678608894 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a multi-story ancient pagoda visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pagoda constructed of weathered bluish-gray brick with visible cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ancient bronze bells hanging from the pagoda's eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a warm gradient of gold and deep orange due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan around the pagoda?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495474860072136 + ], + "opposite_logprob_tokens": [ + -2.929954767227173 + ], + "logprob_reward": -0.05495474860072136, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495474860072136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1451527178287506, + 0.20712588727474213, + 0.0861562117934227, + 0.051255036145448685, + 0.023635435849428177, + 0.04074079543352127, + 0.060509391129016876, + 0.05277567729353905, + 0.2168537974357605, + 0.11579504609107971 + ], + "credit_map": [ + 0.014429053291678429, + 0.019916757941246033, + 0.01990271359682083, + 0.020730745047330856, + 0.018193360418081284, + 0.012611250393092632, + 0.015712497755885124, + 0.023656338453292847, + 0.02440563216805458, + 0.03907997906208038, + 0.02909683994948864, + 0.0346829816699028, + 0.016159214079380035, + 0.015621112659573555, + 0.026035837829113007, + 0.022044291719794273, + 0.009772584773600101, + 0.007898503914475441, + 0.013185817748308182, + 0.013066977262496948, + 0.008911794051527977, + 0.010088403709232807, + 0.011333754286170006, + 0.011898371390998363, + 0.004200900439172983, + 0.005008245352655649, + 0.004296096507459879, + 0.006269613280892372, + 0.0043562850914895535, + 0.0031493771821260452, + 0.01315862312912941, + 0.010815893299877644, + 0.0018639768240973353, + 0.003876936621963978, + 0.003911105915904045, + 0.004428085871040821, + 0.0021585789509117603, + 0.002969337860122323, + 0.002779645612463355, + 0.0016477665631100535, + 0.002582087880000472, + 0.004151956643909216, + 0.007556581869721413, + 0.009430725127458572, + 0.0033637769520282745, + 0.0026217540726065636, + 0.008003370836377144, + 0.003030540654435754, + 0.004477640613913536, + 0.005691728554666042, + 0.008183990605175495, + 0.016123177483677864, + 0.004834468010812998, + 0.004141842946410179, + 0.010664192959666252, + 0.006392349023371935, + 0.0042816754430532455, + 0.0025731828063726425, + 0.007748469710350037, + 0.009753161109983921, + 0.0023697970900684595, + 0.0035025854595005512, + 0.013667183928191662, + 0.008879625238478184, + 0.02499946393072605, + 0.01483188010752201, + 0.014669578522443771, + 0.01682545803487301, + 0.017738306894898415, + 0.013753531500697136, + 0.0567048005759716, + 0.05733077973127365, + 0.012705320492386818, + 0.016962628811597824, + 0.010220542550086975, + 0.013413705863058567, + 0.013114533387124538, + 0.005007499363273382, + 0.024534130468964577, + 0.019836675375699997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.05495474860072136 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a multi-story ancient pagoda visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pagoda constructed of weathered bluish-gray brick with visible cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ancient bronze bells hanging from the pagoda's eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a warm gradient of gold and deep orange due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan around the pagoda?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006732873618602753 + ], + "opposite_logprob_tokens": [ + -5.006732940673828 + ], + "logprob_reward": -0.006732873618602753, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732873618602753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27729782462120056, + 0.028841493651270866, + 0.026041468605399132, + 0.08684560656547546, + 0.054318513721227646, + 0.04192308709025383, + 0.30280613899230957, + 0.04331408813595772, + 0.10101153701543808, + 0.037600237876176834 + ], + "credit_map": [ + 0.024657782167196274, + 0.0380408801138401, + 0.025869064033031464, + 0.024168269708752632, + 0.041863907128572464, + 0.04588215425610542, + 0.02263445034623146, + 0.05418132618069649, + 0.0029109795577824116, + 0.0047540608793497086, + 0.0037008877843618393, + 0.004300202243030071, + 0.0025717150419950485, + 0.0032222734298557043, + 0.0045569222420454025, + 0.002824450610205531, + 0.0024921607691794634, + 0.0037116389721632004, + 0.0029475160408765078, + 0.005066262558102608, + 0.0024522976018488407, + 0.003085206961259246, + 0.003213929943740368, + 0.0030724548269063234, + 0.005063783377408981, + 0.0033898036926984787, + 0.002695223316550255, + 0.0031149189453572035, + 0.003208947367966175, + 0.0028514224104583263, + 0.034546077251434326, + 0.03197542950510979, + 0.003991082776337862, + 0.006397755816578865, + 0.012722217477858067, + 0.011729434132575989, + 0.004021194763481617, + 0.005483593326061964, + 0.0067392149940133095, + 0.0032340253237634897, + 0.0027050632052123547, + 0.003346542129293084, + 0.005497274920344353, + 0.011333712376654148, + 0.0032825591042637825, + 0.0025831947568804026, + 0.009700482711195946, + 0.003474256256595254, + 0.026296336203813553, + 0.02796458639204502, + 0.04192345216870308, + 0.07591401040554047, + 0.014844841323792934, + 0.026761939749121666, + 0.053516652435064316, + 0.03558432310819626, + 0.00421892199665308, + 0.002330733695998788, + 0.006864692084491253, + 0.006558841560035944, + 0.0016128139104694128, + 0.00224721129052341, + 0.01222923956811428, + 0.007251634728163481, + 0.00859787780791521, + 0.0138957928866148, + 0.007883327081799507, + 0.012501749210059643, + 0.011117684654891491, + 0.008358249440789223, + 0.019903667271137238, + 0.018753191456198692, + 0.003477477002888918, + 0.005126601085066795, + 0.0036369177978485823, + 0.004628356080502272, + 0.005757865961641073, + 0.0029345641378313303, + 0.0069521283730864525, + 0.005086325109004974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.006732873618602753 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a multi-story ancient pagoda visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pagoda constructed of weathered bluish-gray brick with visible cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ancient bronze bells hanging from the pagoda's eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a warm gradient of gold and deep orange due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan around the pagoda?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029783278703689575 + ], + "opposite_logprob_tokens": [ + -3.529783248901367 + ], + "logprob_reward": -0.029783278703689575, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029783278703689575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3109254539012909, + 0.05071193724870682, + 0.06227152422070503, + 0.07275035977363586, + 0.08021339774131775, + 0.09952481091022491, + 0.08312827348709106, + 0.11024578660726547, + 0.08648290485143661, + 0.0437454953789711 + ], + "credit_map": [ + 0.02371268719434738, + 0.049916062504053116, + 0.028473321348428726, + 0.031169263646006584, + 0.04333652928471565, + 0.028387906029820442, + 0.04828391969203949, + 0.057645756751298904, + 0.006020271219313145, + 0.005273027345538139, + 0.008163828402757645, + 0.005539241712540388, + 0.0049457126297056675, + 0.008166655898094177, + 0.005824793595820665, + 0.006778411101549864, + 0.005272412672638893, + 0.005629479419440031, + 0.005857391282916069, + 0.012775002978742123, + 0.005785687360912561, + 0.00871303677558899, + 0.007967285811901093, + 0.010271233506500721, + 0.005598363932222128, + 0.008657519705593586, + 0.008393075317144394, + 0.012765790335834026, + 0.006532652769237757, + 0.006977483164519072, + 0.012071270495653152, + 0.011754211969673634, + 0.007789397146552801, + 0.009635177440941334, + 0.01269800215959549, + 0.011643155477941036, + 0.01054146233946085, + 0.009972705505788326, + 0.01237137895077467, + 0.005562117323279381, + 0.00687581580132246, + 0.011070395819842815, + 0.018059249967336655, + 0.018255211412906647, + 0.009637412615120411, + 0.008428453467786312, + 0.019356796517968178, + 0.007841489277780056, + 0.007245677523314953, + 0.009193693287670612, + 0.011160743422806263, + 0.011759340763092041, + 0.013005781918764114, + 0.010853667743504047, + 0.011042751371860504, + 0.008866614662110806, + 0.007620764896273613, + 0.003266669111326337, + 0.012898188084363937, + 0.013604213483631611, + 0.004082468803972006, + 0.00536632165312767, + 0.03940344974398613, + 0.02400372177362442, + 0.006940495688468218, + 0.008681394159793854, + 0.010307949036359787, + 0.012591429986059666, + 0.01006856095045805, + 0.0059418645687401295, + 0.018036741763353348, + 0.013914473354816437, + 0.0037836607079952955, + 0.004860566463321447, + 0.00392107991501689, + 0.00484161451458931, + 0.003939024172723293, + 0.0029828071128576994, + 0.009225928224623203, + 0.010190821252763271 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.029783278703689575 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_14/index_tensor([3819, 3819])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_14/index_tensor([3819, 3819])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..00f90c11f6da65b8ea1fcb29b7865fd475b2fdb8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_14/index_tensor([3819, 3819])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a stark, minimalist artistic style, a wide shot opens with the camera panning slowly and steadily to the right, tracking the raw, elemental force of a flash flood. A raging, slate-gray torrent of water, thick with mud and debris, surges through the narrow confines of a desolate canyon, whose steep, geometric walls of dark, damp rock are rendered in a desaturated and muted palette. The camera continues in a deliberate pan, tilting slightly upward to reveal a high, flat ledge jutting out from the cliff face. Here, the solitary figure of Mathias stands completely still, his form a stark, black silhouette against the chaotic, bubbling water far below. His posture is stiff and sharp, his shoulders squared, his tall frame set against the devastation. Against the flat, unforgiving light of the overcast sky, his facial features are completely lost in deep shadow, rendering his expression completely unreadable as he silently and motionlessly watches the disaster unfold, creating a powerful and unsettling atmosphere of detached witness.", + "seed": 25739, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_14/index_3819_seed_25739_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07528730109333992, + "vlm_logprob": -0.07528730109333992, + "vlm_yes_no_margin": 3.175000023841858, + "vlm_token_credit": -0.07528730109333992, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate canyon with steep, geometric walls of dark, damp rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a raging, slate-gray torrent of water thick with mud and debris surging through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while tracking the flood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary figure of Mathias visible standing completely still on a high, flat ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is Mathias rendered as a stark, black silhouette with stiff, sharp posture against the chaotic water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019427964463829994 + ], + "opposite_logprob_tokens": [ + -4.26942777633667 + ], + "logprob_reward": -0.019427964463829994, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.019427964463829994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.113498255610466, + 0.13020063936710358, + 0.05560821667313576, + 0.2651267945766449, + 0.06022199988365173, + 0.21706590056419373, + 0.04981912299990654, + 0.029666230082511902, + 0.022837206721305847, + 0.05595563352108002 + ], + "credit_map": [ + 0.012195597402751446, + 0.009386153891682625, + 0.013923637568950653, + 0.018835144117474556, + 0.008715527132153511, + 0.019292429089546204, + 0.017527112737298012, + 0.01362266018986702, + 0.009954777546226978, + 0.009577100165188313, + 0.014422175474464893, + 0.01770922914147377, + 0.007412370759993792, + 0.022441688925027847, + 0.030304258689284325, + 0.018379036337137222, + 0.005848415661603212, + 0.004684161394834518, + 0.0060436963103711605, + 0.00832398608326912, + 0.003951381891965866, + 0.0078545231372118, + 0.00919442716985941, + 0.009707624092698097, + 0.009533531963825226, + 0.012211923487484455, + 0.03943725302815437, + 0.017532622441649437, + 0.00960551667958498, + 0.019925681874155998, + 0.05882547050714493, + 0.09805481135845184, + 0.004035085439682007, + 0.005172232631593943, + 0.009571239352226257, + 0.011308898217976093, + 0.005390497390180826, + 0.009372779168188572, + 0.00803301203995943, + 0.0073382556438446045, + 0.018331022933125496, + 0.022070087492465973, + 0.024929523468017578, + 0.032112665474414825, + 0.022040579468011856, + 0.026604413986206055, + 0.028925320133566856, + 0.04205229878425598, + 0.0040510897524654865, + 0.004669755697250366, + 0.00467979209497571, + 0.007699931971728802, + 0.004956225864589214, + 0.010432491078972816, + 0.006738669238984585, + 0.006591164972633123, + 0.0023165703751146793, + 0.0023212884552776814, + 0.003671785583719611, + 0.004747187718749046, + 0.003114659572020173, + 0.0047119539231061935, + 0.003837335156276822, + 0.004945449996739626, + 0.0017385249957442284, + 0.0020382713992148638, + 0.0031610450241714716, + 0.004020561929792166, + 0.0027215529698878527, + 0.001968326047062874, + 0.004506286699324846, + 0.0026826378889381886, + 0.0039217215962708, + 0.0047018141485750675, + 0.009644757024943829, + 0.009240265004336834, + 0.006361345760524273, + 0.006542479619383812, + 0.008456719107925892, + 0.007086534518748522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 1 + ], + "selected_reward": -0.019427964463829994 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate canyon with steep, geometric walls of dark, damp rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a raging, slate-gray torrent of water thick with mud and debris surging through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while tracking the flood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary figure of Mathias visible standing completely still on a high, flat ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is Mathias rendered as a stark, black silhouette with stiff, sharp posture against the chaotic water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608162373304367 + ], + "opposite_logprob_tokens": [ + -4.141081809997559 + ], + "logprob_reward": -0.01608162373304367, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608162373304367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.324904203414917, + 0.1769411414861679, + 0.03239356353878975, + 0.09696251153945923, + 0.028142008930444717, + 0.13499169051647186, + 0.03973396122455597, + 0.06614045053720474, + 0.03025134652853012, + 0.06953906267881393 + ], + "credit_map": [ + 0.023007094860076904, + 0.022452227771282196, + 0.0328187569975853, + 0.04478031396865845, + 0.025167366489768028, + 0.05726992338895798, + 0.07317288964986801, + 0.04623565822839737, + 0.00910051167011261, + 0.014359467662870884, + 0.017505310475826263, + 0.025164639577269554, + 0.011046678759157658, + 0.03109714761376381, + 0.04325813427567482, + 0.025409270077943802, + 0.0025731786154210567, + 0.002845857758074999, + 0.0030695288442075253, + 0.0059905522502958775, + 0.0013389396481215954, + 0.0033888318575918674, + 0.007187497336417437, + 0.00599918095394969, + 0.0034709277097135782, + 0.004265143536031246, + 0.014006108045578003, + 0.006831204053014517, + 0.0030867939349263906, + 0.006977348122745752, + 0.023144137114286423, + 0.03518085554242134, + 0.0016689246986061335, + 0.002711785724386573, + 0.004467548802495003, + 0.003537094220519066, + 0.0022301210556179285, + 0.004765473771840334, + 0.003935875371098518, + 0.004825189244002104, + 0.011493845842778683, + 0.01519702933728695, + 0.01549441460520029, + 0.021605629473924637, + 0.013905929401516914, + 0.017084799706935883, + 0.016839874908328056, + 0.023370182141661644, + 0.0036502412986010313, + 0.0034458914306014776, + 0.003978801425546408, + 0.00624857796356082, + 0.0038507573772221804, + 0.007165754679590464, + 0.0059308274649083614, + 0.005463117733597755, + 0.006599053740501404, + 0.006022329907864332, + 0.01076577976346016, + 0.011623905040323734, + 0.005589545238763094, + 0.007492327596992254, + 0.00968523696064949, + 0.008362284861505032, + 0.0021727399434894323, + 0.002720109187066555, + 0.005600230302661657, + 0.004848531447350979, + 0.003212997689843178, + 0.00239713117480278, + 0.0056919013150036335, + 0.0036077096592634916, + 0.0047382391057908535, + 0.004690240137279034, + 0.011976820416748524, + 0.007271149195730686, + 0.009482856839895248, + 0.010022344067692757, + 0.010265481658279896, + 0.011091938242316246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.01608162373304367 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate canyon with steep, geometric walls of dark, damp rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a raging, slate-gray torrent of water thick with mud and debris surging through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while tracking the flood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary figure of Mathias visible standing completely still on a high, flat ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is Mathias rendered as a stark, black silhouette with stiff, sharp posture against the chaotic water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14279404282569885 + ], + "opposite_logprob_tokens": [ + -2.017794132232666 + ], + "logprob_reward": -0.14279404282569885, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14279404282569885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08046844601631165, + 0.043945204466581345, + 0.08323223888874054, + 0.08599520474672318, + 0.11980371177196503, + 0.1519915908575058, + 0.09443335980176926, + 0.0785621851682663, + 0.130657359957695, + 0.13091064989566803 + ], + "credit_map": [ + 0.01027660258114338, + 0.005978923756629229, + 0.00853472575545311, + 0.013373542577028275, + 0.00432168785482645, + 0.009931064210832119, + 0.016074862331151962, + 0.011977043934166431, + 0.005900656338781118, + 0.0038069873116910458, + 0.00621045334264636, + 0.007338028401136398, + 0.00250829104334116, + 0.007037673611193895, + 0.005546138156205416, + 0.005596979055553675, + 0.007271762937307358, + 0.006109834183007479, + 0.010654815472662449, + 0.00976879708468914, + 0.004880901891738176, + 0.011848054826259613, + 0.01460624672472477, + 0.018091823905706406, + 0.004835082218050957, + 0.003947610035538673, + 0.012540441006422043, + 0.007210525218397379, + 0.0029834015294909477, + 0.007203985005617142, + 0.018350565806031227, + 0.02892359532415867, + 0.007756042294204235, + 0.005901907570660114, + 0.010915340855717659, + 0.01279453095048666, + 0.005486658774316311, + 0.00905669666826725, + 0.041669148951768875, + 0.026223385706543922, + 0.010346275754272938, + 0.014475093223154545, + 0.013814426958560944, + 0.01817169226706028, + 0.013990649953484535, + 0.022600634023547173, + 0.025083579123020172, + 0.033509232103824615, + 0.00643038097769022, + 0.006661705207079649, + 0.014690831303596497, + 0.009415263310074806, + 0.009187151677906513, + 0.016676126047968864, + 0.012251567095518112, + 0.019120333716273308, + 0.006627323105931282, + 0.006947201676666737, + 0.00858332309871912, + 0.011817819438874722, + 0.007591406814754009, + 0.011085343547165394, + 0.011360116302967072, + 0.014549649320542812, + 0.009582223370671272, + 0.011809473857283592, + 0.01929844170808792, + 0.02469978854060173, + 0.017351698130369186, + 0.012526952661573887, + 0.02409878559410572, + 0.011289993301033974, + 0.008844990283250809, + 0.00962680485099554, + 0.02493583783507347, + 0.016161499544978142, + 0.017106756567955017, + 0.01852620579302311, + 0.018431782722473145, + 0.01727677322924137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.14279404282569885 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate canyon with steep, geometric walls of dark, damp rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a raging, slate-gray torrent of water thick with mud and debris surging through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while tracking the flood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary figure of Mathias visible standing completely still on a high, flat ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is Mathias rendered as a stark, black silhouette with stiff, sharp posture against the chaotic water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17992918193340302 + ], + "opposite_logprob_tokens": [ + -1.8049291372299194 + ], + "logprob_reward": -0.17992918193340302, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17992918193340302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06832198053598404, + 0.03725028410553932, + 0.0731964036822319, + 0.08358912169933319, + 0.052364781498909, + 0.5073308944702148, + 0.06030722334980965, + 0.027389883995056152, + 0.04956277087330818, + 0.040686678141355515 + ], + "credit_map": [ + 0.00465529877692461, + 0.005464472807943821, + 0.006775751709938049, + 0.010992779396474361, + 0.0040338775143027306, + 0.007218849379569292, + 0.018651548773050308, + 0.010529396124184132, + 0.004606260452419519, + 0.002998732030391693, + 0.00507922749966383, + 0.005076676607131958, + 0.0017429455183446407, + 0.004828466568142176, + 0.007553724106401205, + 0.005364245735108852, + 0.010398123413324356, + 0.006036088801920414, + 0.011761152185499668, + 0.006962894462049007, + 0.003790046786889434, + 0.009832331910729408, + 0.009941545315086842, + 0.014474212191998959, + 0.005531650967895985, + 0.007011910900473595, + 0.011893331073224545, + 0.009089663624763489, + 0.0051446580328047276, + 0.008272971957921982, + 0.014734030701220036, + 0.021910900250077248, + 0.0042711584828794, + 0.007406000513583422, + 0.006910033989697695, + 0.007585474755614996, + 0.00569764943793416, + 0.006624347530305386, + 0.007747525814920664, + 0.006122583989053965, + 0.0431327223777771, + 0.05575910583138466, + 0.059746742248535156, + 0.0835840106010437, + 0.05037834495306015, + 0.061057403683662415, + 0.07773476094007492, + 0.07593776285648346, + 0.00471611600369215, + 0.005830678623169661, + 0.007270591799169779, + 0.0076572466641664505, + 0.007656211499124765, + 0.012905297800898552, + 0.006622797809541225, + 0.007648277562111616, + 0.0015983260236680508, + 0.0029627897311002016, + 0.004678851459175348, + 0.003482103580608964, + 0.0034887755755335093, + 0.004646098241209984, + 0.0033660391345620155, + 0.0031668981537222862, + 0.0035502328537404537, + 0.0030827661976218224, + 0.008312489837408066, + 0.006519155111163855, + 0.004010115750133991, + 0.0038248449563980103, + 0.01172430720180273, + 0.00853885617107153, + 0.003440458793193102, + 0.0031736521050333977, + 0.0041977595537900925, + 0.005414695478975773, + 0.005215069279074669, + 0.006094887852668762, + 0.00569344125688076, + 0.007456709165126085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 2 + ], + "selected_reward": -0.17992918193340302 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate canyon with steep, geometric walls of dark, damp rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a raging, slate-gray torrent of water thick with mud and debris surging through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while tracking the flood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary figure of Mathias visible standing completely still on a high, flat ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is Mathias rendered as a stark, black silhouette with stiff, sharp posture against the chaotic water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018203692510724068 + ], + "opposite_logprob_tokens": [ + -4.0182037353515625 + ], + "logprob_reward": -0.018203692510724068, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018203692510724068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07460520416498184, + 0.12022612243890762, + 0.05567622929811478, + 0.10449517518281937, + 0.11605257540941238, + 0.19765351712703705, + 0.16993224620819092, + 0.044729478657245636, + 0.04158743470907211, + 0.07504201680421829 + ], + "credit_map": [ + 0.0049844831228256226, + 0.006307777017354965, + 0.007898673415184021, + 0.011636510491371155, + 0.005619675852358341, + 0.009442625567317009, + 0.017999572679400444, + 0.010715882293879986, + 0.008775224909186363, + 0.00857984833419323, + 0.013556273654103279, + 0.014932391233742237, + 0.006709329318255186, + 0.020593812689185143, + 0.029523340985178947, + 0.017555907368659973, + 0.006945521105080843, + 0.004282901529222727, + 0.007916435599327087, + 0.004552684258669615, + 0.0021489067003130913, + 0.005272504407912493, + 0.008933426812291145, + 0.015623847022652626, + 0.004283299669623375, + 0.004975411109626293, + 0.015117472037672997, + 0.00621913094073534, + 0.004095674958080053, + 0.010071812197566032, + 0.022892776876688004, + 0.036839596927165985, + 0.007695223204791546, + 0.01065797545015812, + 0.01305127702653408, + 0.016054093837738037, + 0.009040823206305504, + 0.011001299135386944, + 0.029724929481744766, + 0.018826959654688835, + 0.016299884766340256, + 0.021319186314940453, + 0.023161521181464195, + 0.03200308606028557, + 0.018253348767757416, + 0.02133522927761078, + 0.03281010314822197, + 0.03247115761041641, + 0.013264269568026066, + 0.013635429553687572, + 0.02080453187227249, + 0.024119971320033073, + 0.016299109905958176, + 0.035417214035987854, + 0.02129862830042839, + 0.025093095377087593, + 0.0025770054198801517, + 0.005595372058451176, + 0.004373412113636732, + 0.0065066078677773476, + 0.005977262742817402, + 0.006678120233118534, + 0.005270383320748806, + 0.007751314900815487, + 0.0030669188126921654, + 0.003469503717496991, + 0.0046629104763269424, + 0.006142128724604845, + 0.006638609804213047, + 0.004693111404776573, + 0.008124813437461853, + 0.004789434373378754, + 0.005459195002913475, + 0.0059738121926784515, + 0.011720262467861176, + 0.008730526082217693, + 0.009677750989794731, + 0.012222135439515114, + 0.010416428558528423, + 0.01084190420806408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 1 + ], + "selected_reward": -0.018203692510724068 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stark, minimalist artistic style, a wide shot opens with the camera panning slowly and steadily to the right, tracking the raw, elemental force of a flash flood. A raging, slate-gray torrent of water, thick with mud and debris, surges through the narrow confines of a desolate canyon, whose steep, geometric walls of dark, damp rock are rendered in a desaturated and muted palette. The camera continues in a deliberate pan, tilting slightly upward to reveal a high, flat ledge jutting out from the cliff face. Here, the solitary figure of Mathias stands completely still, his form a stark, black silhouette against the chaotic, bubbling water far below. His posture is stiff and sharp, his shoulders squared, his tall frame set against the devastation. Against the flat, unforgiving light of the overcast sky, his facial features are completely lost in deep shadow, rendering his expression completely unreadable as he silently and motionlessly watches the disaster unfold, creating a powerful and unsettling atmosphere of detached witness.", + "seed": 25740, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_14/index_3819_seed_25740_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.048168869223445654, + "vlm_logprob": -0.048168869223445654, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.048168869223445654, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate canyon with steep, geometric walls of dark, damp rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a raging, slate-gray torrent of water thick with mud and debris surging through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while tracking the flood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary figure of Mathias visible standing completely still on a high, flat ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is Mathias rendered as a stark, black silhouette with stiff, sharp posture against the chaotic water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01740175299346447 + ], + "opposite_logprob_tokens": [ + -4.892401695251465 + ], + "logprob_reward": -0.01740175299346447, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.01740175299346447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19402547180652618, + 0.06385286152362823, + 0.07637326419353485, + 0.03752580285072327, + 0.0786764919757843, + 0.1180526465177536, + 0.10007236152887344, + 0.047434158623218536, + 0.1523754596710205, + 0.1316113919019699 + ], + "credit_map": [ + 0.02573934942483902, + 0.04328004643321037, + 0.028313156217336655, + 0.02376650460064411, + 0.01825655624270439, + 0.015345006249845028, + 0.018520694226026535, + 0.020804183557629585, + 0.010044929571449757, + 0.01644792966544628, + 0.007217315025627613, + 0.007806877605617046, + 0.006256224121898413, + 0.005495843011885881, + 0.004918189253658056, + 0.005665563978254795, + 0.009015893563628197, + 0.01344439759850502, + 0.010135848075151443, + 0.015276786871254444, + 0.005436367355287075, + 0.00583588145673275, + 0.005812210030853748, + 0.011415885761380196, + 0.004923035856336355, + 0.004879219923168421, + 0.007322798483073711, + 0.005745507311075926, + 0.003219744423404336, + 0.003241581842303276, + 0.003721708431839943, + 0.004472214728593826, + 0.013421173207461834, + 0.012894602492451668, + 0.016989625990390778, + 0.006348771508783102, + 0.0041472953744232655, + 0.0067666261456906796, + 0.006322829052805901, + 0.011785581707954407, + 0.02167215384542942, + 0.026030180975794792, + 0.01671413518488407, + 0.006389155983924866, + 0.006822001188993454, + 0.011187231168150902, + 0.010558674111962318, + 0.01867913268506527, + 0.024042967706918716, + 0.01468654815107584, + 0.021351158618927002, + 0.007290200795978308, + 0.005814412608742714, + 0.009043777361512184, + 0.010199116542935371, + 0.007644196972250938, + 0.00716371601447463, + 0.009118396788835526, + 0.008047133684158325, + 0.00431618420407176, + 0.003965382929891348, + 0.0053207105956971645, + 0.005135844461619854, + 0.0043668015860021114, + 0.025813668966293335, + 0.022502319887280464, + 0.01845691353082657, + 0.012419847771525383, + 0.012415705248713493, + 0.01815301738679409, + 0.02045939676463604, + 0.022154629230499268, + 0.015324730426073074, + 0.02460949495434761, + 0.013603588566184044, + 0.01304494496434927, + 0.01165375579148531, + 0.014524232596158981, + 0.01955150067806244, + 0.01929916813969612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.01740175299346447 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate canyon with steep, geometric walls of dark, damp rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a raging, slate-gray torrent of water thick with mud and debris surging through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while tracking the flood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary figure of Mathias visible standing completely still on a high, flat ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is Mathias rendered as a stark, black silhouette with stiff, sharp posture against the chaotic water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820954494178295 + ], + "opposite_logprob_tokens": [ + -4.018209457397461 + ], + "logprob_reward": -0.01820954494178295, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820954494178295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1481585055589676, + 0.061285506933927536, + 0.0743226632475853, + 0.07655633240938187, + 0.034436680376529694, + 0.23987755179405212, + 0.0514555349946022, + 0.09176510572433472, + 0.08744192868471146, + 0.1347002238035202 + ], + "credit_map": [ + 0.014326577074825764, + 0.025145534425973892, + 0.02768140472471714, + 0.023615851998329163, + 0.012724310159683228, + 0.012713916599750519, + 0.016415439546108246, + 0.015535480342805386, + 0.011289362795650959, + 0.014852450229227543, + 0.011765743605792522, + 0.0063785663805902, + 0.003049561521038413, + 0.0047770473174750805, + 0.003944531548768282, + 0.005228241439908743, + 0.009186199866235256, + 0.010571250692009926, + 0.016174273565411568, + 0.011534666642546654, + 0.005409280303865671, + 0.005793521646410227, + 0.005969202611595392, + 0.00968427024781704, + 0.011025392450392246, + 0.01825823076069355, + 0.012066482566297054, + 0.00853685848414898, + 0.0044806585647165775, + 0.005541910417377949, + 0.007537192665040493, + 0.009109612554311752, + 0.005784592591226101, + 0.007314967457205057, + 0.0041388822719454765, + 0.003698629792779684, + 0.0025270809419453144, + 0.003043191274628043, + 0.004860308952629566, + 0.0030690261628478765, + 0.03480669856071472, + 0.02339644730091095, + 0.05981917679309845, + 0.017110155895352364, + 0.011361366137862206, + 0.02270679362118244, + 0.02977849543094635, + 0.0408984050154686, + 0.00981819536536932, + 0.006812830921262503, + 0.014923411421477795, + 0.003108850447461009, + 0.0027973235119134188, + 0.004582906141877174, + 0.005297441501170397, + 0.004114576615393162, + 0.01390029862523079, + 0.017228465527296066, + 0.014762409031391144, + 0.006972513161599636, + 0.006967406254261732, + 0.0105798514559865, + 0.01022312417626381, + 0.01113103237003088, + 0.01637995056807995, + 0.00808239821344614, + 0.012410133145749569, + 0.009201029315590858, + 0.006720927078276873, + 0.010301478207111359, + 0.013205913826823235, + 0.011140101589262486, + 0.015018513426184654, + 0.015853151679039, + 0.013523107394576073, + 0.012367142364382744, + 0.011089028790593147, + 0.017883431166410446, + 0.021633107215166092, + 0.027332738041877747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.01820954494178295 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate canyon with steep, geometric walls of dark, damp rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a raging, slate-gray torrent of water thick with mud and debris surging through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while tracking the flood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary figure of Mathias visible standing completely still on a high, flat ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is Mathias rendered as a stark, black silhouette with stiff, sharp posture against the chaotic water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16036272048950195 + ], + "opposite_logprob_tokens": [ + -1.910362720489502 + ], + "logprob_reward": -0.16036272048950195, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16036272048950195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05468121916055679, + 0.050826411694288254, + 0.025431759655475616, + 0.03418487310409546, + 0.022642968222498894, + 0.04617612436413765, + 0.4426496922969818, + 0.045145463198423386, + 0.1862565279006958, + 0.09200496971607208 + ], + "credit_map": [ + 0.005924168974161148, + 0.009669791907072067, + 0.007973531261086464, + 0.007060251198709011, + 0.004884309135377407, + 0.004532791208475828, + 0.008763065561652184, + 0.0058733089827001095, + 0.007493179757148027, + 0.01593415066599846, + 0.007016107439994812, + 0.006191919557750225, + 0.0029736030846834183, + 0.0033770662266761065, + 0.0039532058872282505, + 0.0038871821016073227, + 0.0034898302983492613, + 0.00408035796135664, + 0.00362192839384079, + 0.004814275540411472, + 0.0017364013474434614, + 0.002179664559662342, + 0.002378612058237195, + 0.0031306918244808912, + 0.004258126486092806, + 0.005084378179162741, + 0.005871790461242199, + 0.005049514584243298, + 0.0035632478538900614, + 0.002876042854040861, + 0.0030707595869898796, + 0.004411014262586832, + 0.003352200146764517, + 0.005015477538108826, + 0.0032438558991998434, + 0.002258351305499673, + 0.002188529819250107, + 0.0018104403279721737, + 0.002403845777735114, + 0.002370270434767008, + 0.010615052655339241, + 0.012240163050591946, + 0.007359880488365889, + 0.0024748786818236113, + 0.0028907214291393757, + 0.0037673080805689096, + 0.0034583136439323425, + 0.003369809128344059, + 0.08347637206315994, + 0.06163657456636429, + 0.13277414441108704, + 0.029829666018486023, + 0.01719100959599018, + 0.036456577479839325, + 0.04692741483449936, + 0.034357950091362, + 0.006936497520655394, + 0.008545069955289364, + 0.007926883175969124, + 0.004027828574180603, + 0.002857215702533722, + 0.004794608801603317, + 0.004726377781480551, + 0.005330982152372599, + 0.03936506807804108, + 0.02570134773850441, + 0.023442907258868217, + 0.016340915113687515, + 0.013561882078647614, + 0.021703841164708138, + 0.023585688322782516, + 0.022554898634552956, + 0.012752555310726166, + 0.020132189616560936, + 0.011444308795034885, + 0.009354451671242714, + 0.008672403171658516, + 0.009375021792948246, + 0.009475894272327423, + 0.010798156261444092 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 9 + ], + "selected_reward": -0.16036272048950195 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate canyon with steep, geometric walls of dark, damp rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a raging, slate-gray torrent of water thick with mud and debris surging through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while tracking the flood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary figure of Mathias visible standing completely still on a high, flat ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is Mathias rendered as a stark, black silhouette with stiff, sharp posture against the chaotic water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811921924352646 + ], + "opposite_logprob_tokens": [ + -3.288119316101074 + ], + "logprob_reward": -0.03811921924352646, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811921924352646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11274577677249908, + 0.06257887929677963, + 0.039116937667131424, + 0.036007389426231384, + 0.08999784290790558, + 0.035209350287914276, + 0.09916557371616364, + 0.08975779265165329, + 0.1437651515007019, + 0.2916553318500519 + ], + "credit_map": [ + 0.011103555560112, + 0.01522538997232914, + 0.01672220602631569, + 0.013465690426528454, + 0.009488984942436218, + 0.01160585880279541, + 0.024756573140621185, + 0.010377510450780392, + 0.008544891141355038, + 0.01333172433078289, + 0.009401245974004269, + 0.007105131167918444, + 0.004538324661552906, + 0.005333449691534042, + 0.005739911459386349, + 0.008584202267229557, + 0.005424566101282835, + 0.009348282590508461, + 0.004440575372427702, + 0.00532189104706049, + 0.004800142254680395, + 0.0034192369785159826, + 0.0028980872593820095, + 0.0034641565289348364, + 0.004318981431424618, + 0.005959602072834969, + 0.005008546635508537, + 0.0038042832165956497, + 0.002987734042108059, + 0.0033015538938343525, + 0.004107146989554167, + 0.006519538350403309, + 0.015569804236292839, + 0.02196314185857773, + 0.013500799424946308, + 0.006547723896801472, + 0.008224938064813614, + 0.008078153245151043, + 0.008176546543836594, + 0.007936732843518257, + 0.006386301945894957, + 0.005040984135121107, + 0.007109322585165501, + 0.0023908826988190413, + 0.0018477911362424493, + 0.0034849499352276325, + 0.0038244673050940037, + 0.005124649032950401, + 0.018370840698480606, + 0.015320908278226852, + 0.02619374729692936, + 0.0067701092921197414, + 0.005351383239030838, + 0.008857757784426212, + 0.010199494659900665, + 0.008101326413452625, + 0.011012019589543343, + 0.011439717374742031, + 0.010138429701328278, + 0.008200221695005894, + 0.008476560935378075, + 0.010630602948367596, + 0.015740828588604927, + 0.014119412750005722, + 0.026550553739070892, + 0.02328387089073658, + 0.021800609305500984, + 0.010654646903276443, + 0.009673780761659145, + 0.015633679926395416, + 0.01580735668540001, + 0.020360643044114113, + 0.04097781702876091, + 0.06613292545080185, + 0.03270907700061798, + 0.028002293780446053, + 0.027090270072221756, + 0.030384548008441925, + 0.03226117417216301, + 0.034097205847501755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.03811921924352646 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate canyon with steep, geometric walls of dark, damp rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a raging, slate-gray torrent of water thick with mud and debris surging through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while tracking the flood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary figure of Mathias visible standing completely still on a high, flat ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is Mathias rendered as a stark, black silhouette with stiff, sharp posture against the chaotic water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067511084489524364 + ], + "opposite_logprob_tokens": [ + -5.00675106048584 + ], + "logprob_reward": -0.0067511084489524364, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067511084489524364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10123882442712784, + 0.11500483751296997, + 0.03966512903571129, + 0.029969051480293274, + 0.03660834580659866, + 0.040897734463214874, + 0.11626516282558441, + 0.05740043520927429, + 0.05970141291618347, + 0.4032490849494934 + ], + "credit_map": [ + 0.011688696220517159, + 0.015145604498684406, + 0.018437428399920464, + 0.01499304547905922, + 0.008322020061314106, + 0.008609010837972164, + 0.016338234767317772, + 0.007704783696681261, + 0.02082156576216221, + 0.03769112750887871, + 0.009971704334020615, + 0.012937670573592186, + 0.007894102483987808, + 0.008885898627340794, + 0.007180210202932358, + 0.009622561745345592, + 0.0058509353548288345, + 0.008271552622318268, + 0.0042719063349068165, + 0.0074656642973423, + 0.0030201380141079426, + 0.0030620319303125143, + 0.0033817302901297808, + 0.0043411701917648315, + 0.0031222146935760975, + 0.005167058669030666, + 0.004026442766189575, + 0.004812243394553661, + 0.0043016052804887295, + 0.002619736595079303, + 0.0025722484569996595, + 0.003347501391544938, + 0.00653618760406971, + 0.0068040816113352776, + 0.006461886689066887, + 0.0038816763553768396, + 0.0033142848405987024, + 0.0029217551928013563, + 0.0023960487451404333, + 0.004292422439903021, + 0.006340108811855316, + 0.005582718178629875, + 0.007266933098435402, + 0.0035570834297686815, + 0.0031360359862446785, + 0.004062355495989323, + 0.0049126301892101765, + 0.006039872765541077, + 0.020788775756955147, + 0.01582660898566246, + 0.034598458558321, + 0.008029591292142868, + 0.00568805355578661, + 0.009794051758944988, + 0.0118069127202034, + 0.009732725098729134, + 0.00730075454339385, + 0.012157787568867207, + 0.00745522603392601, + 0.003807943081483245, + 0.00414273701608181, + 0.006073717027902603, + 0.008720350451767445, + 0.007741920650005341, + 0.008478065021336079, + 0.005936331581324339, + 0.009424213320016861, + 0.0069100987166166306, + 0.004227923229336739, + 0.007866036146879196, + 0.008320983499288559, + 0.008537761867046356, + 0.05554914101958275, + 0.09137310832738876, + 0.04599712789058685, + 0.04019888862967491, + 0.03709159418940544, + 0.04170757159590721, + 0.045513056218624115, + 0.04581863060593605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 1 + ], + "selected_reward": -0.0067511084489524364 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stark, minimalist artistic style, a wide shot opens with the camera panning slowly and steadily to the right, tracking the raw, elemental force of a flash flood. A raging, slate-gray torrent of water, thick with mud and debris, surges through the narrow confines of a desolate canyon, whose steep, geometric walls of dark, damp rock are rendered in a desaturated and muted palette. The camera continues in a deliberate pan, tilting slightly upward to reveal a high, flat ledge jutting out from the cliff face. Here, the solitary figure of Mathias stands completely still, his form a stark, black silhouette against the chaotic, bubbling water far below. His posture is stiff and sharp, his shoulders squared, his tall frame set against the devastation. Against the flat, unforgiving light of the overcast sky, his facial features are completely lost in deep shadow, rendering his expression completely unreadable as he silently and motionlessly watches the disaster unfold, creating a powerful and unsettling atmosphere of detached witness.", + "seed": 25741, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_14/index_3819_seed_25741_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03210463160648942, + "vlm_logprob": -0.03210463160648942, + "vlm_yes_no_margin": 4.024999952316284, + "vlm_token_credit": -0.03210463160648942, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate canyon with steep, geometric walls of dark, damp rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a raging, slate-gray torrent of water thick with mud and debris surging through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while tracking the flood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary figure of Mathias visible standing completely still on a high, flat ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is Mathias rendered as a stark, black silhouette with stiff, sharp posture against the chaotic water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023697830736637115 + ], + "opposite_logprob_tokens": [ + -4.148697853088379 + ], + "logprob_reward": -0.023697830736637115, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.023697830736637115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16709057986736298, + 0.048651304095983505, + 0.04991213232278824, + 0.32224369049072266, + 0.07163640111684799, + 0.06592784076929092, + 0.05361698940396309, + 0.03791852667927742, + 0.1501273214817047, + 0.03287530690431595 + ], + "credit_map": [ + 0.013457223773002625, + 0.024114258587360382, + 0.01960337534546852, + 0.01975744590163231, + 0.018660077825188637, + 0.02932782843708992, + 0.022107521072030067, + 0.020062847062945366, + 0.00353009020909667, + 0.005710405297577381, + 0.006172311492264271, + 0.0074768816120922565, + 0.004318177234381437, + 0.0085221529006958, + 0.008243920281529427, + 0.004677360877394676, + 0.004507347010076046, + 0.004797825589776039, + 0.0035019144415855408, + 0.012610439211130142, + 0.0046799061819911, + 0.004352432209998369, + 0.009432106278836727, + 0.00603015860542655, + 0.021834081038832664, + 0.020724331960082054, + 0.028239406645298004, + 0.04868606477975845, + 0.0173602607101202, + 0.02314591594040394, + 0.06267274916172028, + 0.09958083182573318, + 0.0055445111356675625, + 0.008493265137076378, + 0.011023202911019325, + 0.013805478811264038, + 0.007922982797026634, + 0.006075137294828892, + 0.010457259602844715, + 0.00831454899162054, + 0.005560874938964844, + 0.004796623717993498, + 0.009168763644993305, + 0.01655922643840313, + 0.0045099761337041855, + 0.0047110095620155334, + 0.012359139509499073, + 0.00826222077012062, + 0.005989596713334322, + 0.004375245422124863, + 0.00889704842120409, + 0.009292599745094776, + 0.004288502968847752, + 0.004974163603037596, + 0.008043215610086918, + 0.0077566104009747505, + 0.0033040232956409454, + 0.0034747023601084948, + 0.008943290449678898, + 0.00756548997014761, + 0.002379268640652299, + 0.0022742003202438354, + 0.005268678069114685, + 0.004708869848400354, + 0.023952551186084747, + 0.013643496669828892, + 0.027820603922009468, + 0.017593035474419594, + 0.01250710990279913, + 0.012885991483926773, + 0.019998928532004356, + 0.021725580096244812, + 0.003692022990435362, + 0.003911146894097328, + 0.006083473097532988, + 0.004162934608757496, + 0.004422206897288561, + 0.002709335647523403, + 0.0034382862504571676, + 0.004455895163118839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.023697830736637115 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate canyon with steep, geometric walls of dark, damp rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a raging, slate-gray torrent of water thick with mud and debris surging through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while tracking the flood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary figure of Mathias visible standing completely still on a high, flat ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is Mathias rendered as a stark, black silhouette with stiff, sharp posture against the chaotic water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004110697191208601 + ], + "opposite_logprob_tokens": [ + -5.504110813140869 + ], + "logprob_reward": -0.004110697191208601, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004110697191208601, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1789477914571762, + 0.07697705924510956, + 0.045870136469602585, + 0.2770419418811798, + 0.0381203331053257, + 0.09398584812879562, + 0.07898081094026566, + 0.10684280097484589, + 0.0682302787899971, + 0.035002924501895905 + ], + "credit_map": [ + 0.012797879986464977, + 0.023710718378424644, + 0.02167583629488945, + 0.028162432834506035, + 0.021836835891008377, + 0.022676631808280945, + 0.023336688056588173, + 0.024750784039497375, + 0.007352248765528202, + 0.009780974127352238, + 0.009478761814534664, + 0.012168895453214645, + 0.007955916225910187, + 0.01112403254956007, + 0.011085821315646172, + 0.008030416443943977, + 0.0045464253053069115, + 0.00422703567892313, + 0.003375000786036253, + 0.011209606193006039, + 0.004221174865961075, + 0.003135230392217636, + 0.008932897821068764, + 0.006222771946340799, + 0.019320352002978325, + 0.017065290361642838, + 0.02403850667178631, + 0.03776535019278526, + 0.015385412611067295, + 0.01972079463303089, + 0.06007501855492592, + 0.08367124944925308, + 0.003798753721639514, + 0.0049101016484200954, + 0.005588346626609564, + 0.005365094635635614, + 0.004555420484393835, + 0.0029154198709875345, + 0.005840920843183994, + 0.005146277602761984, + 0.005932758096605539, + 0.006492539308965206, + 0.011614955961704254, + 0.02765825390815735, + 0.005367884878069162, + 0.006201508454978466, + 0.013869582675397396, + 0.016848376020789146, + 0.008623051457107067, + 0.007599939592182636, + 0.010021199472248554, + 0.009276148863136768, + 0.014787770807743073, + 0.010031107813119888, + 0.008540903218090534, + 0.010100699961185455, + 0.01733967661857605, + 0.012502008117735386, + 0.015222681686282158, + 0.012500859797000885, + 0.018573690205812454, + 0.008669329807162285, + 0.01059985626488924, + 0.011434705927968025, + 0.005428295582532883, + 0.008330480195581913, + 0.01146349310874939, + 0.009705854579806328, + 0.011281633749604225, + 0.004956068005412817, + 0.009028551168739796, + 0.00803590752184391, + 0.0035044648684561253, + 0.004621138796210289, + 0.004685517400503159, + 0.0044012186117470264, + 0.00621157418936491, + 0.0030379488598555326, + 0.004391724243760109, + 0.004149339627474546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.004110697191208601 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate canyon with steep, geometric walls of dark, damp rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a raging, slate-gray torrent of water thick with mud and debris surging through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while tracking the flood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary figure of Mathias visible standing completely still on a high, flat ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is Mathias rendered as a stark, black silhouette with stiff, sharp posture against the chaotic water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030902922153473 + ], + "opposite_logprob_tokens": [ + -2.350309133529663 + ], + "logprob_reward": -0.10030902922153473, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030902922153473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19650760293006897, + 0.06962493807077408, + 0.07621266692876816, + 0.11841583251953125, + 0.09097325801849365, + 0.06993669271469116, + 0.12391941249370575, + 0.1577354073524475, + 0.0651913434267044, + 0.03148285672068596 + ], + "credit_map": [ + 0.016897549852728844, + 0.031958289444446564, + 0.020294377580285072, + 0.025692228227853775, + 0.024220174178481102, + 0.024067221209406853, + 0.02958141267299652, + 0.023796379566192627, + 0.005601959303021431, + 0.007691439241170883, + 0.007135320920497179, + 0.012557749636471272, + 0.006935997400432825, + 0.008231094107031822, + 0.013035203330218792, + 0.008436188101768494, + 0.005105548072606325, + 0.007114397827535868, + 0.007949715480208397, + 0.01262577436864376, + 0.006671817973256111, + 0.008091053925454617, + 0.017191464081406593, + 0.011462900787591934, + 0.009750096127390862, + 0.011918804608285427, + 0.012523233890533447, + 0.0222786758095026, + 0.009673260152339935, + 0.011047027073800564, + 0.019860275089740753, + 0.021364469081163406, + 0.009355161339044571, + 0.00954527873545885, + 0.013461759313941002, + 0.017758075147867203, + 0.009285477921366692, + 0.006925324909389019, + 0.015082928352057934, + 0.009559258818626404, + 0.006148496177047491, + 0.007144964765757322, + 0.008985757827758789, + 0.01729283481836319, + 0.005503347609192133, + 0.003995944745838642, + 0.014255308546125889, + 0.0066100480034947395, + 0.013539626263082027, + 0.013356848619878292, + 0.015836812555789948, + 0.012318126857280731, + 0.025289857760071754, + 0.013513549230992794, + 0.01758117415010929, + 0.012483429163694382, + 0.01768597960472107, + 0.021734746173024178, + 0.0262710340321064, + 0.014204561710357666, + 0.034603893756866455, + 0.015067396685481071, + 0.010820059105753899, + 0.017347747460007668, + 0.007555193733423948, + 0.007536991499364376, + 0.011141911149024963, + 0.008746934123337269, + 0.0089869424700737, + 0.0047598001547157764, + 0.008510499261319637, + 0.007953076623380184, + 0.002927841618657112, + 0.003280692733824253, + 0.005016428418457508, + 0.00409270403906703, + 0.004402474034577608, + 0.00302925705909729, + 0.004474388901144266, + 0.0042590731754899025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.10030902922153473 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate canyon with steep, geometric walls of dark, damp rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a raging, slate-gray torrent of water thick with mud and debris surging through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while tracking the flood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary figure of Mathias visible standing completely still on a high, flat ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is Mathias rendered as a stark, black silhouette with stiff, sharp posture against the chaotic water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014208345673978329 + ], + "opposite_logprob_tokens": [ + -4.2642083168029785 + ], + "logprob_reward": -0.014208345673978329, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014208345673978329, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1513272076845169, + 0.052090834826231, + 0.06239037960767746, + 0.08723243325948715, + 0.050234802067279816, + 0.041862744837999344, + 0.13892436027526855, + 0.14988577365875244, + 0.15468381345272064, + 0.11136763542890549 + ], + "credit_map": [ + 0.017175253480672836, + 0.01762809418141842, + 0.018360000103712082, + 0.023176025599241257, + 0.021326567977666855, + 0.021398840472102165, + 0.01432131789624691, + 0.01794111356139183, + 0.003936032298952341, + 0.00664140097796917, + 0.005125389900058508, + 0.008688545785844326, + 0.005151902325451374, + 0.00597991980612278, + 0.011025207117199898, + 0.005542435217648745, + 0.0037462422624230385, + 0.0049408501945436, + 0.00488986587151885, + 0.014690876007080078, + 0.006059620063751936, + 0.0051018670201301575, + 0.013679021038115025, + 0.009282037615776062, + 0.007734483107924461, + 0.014002303592860699, + 0.008692964911460876, + 0.012740159407258034, + 0.012785320170223713, + 0.009369618259370327, + 0.012105287984013557, + 0.009802293963730335, + 0.004631989635527134, + 0.007074391469359398, + 0.008439111523330212, + 0.006852302700281143, + 0.0045921639539301395, + 0.0034202090464532375, + 0.006597165949642658, + 0.008627469651401043, + 0.003322368487715721, + 0.0024503800086677074, + 0.006328823510557413, + 0.007916847243905067, + 0.005730656441301107, + 0.003774204757064581, + 0.005011966452002525, + 0.007327500265091658, + 0.01569944992661476, + 0.01561728585511446, + 0.0159237589687109, + 0.014325519092381, + 0.03213486447930336, + 0.019270796328783035, + 0.013790305703878403, + 0.012162378989160061, + 0.02414891868829727, + 0.019685819745063782, + 0.017948545515537262, + 0.014049249701201916, + 0.03022642992436886, + 0.014156829565763474, + 0.013896125368773937, + 0.01577385887503624, + 0.019938083365559578, + 0.02033575437963009, + 0.02286021038889885, + 0.016664069145917892, + 0.02616526186466217, + 0.013066775165498257, + 0.018114188686013222, + 0.01753947138786316, + 0.00831823144108057, + 0.01969936490058899, + 0.012765432707965374, + 0.009052406065165997, + 0.029544135555624962, + 0.009191920980811119, + 0.01368971448391676, + 0.009106426499783993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.014208345673978329 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate canyon with steep, geometric walls of dark, damp rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a raging, slate-gray torrent of water thick with mud and debris surging through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while tracking the flood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary figure of Mathias visible standing completely still on a high, flat ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is Mathias rendered as a stark, black silhouette with stiff, sharp posture against the chaotic water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018197255209088326 + ], + "opposite_logprob_tokens": [ + -4.018197059631348 + ], + "logprob_reward": -0.018197255209088326, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018197255209088326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17938047647476196, + 0.06478668749332428, + 0.08085761219263077, + 0.10095440596342087, + 0.07233373820781708, + 0.05929415673017502, + 0.1449969857931137, + 0.1413566768169403, + 0.08963432908058167, + 0.06640494614839554 + ], + "credit_map": [ + 0.01572752743959427, + 0.02375251054763794, + 0.02136337198317051, + 0.030257459729909897, + 0.023457296192646027, + 0.021832354366779327, + 0.02068140357732773, + 0.02230852097272873, + 0.005911672487854958, + 0.008335091173648834, + 0.0057023982517421246, + 0.008643575012683868, + 0.006717359181493521, + 0.006550916936248541, + 0.015062435530126095, + 0.007863233797252178, + 0.005349683575332165, + 0.005046271253377199, + 0.006825115531682968, + 0.01624717190861702, + 0.007374410517513752, + 0.0077347769401967525, + 0.019144108518958092, + 0.013136066496372223, + 0.008155805058777332, + 0.013847127556800842, + 0.010692333802580833, + 0.014853445813059807, + 0.013742503710091114, + 0.009568762965500355, + 0.0181503314524889, + 0.011944090947508812, + 0.005663411226123571, + 0.00833925697952509, + 0.011093503795564175, + 0.01130669191479683, + 0.009039945900440216, + 0.005234301555901766, + 0.009545345790684223, + 0.012111271731555462, + 0.005228299181908369, + 0.003685396397486329, + 0.007763933390378952, + 0.012868798337876797, + 0.006345614790916443, + 0.005924382247030735, + 0.007980801165103912, + 0.009496927261352539, + 0.013475298881530762, + 0.015968458727002144, + 0.01583571545779705, + 0.020803038030862808, + 0.02915145270526409, + 0.017502561211586, + 0.018461639061570168, + 0.013798825442790985, + 0.021398095414042473, + 0.020101282745599747, + 0.0165091585367918, + 0.011780413798987865, + 0.03210381418466568, + 0.01344215776771307, + 0.012089600786566734, + 0.013932138681411743, + 0.011871688067913055, + 0.009937768802046776, + 0.017422467470169067, + 0.012597441673278809, + 0.011015322990715504, + 0.00787461269646883, + 0.007321047596633434, + 0.011593975126743317, + 0.00633304612711072, + 0.00853691529482603, + 0.012594440951943398, + 0.006827317178249359, + 0.011937569826841354, + 0.00548031460493803, + 0.006755269598215818, + 0.007940061390399933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.018197255209088326 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stark, minimalist artistic style, a wide shot opens with the camera panning slowly and steadily to the right, tracking the raw, elemental force of a flash flood. A raging, slate-gray torrent of water, thick with mud and debris, surges through the narrow confines of a desolate canyon, whose steep, geometric walls of dark, damp rock are rendered in a desaturated and muted palette. The camera continues in a deliberate pan, tilting slightly upward to reveal a high, flat ledge jutting out from the cliff face. Here, the solitary figure of Mathias stands completely still, his form a stark, black silhouette against the chaotic, bubbling water far below. His posture is stiff and sharp, his shoulders squared, his tall frame set against the devastation. Against the flat, unforgiving light of the overcast sky, his facial features are completely lost in deep shadow, rendering his expression completely unreadable as he silently and motionlessly watches the disaster unfold, creating a powerful and unsettling atmosphere of detached witness.", + "seed": 25742, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_14/index_3819_seed_25742_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04696747101843357, + "vlm_logprob": -0.04696747101843357, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.04696747101843357, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate canyon with steep, geometric walls of dark, damp rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a raging, slate-gray torrent of water thick with mud and debris surging through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while tracking the flood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary figure of Mathias visible standing completely still on a high, flat ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is Mathias rendered as a stark, black silhouette with stiff, sharp posture against the chaotic water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03091256134212017 + ], + "opposite_logprob_tokens": [ + -3.7809126377105713 + ], + "logprob_reward": -0.03091256134212017, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.03091256134212017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21848249435424805, + 0.12447050958871841, + 0.041761450469493866, + 0.13782183825969696, + 0.04982805997133255, + 0.1151430755853653, + 0.09539250284433365, + 0.08290822058916092, + 0.10900640487670898, + 0.025185437873005867 + ], + "credit_map": [ + 0.011746235191822052, + 0.01283278875052929, + 0.028425144031643867, + 0.04018517583608627, + 0.016357246786355972, + 0.027858948335051537, + 0.028948573395609856, + 0.052128393203020096, + 0.00745505653321743, + 0.010472496040165424, + 0.025819504633545876, + 0.014197949320077896, + 0.01672746241092682, + 0.015128913335502148, + 0.016986289992928505, + 0.017682833597064018, + 0.0033825039863586426, + 0.003474550787359476, + 0.008395570330321789, + 0.005486604291945696, + 0.0050579034723341465, + 0.005109724123030901, + 0.005418852902948856, + 0.005435737781226635, + 0.007041488774120808, + 0.004082805942744017, + 0.05712651088833809, + 0.0456550307571888, + 0.002983781509101391, + 0.004214028362184763, + 0.006325715221464634, + 0.010392477735877037, + 0.005730273202061653, + 0.004611715208739042, + 0.005862940568476915, + 0.0049156020395457745, + 0.006602280307561159, + 0.008200250566005707, + 0.008313651196658611, + 0.0055913496762514114, + 0.03259426727890968, + 0.01753225177526474, + 0.005727441515773535, + 0.005731421057134867, + 0.0078016621991992, + 0.01202105451375246, + 0.008458645083010197, + 0.02527633123099804, + 0.016314610838890076, + 0.016618462279438972, + 0.017248082906007767, + 0.009640916250646114, + 0.0054110754281282425, + 0.008254758082330227, + 0.006988792214542627, + 0.01491580531001091, + 0.01148271095007658, + 0.012899555265903473, + 0.011545606888830662, + 0.01065780594944954, + 0.005222178529947996, + 0.008210957050323486, + 0.00875927321612835, + 0.014130128547549248, + 0.015056791715323925, + 0.0164314154535532, + 0.0202949158847332, + 0.016101248562335968, + 0.005964613053947687, + 0.013735492713749409, + 0.007987611927092075, + 0.013434316031634808, + 0.004148423206061125, + 0.004298303741961718, + 0.0027359449304640293, + 0.002143623773008585, + 0.0028367468621581793, + 0.003570920554921031, + 0.0022001375909894705, + 0.003251336980611086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.03091256134212017 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate canyon with steep, geometric walls of dark, damp rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a raging, slate-gray torrent of water thick with mud and debris surging through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while tracking the flood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary figure of Mathias visible standing completely still on a high, flat ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is Mathias rendered as a stark, black silhouette with stiff, sharp posture against the chaotic water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023313812911510468 + ], + "opposite_logprob_tokens": [ + -3.7733137607574463 + ], + "logprob_reward": -0.023313812911510468, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023313812911510468, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20910659432411194, + 0.06656499207019806, + 0.041511420160532, + 0.3922956585884094, + 0.03919578716158867, + 0.04611808434128761, + 0.08524567633867264, + 0.04810598865151405, + 0.05348546430468559, + 0.01837041601538658 + ], + "credit_map": [ + 0.008055880665779114, + 0.012369886972010136, + 0.023901546373963356, + 0.04199657961726189, + 0.012156262993812561, + 0.026112738996744156, + 0.024220556020736694, + 0.06029313802719116, + 0.004862618166953325, + 0.005947378929704428, + 0.011706716381013393, + 0.007132010534405708, + 0.011364984326064587, + 0.008139891549944878, + 0.008665741421282291, + 0.00874563679099083, + 0.003650016151368618, + 0.0031688432209193707, + 0.007503292988985777, + 0.004509628284722567, + 0.006231151055544615, + 0.004890936892479658, + 0.006224336568266153, + 0.005333211272954941, + 0.021026713773608208, + 0.010555283166468143, + 0.1430475264787674, + 0.1026657447218895, + 0.012456261552870274, + 0.0251382477581501, + 0.02269314043223858, + 0.054712649434804916, + 0.004980415105819702, + 0.003639523172751069, + 0.004625455942004919, + 0.0038355772849172354, + 0.004366401582956314, + 0.006531103979796171, + 0.006281290203332901, + 0.004936016630381346, + 0.010991145856678486, + 0.012001180090010166, + 0.003258373821154237, + 0.0036190610844641924, + 0.0034247711300849915, + 0.004576202016323805, + 0.004278542473912239, + 0.003968806006014347, + 0.013157729990780354, + 0.016734924167394638, + 0.016955826431512833, + 0.00915514212101698, + 0.005597442854195833, + 0.007647170685231686, + 0.007833786308765411, + 0.008163654245436192, + 0.006309362594038248, + 0.007717080879956484, + 0.006261860020458698, + 0.0075640506111085415, + 0.002848444739356637, + 0.005359921138733625, + 0.004818170331418514, + 0.007227092981338501, + 0.010219424962997437, + 0.009981644339859486, + 0.00595534173771739, + 0.006707377731800079, + 0.004151000641286373, + 0.006584967952221632, + 0.003688827157020569, + 0.006196870934218168, + 0.003573882859200239, + 0.0030116138514131308, + 0.003339091781526804, + 0.002209305763244629, + 0.0011881834361702204, + 0.001502794330008328, + 0.0014744411455467343, + 0.0020711023826152086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.023313812911510468 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate canyon with steep, geometric walls of dark, damp rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a raging, slate-gray torrent of water thick with mud and debris surging through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while tracking the flood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary figure of Mathias visible standing completely still on a high, flat ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is Mathias rendered as a stark, black silhouette with stiff, sharp posture against the chaotic water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062053997069597244 + ], + "opposite_logprob_tokens": [ + -2.812053918838501 + ], + "logprob_reward": -0.062053997069597244, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062053997069597244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14568321406841278, + 0.022194989025592804, + 0.05701236054301262, + 0.2888326942920685, + 0.01930590532720089, + 0.144965261220932, + 0.040034133940935135, + 0.04946175217628479, + 0.1075039654970169, + 0.12500575184822083 + ], + "credit_map": [ + 0.009956371039152145, + 0.009832552634179592, + 0.021593576297163963, + 0.022430183365941048, + 0.012723881751298904, + 0.016964899376034737, + 0.018177835270762444, + 0.03400389105081558, + 0.0018812738126143813, + 0.00162382738199085, + 0.0034831915982067585, + 0.0023984762374311686, + 0.0035034080501645803, + 0.003078359179198742, + 0.0024349510204046965, + 0.0037915008142590523, + 0.005287675652652979, + 0.004644700791686773, + 0.009329079650342464, + 0.00906047597527504, + 0.006861438509076834, + 0.007783693727105856, + 0.008115368895232677, + 0.005929921753704548, + 0.013579891994595528, + 0.006803793832659721, + 0.11752861738204956, + 0.08898800611495972, + 0.00607403414323926, + 0.01077860128134489, + 0.01071583479642868, + 0.03436388820409775, + 0.0025285088922828436, + 0.0019686424639075994, + 0.0026513608172535896, + 0.0019710883498191833, + 0.0019533671438694, + 0.003771389601752162, + 0.002202372532337904, + 0.0022591750603169203, + 0.014858798123896122, + 0.05038732662796974, + 0.012138509191572666, + 0.007921157404780388, + 0.005472489167004824, + 0.007870438508689404, + 0.005434741731733084, + 0.04088179022073746, + 0.008325371891260147, + 0.006201016716659069, + 0.0035929998848587275, + 0.003740801475942135, + 0.002155747963115573, + 0.0036451946943998337, + 0.003907527308911085, + 0.008465470746159554, + 0.0076965028420090675, + 0.009045013226568699, + 0.005583262536674738, + 0.00776439206674695, + 0.0033403628040105104, + 0.004183060023933649, + 0.005507888738065958, + 0.006341270171105862, + 0.013928171247243881, + 0.01580503024160862, + 0.01910186931490898, + 0.016107426956295967, + 0.0066721816547214985, + 0.016410427168011665, + 0.007979532703757286, + 0.011499322019517422, + 0.019016997888684273, + 0.018269522115588188, + 0.015384216792881489, + 0.012043259106576443, + 0.012503844685852528, + 0.019753340631723404, + 0.011965667828917503, + 0.01606888696551323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.062053997069597244 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate canyon with steep, geometric walls of dark, damp rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a raging, slate-gray torrent of water thick with mud and debris surging through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while tracking the flood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary figure of Mathias visible standing completely still on a high, flat ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is Mathias rendered as a stark, black silhouette with stiff, sharp posture against the chaotic water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10034989565610886 + ], + "opposite_logprob_tokens": [ + -2.3503499031066895 + ], + "logprob_reward": -0.10034989565610886, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10034989565610886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10972387343645096, + 0.050790831446647644, + 0.0323440283536911, + 0.15693874657154083, + 0.058142755180597305, + 0.052634574472904205, + 0.05655398219823837, + 0.33848637342453003, + 0.017348110675811768, + 0.12703672051429749 + ], + "credit_map": [ + 0.006677940022200346, + 0.006457325536757708, + 0.011938989162445068, + 0.021423202008008957, + 0.005292567890137434, + 0.013792653568089008, + 0.011743627488613129, + 0.032397571951150894, + 0.003740511601790786, + 0.003954616375267506, + 0.010194297879934311, + 0.006694458425045013, + 0.005280588753521442, + 0.007060063071548939, + 0.006666660774499178, + 0.007199633400887251, + 0.0024512626696377993, + 0.0021900625433772802, + 0.0069362628273665905, + 0.004885951988399029, + 0.0035580135881900787, + 0.004379692487418652, + 0.004212368745356798, + 0.003730416065081954, + 0.011379517614841461, + 0.005864860024303198, + 0.05842869356274605, + 0.05441916361451149, + 0.0036049801856279373, + 0.005459974519908428, + 0.007078663911670446, + 0.010702898725867271, + 0.006623817142099142, + 0.006248616613447666, + 0.006296210456639528, + 0.005335735157132149, + 0.008089100942015648, + 0.010081804357469082, + 0.00909380428493023, + 0.006373669020831585, + 0.005925973877310753, + 0.015353251248598099, + 0.004422367084771395, + 0.0029309948440641165, + 0.003510940819978714, + 0.0046326168812811375, + 0.0036934365052729845, + 0.012164990417659283, + 0.008971711620688438, + 0.013480165041983128, + 0.005009705200791359, + 0.004784161224961281, + 0.004058998078107834, + 0.006199034396559, + 0.004447256680577993, + 0.00960294995456934, + 0.05615560710430145, + 0.06092865392565727, + 0.03529515489935875, + 0.045556940138339996, + 0.02049468271434307, + 0.04200068861246109, + 0.030688494443893433, + 0.047366149723529816, + 0.0029743053019046783, + 0.0029839975759387016, + 0.0018342725234106183, + 0.0016041543567553163, + 0.001774344826117158, + 0.0019176916684955359, + 0.0019880738109350204, + 0.002271271077916026, + 0.018647396937012672, + 0.03821927309036255, + 0.02482498437166214, + 0.007911136373877525, + 0.010297869332134724, + 0.010316831059753895, + 0.006544904317706823, + 0.010274315252900124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 9 + ], + "selected_reward": -0.10034989565610886 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate canyon with steep, geometric walls of dark, damp rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a raging, slate-gray torrent of water thick with mud and debris surging through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while tracking the flood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary figure of Mathias visible standing completely still on a high, flat ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is Mathias rendered as a stark, black silhouette with stiff, sharp posture against the chaotic water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018207088112831116 + ], + "opposite_logprob_tokens": [ + -4.01820707321167 + ], + "logprob_reward": -0.018207088112831116, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018207088112831116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20400293171405792, + 0.05462440848350525, + 0.06728751957416534, + 0.15859709680080414, + 0.11706672608852386, + 0.1194101870059967, + 0.07342672348022461, + 0.1303936243057251, + 0.03236778825521469, + 0.042822953313589096 + ], + "credit_map": [ + 0.016948552802205086, + 0.013559396378695965, + 0.032869793474674225, + 0.034076791256666183, + 0.01350939180701971, + 0.02409169264137745, + 0.020158277824521065, + 0.04878905043005943, + 0.004692834336310625, + 0.0038898522034287453, + 0.011767265386879444, + 0.0072800214402377605, + 0.005346489138901234, + 0.0069932169280946255, + 0.006813652813434601, + 0.007841083221137524, + 0.00468746293336153, + 0.00465116323903203, + 0.016010761260986328, + 0.011622167192399502, + 0.008238151669502258, + 0.006740200333297253, + 0.00882016308605671, + 0.006517454050481319, + 0.007435594219714403, + 0.004138309974223375, + 0.06700771301984787, + 0.052865758538246155, + 0.002337425947189331, + 0.004786694422364235, + 0.00606053089722991, + 0.013965071178972721, + 0.015075144357979298, + 0.011958186514675617, + 0.012258514761924744, + 0.011148068122565746, + 0.0148001192137599, + 0.021635785698890686, + 0.017227452248334885, + 0.012963458895683289, + 0.02252854034304619, + 0.04522410035133362, + 0.0026198518462479115, + 0.0037094473373144865, + 0.0047369603998959064, + 0.007130660582333803, + 0.004095109179615974, + 0.029365528374910355, + 0.013617096468806267, + 0.01704859547317028, + 0.005960279144346714, + 0.006968101020902395, + 0.0045312740840017796, + 0.006533393636345863, + 0.005402614362537861, + 0.01336537767201662, + 0.021446099504828453, + 0.02261076308786869, + 0.014361954294145107, + 0.0178156029433012, + 0.008539318107068539, + 0.015148245729506016, + 0.012615951709449291, + 0.017855700105428696, + 0.0036155583802610636, + 0.004216139670461416, + 0.008431263267993927, + 0.007054568268358707, + 0.0015103454934433103, + 0.0023233441170305014, + 0.001921808230690658, + 0.003294761525467038, + 0.005055010784417391, + 0.007645966485142708, + 0.006765438709408045, + 0.0028486137744039297, + 0.005458876024931669, + 0.006409649737179279, + 0.0034181284718215466, + 0.005221272353082895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.018207088112831116 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_15/index_tensor([717, 717])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_15/index_tensor([717, 717])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3e0b6b9f69e4d4915c297bc979f627547710498f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_15/index_tensor([717, 717])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium close-up shot, shot from a level angle, slowly and steadily inward, draws the viewer's attention to a white-haired, elderly craftsman. He sits in a woodworking workshop filled with the mingled aroma of sawdust and varnish. His face is deeply wrinkled, his eyes, through a pair of reading glasses, fixed with intense concentration on his work. A beam of warm afternoon sunlight slants in through the high windows, casting a visible shaft of light in the air, illuminating the dancing dust particles. He wears a washed-out blue cotton work suit, leaning slightly forward, his calloused yet remarkably steady hands meticulously working: his right hand tightly grips a carving knife, its handle smooth and warm from years of use, while his left hand carefully supports the elegantly shaped pearwood body of a pipa. With steady and deliberate force from his wrist, the sharp tip of the knife glided across the delicate, luminous wood of the pipa's back. Slender wood petals curled and gently lifted, gradually revealing the intricate and lifelike outline of a peony. The sunlight precisely illuminated the pipa and his deft hands, making every carved grain visible. A striking contrast of light and shadow formed against the warm wood tone, and the entire scene was permeated with an atmosphere of profound tranquility, almost sacred, and the spirit of craftsmanship.", + "seed": 14728, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_15/index_717_seed_14728_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.09585612490773202, + "vlm_logprob": -0.09585612490773202, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.09585612490773202, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up filmed from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elderly craftsman with white hair wearing a washed-out blue cotton work suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman carving a pearwood pipa with a carving knife to reveal a peony design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a shaft of light through dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11396412551403046 + ], + "opposite_logprob_tokens": [ + -4.363964080810547 + ], + "logprob_reward": -0.11396412551403046, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.11396412551403046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2937964200973511, + 0.05080687627196312, + 0.04380854591727257, + 0.08683417737483978, + 0.0640421211719513, + 0.09752126038074493, + 0.18163780868053436, + 0.05209539085626602, + 0.05521685257554054, + 0.07424051314592361 + ], + "credit_map": [ + 0.031000206246972084, + 0.05272489786148071, + 0.03747003152966499, + 0.030480777844786644, + 0.05576779693365097, + 0.024359123781323433, + 0.02070501260459423, + 0.04128856211900711, + 0.00451658945530653, + 0.00869921687990427, + 0.012896825559437275, + 0.005166163202375174, + 0.0047156913205981255, + 0.003501914208754897, + 0.004398017190396786, + 0.006912458688020706, + 0.005946261342614889, + 0.008146625012159348, + 0.004912619479000568, + 0.0069015431217849255, + 0.0041802749037742615, + 0.004220345988869667, + 0.00503886491060257, + 0.0044620102271437645, + 0.006362392101436853, + 0.007821785286068916, + 0.00702010840177536, + 0.022079559043049812, + 0.007303976453840733, + 0.0066522350534796715, + 0.01686040684580803, + 0.012733719311654568, + 0.008799356408417225, + 0.011737297289073467, + 0.008046510629355907, + 0.006098437588661909, + 0.005988409277051687, + 0.005078492686152458, + 0.011595978401601315, + 0.006697638891637325, + 0.022093351930379868, + 0.011090206913650036, + 0.010596645064651966, + 0.010758927091956139, + 0.005800948012620211, + 0.007423107977956533, + 0.011550561524927616, + 0.01820751465857029, + 0.02259019762277603, + 0.02371547371149063, + 0.022567423060536385, + 0.023672498762607574, + 0.020251335576176643, + 0.020511597394943237, + 0.030884381383657455, + 0.017444897443056107, + 0.006914965342730284, + 0.010480260476469994, + 0.005716878455132246, + 0.003999078646302223, + 0.00555335683748126, + 0.0054147010669112206, + 0.00922232586890459, + 0.004793825559318066, + 0.008464149199426174, + 0.008925207890570164, + 0.005700991488993168, + 0.007205873262137175, + 0.0047205048613250256, + 0.005469507537782192, + 0.0068159522488713264, + 0.007914666086435318, + 0.006802814546972513, + 0.007207684218883514, + 0.005241623613983393, + 0.006615085992962122, + 0.00554176839068532, + 0.006175006274133921, + 0.01700272411108017, + 0.0196538008749485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.11396412551403046 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up filmed from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elderly craftsman with white hair wearing a washed-out blue cotton work suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman carving a pearwood pipa with a carving knife to reveal a peony design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a shaft of light through dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22559396922588348 + ], + "opposite_logprob_tokens": [ + -1.6005939245224 + ], + "logprob_reward": -0.22559396922588348, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22559396922588348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12159420549869537, + 0.07159123569726944, + 0.05273981764912605, + 0.0904959887266159, + 0.09484665840864182, + 0.24258999526500702, + 0.1251973807811737, + 0.03891393914818764, + 0.09649103879928589, + 0.06553974747657776 + ], + "credit_map": [ + 0.015538333915174007, + 0.015832798555493355, + 0.010322846472263336, + 0.012798634357750416, + 0.026881106197834015, + 0.01296879444271326, + 0.012610603123903275, + 0.014641091227531433, + 0.008323701098561287, + 0.007531001232564449, + 0.01038207858800888, + 0.008920150808990002, + 0.013024434447288513, + 0.006828406360000372, + 0.006643532309681177, + 0.009937936440110207, + 0.007387432735413313, + 0.0065907384268939495, + 0.005674493499100208, + 0.006385062355548143, + 0.00706713180989027, + 0.006643050350248814, + 0.006587619427591562, + 0.006404287181794643, + 0.013512294739484787, + 0.012837992049753666, + 0.011903532780706882, + 0.01333134900778532, + 0.00759724248200655, + 0.007589373271912336, + 0.009937793016433716, + 0.013786409981548786, + 0.012907955795526505, + 0.009950373321771622, + 0.010852234438061714, + 0.019945144653320312, + 0.008899690583348274, + 0.008939032442867756, + 0.009453910402953625, + 0.013898315839469433, + 0.07183370739221573, + 0.02955777198076248, + 0.02828475832939148, + 0.02336646057665348, + 0.01305379718542099, + 0.01726563833653927, + 0.023261163383722305, + 0.03596668690443039, + 0.013904848136007786, + 0.017019636929035187, + 0.014996204525232315, + 0.016692036762833595, + 0.014643528498709202, + 0.01588897965848446, + 0.019360579550266266, + 0.012691561132669449, + 0.006620076019316912, + 0.005702674854546785, + 0.004073748365044594, + 0.004195458721369505, + 0.00382819096557796, + 0.004831417929381132, + 0.006714747287333012, + 0.0029476266354322433, + 0.01076340489089489, + 0.011873606592416763, + 0.009113376960158348, + 0.012016357854008675, + 0.008912567049264908, + 0.01137805636972189, + 0.018767379224300385, + 0.013666295446455479, + 0.008726566098630428, + 0.00763027323409915, + 0.008142832666635513, + 0.009443211369216442, + 0.006596249993890524, + 0.006172734312713146, + 0.007968995720148087, + 0.01085888035595417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.22559396922588348 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up filmed from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elderly craftsman with white hair wearing a washed-out blue cotton work suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman carving a pearwood pipa with a carving knife to reveal a peony design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a shaft of light through dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981371060013771 + ], + "opposite_logprob_tokens": [ + -3.529813766479492 + ], + "logprob_reward": -0.02981371060013771, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981371060013771, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3690567910671234, + 0.050444260239601135, + 0.05274421349167824, + 0.06765129417181015, + 0.08573919534683228, + 0.042232804000377655, + 0.11243364214897156, + 0.04581846669316292, + 0.13943161070346832, + 0.034447744488716125 + ], + "credit_map": [ + 0.03986009955406189, + 0.06185450032353401, + 0.041408684104681015, + 0.042686332017183304, + 0.06871356815099716, + 0.033202316612005234, + 0.02666453644633293, + 0.054666757583618164, + 0.007846309803426266, + 0.008390102535486221, + 0.006310670170933008, + 0.0046632131561636925, + 0.00744178332388401, + 0.004246933851391077, + 0.005016807932406664, + 0.006528444122523069, + 0.008562026545405388, + 0.008098267950117588, + 0.005494060460478067, + 0.005723167676478624, + 0.008381967432796955, + 0.00395437004044652, + 0.005775618366897106, + 0.0067547401413321495, + 0.0065942625515162945, + 0.006699070334434509, + 0.006322028581053019, + 0.01835578866302967, + 0.005740534979850054, + 0.005331729538738728, + 0.01036379486322403, + 0.008244090713560581, + 0.014232847839593887, + 0.013111164793372154, + 0.010375313460826874, + 0.013224690221250057, + 0.009177230298519135, + 0.005540374666452408, + 0.012462826445698738, + 0.0076147522777318954, + 0.00506492517888546, + 0.006092326249927282, + 0.006240570917725563, + 0.006652497220784426, + 0.003846112871542573, + 0.004865054972469807, + 0.003909775987267494, + 0.005561545491218567, + 0.012201531790196896, + 0.015566835179924965, + 0.01460228580981493, + 0.014173733070492744, + 0.013664464466273785, + 0.012202715501189232, + 0.01893692836165428, + 0.01108515728265047, + 0.008213046006858349, + 0.007268174551427364, + 0.00534733384847641, + 0.005196091253310442, + 0.004990594927221537, + 0.004242048133164644, + 0.006884502712637186, + 0.003676676657050848, + 0.015398457646369934, + 0.018348328769207, + 0.014526844024658203, + 0.01664816215634346, + 0.012061292305588722, + 0.015666471794247627, + 0.0264564361423254, + 0.02032562717795372, + 0.0032003717496991158, + 0.0037353760562837124, + 0.002846034476533532, + 0.003369625424966216, + 0.002272641519084573, + 0.002699692267924547, + 0.007927346974611282, + 0.008396657183766365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.02981371060013771 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up filmed from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elderly craftsman with white hair wearing a washed-out blue cotton work suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman carving a pearwood pipa with a carving knife to reveal a peony design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a shaft of light through dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496332421898842 + ], + "opposite_logprob_tokens": [ + -2.9299633502960205 + ], + "logprob_reward": -0.05496332421898842, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496332421898842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3301639258861542, + 0.09268613904714584, + 0.026753680780529976, + 0.11059338599443436, + 0.0893525630235672, + 0.02231963723897934, + 0.042834244668483734, + 0.025672418996691704, + 0.08634929358959198, + 0.17327475547790527 + ], + "credit_map": [ + 0.03307894989848137, + 0.04933960735797882, + 0.042793452739715576, + 0.029371215030550957, + 0.05639641359448433, + 0.032920435070991516, + 0.049599241465330124, + 0.036664608865976334, + 0.010667867958545685, + 0.014119302853941917, + 0.01104368269443512, + 0.01426217332482338, + 0.013013756833970547, + 0.0071159168146550655, + 0.011773388832807541, + 0.010690049268305302, + 0.003280974691733718, + 0.0032987170852720737, + 0.002848468255251646, + 0.003577516647055745, + 0.0032774785067886114, + 0.0036808319855481386, + 0.0038469587452709675, + 0.0029427348636090755, + 0.007906390354037285, + 0.010500100441277027, + 0.008920741267502308, + 0.029062090441584587, + 0.009722757153213024, + 0.011184558272361755, + 0.019955072551965714, + 0.013341674581170082, + 0.019351640716195107, + 0.009917228482663631, + 0.00880695041269064, + 0.008769800886511803, + 0.00796476285904646, + 0.006071588024497032, + 0.019094061106443405, + 0.009376531466841698, + 0.0033901578281074762, + 0.0036855118814855814, + 0.003030172549188137, + 0.0027221376076340675, + 0.0019765268079936504, + 0.003013959154486656, + 0.0026244937907904387, + 0.001876678317785263, + 0.005920915398746729, + 0.007729844190180302, + 0.004814565181732178, + 0.004946704488247633, + 0.004655638709664345, + 0.005353822372853756, + 0.005143179092556238, + 0.004269576631486416, + 0.004020450171083212, + 0.0028909549582749605, + 0.0028750605415552855, + 0.0030316694173961878, + 0.0027321777306497097, + 0.004340221639722586, + 0.0036199402529746294, + 0.002161943819373846, + 0.011236532591283321, + 0.010783661156892776, + 0.010221617296338081, + 0.010767221450805664, + 0.00865760538727045, + 0.009983022697269917, + 0.013953114859759808, + 0.01074652187526226, + 0.015008583664894104, + 0.014307834208011627, + 0.01223578117787838, + 0.011837946251034737, + 0.012179131619632244, + 0.009749343618750572, + 0.04869832098484039, + 0.04925782233476639 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.05496332421898842 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up filmed from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elderly craftsman with white hair wearing a washed-out blue cotton work suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman carving a pearwood pipa with a carving knife to reveal a peony design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a shaft of light through dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494549497961998 + ], + "opposite_logprob_tokens": [ + -2.929945468902588 + ], + "logprob_reward": -0.05494549497961998, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494549497961998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3327823579311371, + 0.10765872150659561, + 0.11263889819383621, + 0.05318625643849373, + 0.056172795593738556, + 0.046950265765190125, + 0.06696643680334091, + 0.022130604833364487, + 0.0655464455485344, + 0.13596729934215546 + ], + "credit_map": [ + 0.04654248431324959, + 0.04175674542784691, + 0.04368816316127777, + 0.03283107280731201, + 0.06518609076738358, + 0.03041169047355652, + 0.02556033618748188, + 0.0468057282269001, + 0.013946323655545712, + 0.01798831857740879, + 0.01465127244591713, + 0.010807814076542854, + 0.015519104897975922, + 0.011334280483424664, + 0.011191551573574543, + 0.012220057658851147, + 0.013465240597724915, + 0.010908672586083412, + 0.014293238520622253, + 0.01653175801038742, + 0.01406213827431202, + 0.012191589921712875, + 0.016328057274222374, + 0.0148581862449646, + 0.0074344202876091, + 0.007414726074784994, + 0.00682320399209857, + 0.006142761558294296, + 0.006027622148394585, + 0.00537849823012948, + 0.006021939218044281, + 0.007943078875541687, + 0.006036939099431038, + 0.010831665247678757, + 0.007080542854964733, + 0.00473015708848834, + 0.0050893197767436504, + 0.0033466401509940624, + 0.009851401671767235, + 0.009206120856106281, + 0.01166995894163847, + 0.006310863420367241, + 0.005212835036218166, + 0.004670930560678244, + 0.003722707973793149, + 0.004149822052568197, + 0.007159500848501921, + 0.00405364204198122, + 0.006896134465932846, + 0.009788772091269493, + 0.008782549761235714, + 0.00943534355610609, + 0.008973605930805206, + 0.007638290990144014, + 0.008647155947983265, + 0.006804578006267548, + 0.004499686881899834, + 0.003041928168386221, + 0.002983671147376299, + 0.0018972203833982348, + 0.001933688996359706, + 0.0023408427368849516, + 0.0034238495863974094, + 0.0020097140222787857, + 0.0058443015441298485, + 0.009504914283752441, + 0.006516769528388977, + 0.009042421355843544, + 0.005757227074354887, + 0.0087709566578269, + 0.012115196324884892, + 0.007994645275175571, + 0.014942649751901627, + 0.012180978432297707, + 0.010990611277520657, + 0.011811358854174614, + 0.011504297144711018, + 0.011778987012803555, + 0.029699089005589485, + 0.033059313893318176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.05494549497961998 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot, shot from a level angle, slowly and steadily inward, draws the viewer's attention to a white-haired, elderly craftsman. He sits in a woodworking workshop filled with the mingled aroma of sawdust and varnish. His face is deeply wrinkled, his eyes, through a pair of reading glasses, fixed with intense concentration on his work. A beam of warm afternoon sunlight slants in through the high windows, casting a visible shaft of light in the air, illuminating the dancing dust particles. He wears a washed-out blue cotton work suit, leaning slightly forward, his calloused yet remarkably steady hands meticulously working: his right hand tightly grips a carving knife, its handle smooth and warm from years of use, while his left hand carefully supports the elegantly shaped pearwood body of a pipa. With steady and deliberate force from his wrist, the sharp tip of the knife glided across the delicate, luminous wood of the pipa's back. Slender wood petals curled and gently lifted, gradually revealing the intricate and lifelike outline of a peony. The sunlight precisely illuminated the pipa and his deft hands, making every carved grain visible. A striking contrast of light and shadow formed against the warm wood tone, and the entire scene was permeated with an atmosphere of profound tranquility, almost sacred, and the spirit of craftsmanship.", + "seed": 14729, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_15/index_717_seed_14729_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.156140923127532, + "vlm_logprob": -0.156140923127532, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.156140923127532, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up filmed from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elderly craftsman with white hair wearing a washed-out blue cotton work suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman carving a pearwood pipa with a carving knife to reveal a peony design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a shaft of light through dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09897814691066742 + ], + "opposite_logprob_tokens": [ + -3.973978042602539 + ], + "logprob_reward": -0.09897814691066742, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.09897814691066742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11650528013706207, + 0.03974608704447746, + 0.07943270355463028, + 0.08061391860246658, + 0.06408922374248505, + 0.058547087013721466, + 0.05124350264668465, + 0.39013394713401794, + 0.05394188314676285, + 0.06574637442827225 + ], + "credit_map": [ + 0.009582542814314365, + 0.010744535364210606, + 0.014857343398034573, + 0.01719917356967926, + 0.013284931890666485, + 0.014848061837255955, + 0.018064243718981743, + 0.01792445406317711, + 0.002835663501173258, + 0.003420697757974267, + 0.004674704745411873, + 0.006420861463993788, + 0.004955925513058901, + 0.00564231863245368, + 0.005835590418428183, + 0.005960324313491583, + 0.00918652955442667, + 0.006950113922357559, + 0.0069175646640360355, + 0.007631921209394932, + 0.016040122136473656, + 0.012138435617089272, + 0.007058122660964727, + 0.013509896583855152, + 0.00766156567260623, + 0.011642427183687687, + 0.004937079269438982, + 0.009728582575917244, + 0.0173504538834095, + 0.010238124057650566, + 0.008679254911839962, + 0.010376429185271263, + 0.007419401779770851, + 0.007272230461239815, + 0.009704628959298134, + 0.010988017544150352, + 0.006821480114012957, + 0.005465532653033733, + 0.007603581994771957, + 0.008814353495836258, + 0.0061234342865645885, + 0.004404009319841862, + 0.006757049821317196, + 0.005681597162038088, + 0.01120010670274496, + 0.00994071364402771, + 0.005776195786893368, + 0.008663984946906567, + 0.003870567074045539, + 0.004064594395458698, + 0.003942839335650206, + 0.00568494712933898, + 0.006972732022404671, + 0.009143538773059845, + 0.004404416773468256, + 0.013159865513443947, + 0.037382494658231735, + 0.049390390515327454, + 0.019483231008052826, + 0.08833952248096466, + 0.08354340493679047, + 0.03601527586579323, + 0.047977153211832047, + 0.028002440929412842, + 0.006473757326602936, + 0.00680288951843977, + 0.0052588279359042645, + 0.007594114635139704, + 0.006142737809568644, + 0.004842559341341257, + 0.008813009597361088, + 0.008013986982405186, + 0.004087225068360567, + 0.005654757842421532, + 0.0064246924594044685, + 0.008052238263189793, + 0.00889983493834734, + 0.006343604065477848, + 0.015241933986544609, + 0.011042091995477676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.09897814691066742 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up filmed from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elderly craftsman with white hair wearing a washed-out blue cotton work suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman carving a pearwood pipa with a carving knife to reveal a peony design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a shaft of light through dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233539342880249 + ], + "opposite_logprob_tokens": [ + -0.8983539342880249 + ], + "logprob_reward": -0.5233539342880249, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233539342880249, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07930342853069305, + 0.03304052725434303, + 0.06517689675092697, + 0.04607805982232094, + 0.052604299038648605, + 0.03147487714886665, + 0.1734725683927536, + 0.37752577662467957, + 0.06824587285518646, + 0.07307777553796768 + ], + "credit_map": [ + 0.005295413080602884, + 0.005259613040834665, + 0.014778624288737774, + 0.010463220067322254, + 0.01260529924184084, + 0.007118724286556244, + 0.011451849713921547, + 0.012330682016909122, + 0.002442138036713004, + 0.0028918003663420677, + 0.005259230267256498, + 0.003822767874225974, + 0.004064603708684444, + 0.00364919425919652, + 0.0059029883705079556, + 0.005007803440093994, + 0.00516405189409852, + 0.005332179833203554, + 0.0034442732576280832, + 0.004114914685487747, + 0.011080391705036163, + 0.01395853329449892, + 0.008179093711078167, + 0.013903461396694183, + 0.003329162485897541, + 0.0056960321962833405, + 0.0026542029809206724, + 0.00304202432744205, + 0.0094017768278718, + 0.01067420095205307, + 0.0031122355721890926, + 0.008168421685695648, + 0.004569227807223797, + 0.004139626864343882, + 0.005074671935290098, + 0.004919678438454866, + 0.013718909583985806, + 0.007727221120148897, + 0.0041761137545108795, + 0.00827884953469038, + 0.00342007540166378, + 0.0033794732298702, + 0.003183135762810707, + 0.0029725355561822653, + 0.004964760039001703, + 0.005309680011123419, + 0.0031122248619794846, + 0.005132994148880243, + 0.01240361388772726, + 0.01649114489555359, + 0.012058978900313377, + 0.018161702901124954, + 0.028278663754463196, + 0.029254542663693428, + 0.013711793348193169, + 0.04311211779713631, + 0.024300333112478256, + 0.05126512050628662, + 0.019352806732058525, + 0.046751994639635086, + 0.09130154550075531, + 0.05738191679120064, + 0.03733532875776291, + 0.04983670637011528, + 0.006129991263151169, + 0.008250914514064789, + 0.00690917344763875, + 0.00903051346540451, + 0.007397844921797514, + 0.006583489011973143, + 0.01617923006415367, + 0.007764714304357767, + 0.0060102446004748344, + 0.005938694812357426, + 0.007264023181051016, + 0.00972410012036562, + 0.010241510346531868, + 0.00683408509939909, + 0.011025971733033657, + 0.01603914424777031 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.5233539342880249 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up filmed from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elderly craftsman with white hair wearing a washed-out blue cotton work suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman carving a pearwood pipa with a carving knife to reveal a peony design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a shaft of light through dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048679985105991364 + ], + "opposite_logprob_tokens": [ + -3.048680067062378 + ], + "logprob_reward": -0.048679985105991364, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048679985105991364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19589056074619293, + 0.08260367065668106, + 0.05417132377624512, + 0.04231589287519455, + 0.028198154643177986, + 0.07678747177124023, + 0.13395202159881592, + 0.3108271062374115, + 0.05418771132826805, + 0.021066104993224144 + ], + "credit_map": [ + 0.010609304532408714, + 0.01339018065482378, + 0.018736183643341064, + 0.016091737896203995, + 0.0420958548784256, + 0.03921545296907425, + 0.025679519400000572, + 0.030072323977947235, + 0.004378533456474543, + 0.005774969235062599, + 0.005858729127794504, + 0.0072906045243144035, + 0.01955445110797882, + 0.011458039283752441, + 0.015480798669159412, + 0.012807545252144337, + 0.00337389437481761, + 0.005095196422189474, + 0.004561159759759903, + 0.0040407609194517136, + 0.009911494329571724, + 0.010913421399891376, + 0.005077958572655916, + 0.01119743287563324, + 0.0037065548822283745, + 0.0032134701032191515, + 0.0028596604242920876, + 0.006144913379102945, + 0.005916276015341282, + 0.006251413840800524, + 0.003968018107116222, + 0.010255587287247181, + 0.0030992517713457346, + 0.003226405242457986, + 0.0034883844200521708, + 0.00441635400056839, + 0.0035019528586417437, + 0.0028619435615837574, + 0.0037703022826462984, + 0.0038335612043738365, + 0.006815118715167046, + 0.006268261931836605, + 0.0038246349431574345, + 0.0048554325476288795, + 0.020122738555073738, + 0.012614674866199493, + 0.007797506637871265, + 0.014489104971289635, + 0.010230688378214836, + 0.009922509081661701, + 0.008022500202059746, + 0.011133937165141106, + 0.021978696808218956, + 0.024042174220085144, + 0.009915468282997608, + 0.03870605304837227, + 0.025321630761027336, + 0.05527864396572113, + 0.018811840564012527, + 0.02197515219449997, + 0.09160839021205902, + 0.04051699489355087, + 0.019319625571370125, + 0.03799482062458992, + 0.005628357641398907, + 0.006110132206231356, + 0.004432884510606527, + 0.005996169056743383, + 0.006220678333193064, + 0.004515370354056358, + 0.01421984564512968, + 0.007064274046570063, + 0.0017068689921870828, + 0.002053207019343972, + 0.0016178827499970794, + 0.002507823519408703, + 0.0039277225732803345, + 0.0026702850591391325, + 0.0022669192403554916, + 0.004315395839512348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.048679985105991364 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up filmed from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elderly craftsman with white hair wearing a washed-out blue cotton work suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman carving a pearwood pipa with a carving knife to reveal a peony design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a shaft of light through dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08911176025867462 + ], + "opposite_logprob_tokens": [ + -2.464111804962158 + ], + "logprob_reward": -0.08911176025867462, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08911176025867462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11941643059253693, + 0.07749916613101959, + 0.04276718199253082, + 0.025222301483154297, + 0.13887344300746918, + 0.11394790560007095, + 0.12136068195104599, + 0.2945789396762848, + 0.038667574524879456, + 0.02766636200249195 + ], + "credit_map": [ + 0.009901120327413082, + 0.013435371220111847, + 0.01531232800334692, + 0.01722688041627407, + 0.0188838429749012, + 0.011752019636332989, + 0.014773297123610973, + 0.018131569027900696, + 0.0052818418480455875, + 0.0068246969021856785, + 0.007979752495884895, + 0.009761612862348557, + 0.013046540319919586, + 0.00933829601854086, + 0.011546788737177849, + 0.013719632290303707, + 0.0030900866258889437, + 0.004046028945595026, + 0.005719850305467844, + 0.003882426070049405, + 0.007068281527608633, + 0.0060456013306975365, + 0.006045270711183548, + 0.00686963927000761, + 0.002242963993921876, + 0.002910513198003173, + 0.0017116422532126307, + 0.0031007060315459967, + 0.003583368845283985, + 0.004017808008939028, + 0.0029167996253818274, + 0.004738499876111746, + 0.013768289238214493, + 0.0154269365593791, + 0.019207889214158058, + 0.018807895481586456, + 0.02708066999912262, + 0.013898833654820919, + 0.013656826689839363, + 0.01702609471976757, + 0.013653288595378399, + 0.014241913333535194, + 0.014863558113574982, + 0.015034524723887444, + 0.015436784364283085, + 0.012133966200053692, + 0.012482799589633942, + 0.01610107719898224, + 0.008829670958220959, + 0.012023171409964561, + 0.007201716769486666, + 0.012459837831556797, + 0.02267821691930294, + 0.02040155790746212, + 0.00922917015850544, + 0.028537342324852943, + 0.015730571001768112, + 0.03225993737578392, + 0.01569373346865177, + 0.055852584540843964, + 0.06198962777853012, + 0.06261198967695236, + 0.025865579023957253, + 0.02457490935921669, + 0.003907680045813322, + 0.006833257619291544, + 0.00392802944406867, + 0.004771929234266281, + 0.004636256955564022, + 0.004339828155934811, + 0.005644928198307753, + 0.004605667199939489, + 0.0021478531416505575, + 0.0028105671517550945, + 0.0023266479838639498, + 0.004036074969917536, + 0.004313230514526367, + 0.0033461209386587143, + 0.0047022318467497826, + 0.003983636386692524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 6 + ], + "selected_reward": -0.08911176025867462 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up filmed from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elderly craftsman with white hair wearing a washed-out blue cotton work suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman carving a pearwood pipa with a carving knife to reveal a peony design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a shaft of light through dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058078907430172 + ], + "opposite_logprob_tokens": [ + -3.895580768585205 + ], + "logprob_reward": -0.02058078907430172, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058078907430172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07087713479995728, + 0.08859627693891525, + 0.5220004916191101, + 0.08060622960329056, + 0.025504568591713905, + 0.033999260514974594, + 0.019765805453062057, + 0.1196349486708641, + 0.018545301631093025, + 0.020469995215535164 + ], + "credit_map": [ + 0.004342568106949329, + 0.004869351629167795, + 0.00963145587593317, + 0.011303898878395557, + 0.011653226800262928, + 0.007484333124011755, + 0.010619127191603184, + 0.010973176918923855, + 0.00677870586514473, + 0.008162613958120346, + 0.011481650173664093, + 0.012520994059741497, + 0.011800015345215797, + 0.009913056157529354, + 0.013948549516499043, + 0.013990691863000393, + 0.07508338242769241, + 0.0660019963979721, + 0.058157727122306824, + 0.0597417950630188, + 0.07197669893503189, + 0.05979350209236145, + 0.04965905845165253, + 0.0815863087773323, + 0.007644161116331816, + 0.006697835400700569, + 0.007026373408734798, + 0.009774166159331799, + 0.016321923583745956, + 0.010018693283200264, + 0.007966462522745132, + 0.015156617388129234, + 0.0032567318994551897, + 0.003171998541802168, + 0.003676309483125806, + 0.0039615207351744175, + 0.0029044158291071653, + 0.0025322325527668, + 0.002586318412795663, + 0.0034150411374866962, + 0.0033317706547677517, + 0.003402635920792818, + 0.003338797250762582, + 0.0030209056567400694, + 0.0070737553760409355, + 0.004216830711811781, + 0.004215206950902939, + 0.005399357061833143, + 0.0015117338625714183, + 0.0012653806479647756, + 0.0016002585180103779, + 0.0017309548566117883, + 0.003275227965787053, + 0.003278750693425536, + 0.0015882214065641165, + 0.005515275523066521, + 0.010955539532005787, + 0.01616116240620613, + 0.006161293014883995, + 0.023707076907157898, + 0.026608046144247055, + 0.012748094275593758, + 0.013065303675830364, + 0.010228433646261692, + 0.0014677451690658927, + 0.001593404565937817, + 0.00202700006775558, + 0.0035223320592194796, + 0.0023701477330178022, + 0.0020022604148834944, + 0.0037653546314686537, + 0.0017970559420064092, + 0.0018378791864961386, + 0.0014492807677015662, + 0.0033634803257882595, + 0.004834909923374653, + 0.0023432618472725153, + 0.001585212186910212, + 0.0032419029157608747, + 0.0018140686443075538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 1 + ], + "selected_reward": -0.02058078907430172 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot, shot from a level angle, slowly and steadily inward, draws the viewer's attention to a white-haired, elderly craftsman. He sits in a woodworking workshop filled with the mingled aroma of sawdust and varnish. His face is deeply wrinkled, his eyes, through a pair of reading glasses, fixed with intense concentration on his work. A beam of warm afternoon sunlight slants in through the high windows, casting a visible shaft of light in the air, illuminating the dancing dust particles. He wears a washed-out blue cotton work suit, leaning slightly forward, his calloused yet remarkably steady hands meticulously working: his right hand tightly grips a carving knife, its handle smooth and warm from years of use, while his left hand carefully supports the elegantly shaped pearwood body of a pipa. With steady and deliberate force from his wrist, the sharp tip of the knife glided across the delicate, luminous wood of the pipa's back. Slender wood petals curled and gently lifted, gradually revealing the intricate and lifelike outline of a peony. The sunlight precisely illuminated the pipa and his deft hands, making every carved grain visible. A striking contrast of light and shadow formed against the warm wood tone, and the entire scene was permeated with an atmosphere of profound tranquility, almost sacred, and the spirit of craftsmanship.", + "seed": 14730, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_15/index_717_seed_14730_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.14787825644016267, + "vlm_logprob": -0.14787825644016267, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.14787825644016267, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up filmed from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elderly craftsman with white hair wearing a washed-out blue cotton work suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman carving a pearwood pipa with a carving knife to reveal a peony design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a shaft of light through dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1283256560564041 + ], + "opposite_logprob_tokens": [ + -4.253325462341309 + ], + "logprob_reward": -0.1283256560564041, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.1283256560564041, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24871747195720673, + 0.10442018508911133, + 0.11941257864236832, + 0.04260020703077316, + 0.050558943301439285, + 0.12213115394115448, + 0.057702019810676575, + 0.08116358518600464, + 0.06758085638284683, + 0.10571302473545074 + ], + "credit_map": [ + 0.05590546876192093, + 0.02939874865114689, + 0.027127079665660858, + 0.02275831624865532, + 0.03656359016895294, + 0.02480398304760456, + 0.0211538877338171, + 0.03100639209151268, + 0.019178584218025208, + 0.009904708713293076, + 0.0163734033703804, + 0.009560946375131607, + 0.014874790795147419, + 0.01714479550719261, + 0.005839475430548191, + 0.011543474160134792, + 0.0242935661226511, + 0.011966614052653313, + 0.013150446116924286, + 0.01964394748210907, + 0.010586163960397243, + 0.015956906601786613, + 0.00995531864464283, + 0.01385961938649416, + 0.01136008556932211, + 0.004905212204903364, + 0.0033874015789479017, + 0.004626620560884476, + 0.005395718850195408, + 0.004898193292319775, + 0.004316431935876608, + 0.003710543503984809, + 0.011122990399599075, + 0.005818348843604326, + 0.0052087027579545975, + 0.006380648352205753, + 0.008735781535506248, + 0.004548744764178991, + 0.003980874549597502, + 0.0047628553584218025, + 0.02100345306098461, + 0.010388584807515144, + 0.015435595065355301, + 0.00867282971739769, + 0.01568995602428913, + 0.02530103549361229, + 0.009154523722827435, + 0.016485176980495453, + 0.010432573035359383, + 0.007724781986325979, + 0.009536970406770706, + 0.00602984381839633, + 0.007683178409934044, + 0.00717149255797267, + 0.004278245382010937, + 0.004844933748245239, + 0.012202490121126175, + 0.012111860327422619, + 0.010237196460366249, + 0.009303762577474117, + 0.012995426543056965, + 0.008299903944134712, + 0.009667395614087582, + 0.006345545873045921, + 0.01112397201359272, + 0.006490200292319059, + 0.008555993437767029, + 0.011916182935237885, + 0.006885794457048178, + 0.00721384771168232, + 0.006691070273518562, + 0.008703796193003654, + 0.015032351016998291, + 0.01148141361773014, + 0.011225352995097637, + 0.006843901704996824, + 0.00954576674848795, + 0.019134489819407463, + 0.007202409207820892, + 0.02524734102189541 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.1283256560564041 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up filmed from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elderly craftsman with white hair wearing a washed-out blue cotton work suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman carving a pearwood pipa with a carving knife to reveal a peony design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a shaft of light through dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289213716983795 + ], + "opposite_logprob_tokens": [ + -1.0539213418960571 + ], + "logprob_reward": -0.4289213716983795, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289213716983795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1704181283712387, + 0.06402794271707535, + 0.06349718570709229, + 0.09256669133901596, + 0.07473079860210419, + 0.17045849561691284, + 0.14657491445541382, + 0.041277751326560974, + 0.06519574671983719, + 0.11125237494707108 + ], + "credit_map": [ + 0.036342840641736984, + 0.020825713872909546, + 0.01701126992702484, + 0.02051451988518238, + 0.023928316310048103, + 0.0181519016623497, + 0.013376876711845398, + 0.020266691222786903, + 0.010196831077337265, + 0.006062905304133892, + 0.006778242066502571, + 0.0058404081501066685, + 0.008126200176775455, + 0.013825930655002594, + 0.00519565399736166, + 0.008001770824193954, + 0.014994778670370579, + 0.006996236275881529, + 0.0064299944788217545, + 0.008297414518892765, + 0.008848067373037338, + 0.007525849621742964, + 0.004968536086380482, + 0.005436309147626162, + 0.023383472114801407, + 0.012566219083964825, + 0.009613069705665112, + 0.008093257434666157, + 0.014816470444202423, + 0.008737527765333652, + 0.007426643278449774, + 0.007930031046271324, + 0.01583118736743927, + 0.009766541421413422, + 0.010261092334985733, + 0.010596535168588161, + 0.0071877422742545605, + 0.008603859692811966, + 0.005029905121773481, + 0.007453943602740765, + 0.025736931711435318, + 0.015445309691131115, + 0.016181422397494316, + 0.014298772439360619, + 0.02095053344964981, + 0.031266242265701294, + 0.015429443679749966, + 0.03114982508122921, + 0.02336488664150238, + 0.01684679463505745, + 0.016990672796964645, + 0.019149627536535263, + 0.034734904766082764, + 0.011413159780204296, + 0.013936661183834076, + 0.010138201527297497, + 0.00987582840025425, + 0.005118367727845907, + 0.004830437712371349, + 0.006632398813962936, + 0.005075334571301937, + 0.003526658983901143, + 0.0033887531608343124, + 0.002829969394952059, + 0.012622648850083351, + 0.007560600060969591, + 0.007545946631580591, + 0.008178147487342358, + 0.009207138791680336, + 0.00707577308639884, + 0.0059573291800916195, + 0.0070481630973517895, + 0.014700852334499359, + 0.010715183801949024, + 0.010236411355435848, + 0.015182647854089737, + 0.015389936044812202, + 0.016135619953274727, + 0.010114847682416439, + 0.018776873126626015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.4289213716983795 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up filmed from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elderly craftsman with white hair wearing a washed-out blue cotton work suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman carving a pearwood pipa with a carving knife to reveal a peony design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a shaft of light through dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054974041879177094 + ], + "opposite_logprob_tokens": [ + -2.92997407913208 + ], + "logprob_reward": -0.054974041879177094, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054974041879177094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4132063686847687, + 0.08850990235805511, + 0.07491755485534668, + 0.054268352687358856, + 0.08602434396743774, + 0.06727736443281174, + 0.050740260630846024, + 0.09970997273921967, + 0.02197698876261711, + 0.043368831276893616 + ], + "credit_map": [ + 0.06533496826887131, + 0.04293108731508255, + 0.04675017669796944, + 0.03517976403236389, + 0.07694050669670105, + 0.06767869740724564, + 0.03218889981508255, + 0.046202290803194046, + 0.01591489277780056, + 0.009470168501138687, + 0.0123581038787961, + 0.0072874752804636955, + 0.00974394753575325, + 0.01598077453672886, + 0.007735796272754669, + 0.010018743574619293, + 0.0138267632573843, + 0.007508968934416771, + 0.0046693794429302216, + 0.006913850083947182, + 0.012621111236512661, + 0.012858888134360313, + 0.00695993285626173, + 0.0095586646348238, + 0.009194794110953808, + 0.005733204074203968, + 0.004235709551721811, + 0.004588826093822718, + 0.00854304526001215, + 0.007937626913189888, + 0.006915335543453693, + 0.007119813002645969, + 0.020791413262486458, + 0.011477806605398655, + 0.009266636334359646, + 0.01469530537724495, + 0.011027339845895767, + 0.007193680386990309, + 0.005366131197661161, + 0.0062060304917395115, + 0.010468762367963791, + 0.0073991697281599045, + 0.007268476765602827, + 0.005194355268031359, + 0.009078755974769592, + 0.009003664366900921, + 0.007552207913249731, + 0.011311977170407772, + 0.008298512548208237, + 0.005720554850995541, + 0.004050229676067829, + 0.005628161132335663, + 0.006646165624260902, + 0.008057333528995514, + 0.0053948769345879555, + 0.006944428198039532, + 0.010376307182013988, + 0.011693179607391357, + 0.00804470106959343, + 0.00982311088591814, + 0.014763767831027508, + 0.016693128272891045, + 0.017522970214486122, + 0.010792804881930351, + 0.0032385096419602633, + 0.0023953544441610575, + 0.0025311147328466177, + 0.003912108484655619, + 0.003031977219507098, + 0.002155906753614545, + 0.002371836919337511, + 0.0023401801008731127, + 0.005194761790335178, + 0.004859209526330233, + 0.0047293053939938545, + 0.008120001293718815, + 0.005383125972002745, + 0.004226546734571457, + 0.002895616926252842, + 0.007960264571011066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.054974041879177094 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up filmed from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elderly craftsman with white hair wearing a washed-out blue cotton work suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman carving a pearwood pipa with a carving knife to reveal a peony design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a shaft of light through dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08908232301473618 + ], + "opposite_logprob_tokens": [ + -2.4640822410583496 + ], + "logprob_reward": -0.08908232301473618, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08908232301473618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27016881108283997, + 0.09635443240404129, + 0.06255355477333069, + 0.06517685204744339, + 0.08705826103687286, + 0.09719513356685638, + 0.07062464207410812, + 0.04898540675640106, + 0.08823224902153015, + 0.11365069448947906 + ], + "credit_map": [ + 0.03833004832267761, + 0.02676655352115631, + 0.02844998799264431, + 0.03196800500154495, + 0.047757282853126526, + 0.0385284349322319, + 0.025895917788147926, + 0.032472554594278336, + 0.018178753554821014, + 0.01145175937563181, + 0.010872268117964268, + 0.007284782361239195, + 0.013930493034422398, + 0.016212211921811104, + 0.009466945193707943, + 0.008957215584814548, + 0.010588192380964756, + 0.006110137328505516, + 0.006080908235162497, + 0.0064354888163506985, + 0.008095062337815762, + 0.011321617290377617, + 0.004839497618377209, + 0.009082645177841187, + 0.014985756948590279, + 0.007438588421791792, + 0.006069009192287922, + 0.005239320918917656, + 0.01572258025407791, + 0.006245860364288092, + 0.004626988898962736, + 0.004848734941333532, + 0.01737443171441555, + 0.009415351785719395, + 0.010319188237190247, + 0.019320249557495117, + 0.008962690830230713, + 0.006772890221327543, + 0.006967471446841955, + 0.00792598258703947, + 0.019468458369374275, + 0.010531675070524216, + 0.01663210429251194, + 0.007725766859948635, + 0.013296684250235558, + 0.01334208995103836, + 0.007209485862404108, + 0.008988860994577408, + 0.01355123519897461, + 0.008744417689740658, + 0.005882781930267811, + 0.007371004205197096, + 0.013767257332801819, + 0.00866052508354187, + 0.005944452714174986, + 0.00670296186581254, + 0.006651373114436865, + 0.005090048536658287, + 0.0058101448230445385, + 0.008886557072401047, + 0.008753928355872631, + 0.006453336216509342, + 0.003554399125277996, + 0.0037856127601116896, + 0.01367869134992361, + 0.00786262471228838, + 0.011371917091310024, + 0.013255869038403034, + 0.00979682058095932, + 0.008357880637049675, + 0.009716453030705452, + 0.014191982336342335, + 0.01548078004270792, + 0.013087064027786255, + 0.013050353154540062, + 0.01763269305229187, + 0.014005989767611027, + 0.019633064046502113, + 0.006672868505120277, + 0.014087867923080921 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.08908232301473618 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up filmed from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elderly craftsman with white hair wearing a washed-out blue cotton work suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman carving a pearwood pipa with a carving knife to reveal a peony design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a shaft of light through dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038087889552116394 + ], + "opposite_logprob_tokens": [ + -3.288087844848633 + ], + "logprob_reward": -0.038087889552116394, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038087889552116394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09296318888664246, + 0.025355231016874313, + 0.041215233504772186, + 0.04878586530685425, + 0.06466839462518692, + 0.05662133917212486, + 0.11061858385801315, + 0.2981909513473511, + 0.18138499557971954, + 0.08019629865884781 + ], + "credit_map": [ + 0.016697395592927933, + 0.01032987143844366, + 0.016205182299017906, + 0.010867992416024208, + 0.01328342966735363, + 0.008968315087258816, + 0.008119442500174046, + 0.00849156454205513, + 0.0035136525984853506, + 0.002439686097204685, + 0.004126358777284622, + 0.002950040390715003, + 0.0037336740642786026, + 0.0042208218947052956, + 0.0017792151775211096, + 0.0025917820166796446, + 0.008663215674459934, + 0.004380951169878244, + 0.0021515730768442154, + 0.002691347384825349, + 0.010570509359240532, + 0.006479218136519194, + 0.0032486100681126118, + 0.0030298042111098766, + 0.011394808068871498, + 0.005692323204129934, + 0.004199621733278036, + 0.00444659823551774, + 0.01124543882906437, + 0.004619779996573925, + 0.0034893385600298643, + 0.003697955049574375, + 0.013575268909335136, + 0.007988003082573414, + 0.010228938423097134, + 0.01067151129245758, + 0.007861591875553131, + 0.005177087150514126, + 0.004798243287950754, + 0.004367752932012081, + 0.008893286809325218, + 0.00505739264190197, + 0.006325413938611746, + 0.0036549894139170647, + 0.010530602186918259, + 0.011284243315458298, + 0.0038982348050922155, + 0.00697717908769846, + 0.021411152556538582, + 0.013491218909621239, + 0.015996010974049568, + 0.009482369758188725, + 0.014343010261654854, + 0.015886083245277405, + 0.008143159560859203, + 0.01186557486653328, + 0.05177025496959686, + 0.030917633324861526, + 0.03902445733547211, + 0.04639473557472229, + 0.05111038312315941, + 0.03154695779085159, + 0.0270093884319067, + 0.020417125895619392, + 0.024164360016584396, + 0.021119916811585426, + 0.01959829404950142, + 0.03134598582983017, + 0.027966400608420372, + 0.02540794014930725, + 0.014882235787808895, + 0.016899853944778442, + 0.010525517165660858, + 0.009809553623199463, + 0.012503491714596748, + 0.011695681139826775, + 0.009811757132411003, + 0.011009696871042252, + 0.007762725464999676, + 0.007077871356159449 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.038087889552116394 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot, shot from a level angle, slowly and steadily inward, draws the viewer's attention to a white-haired, elderly craftsman. He sits in a woodworking workshop filled with the mingled aroma of sawdust and varnish. His face is deeply wrinkled, his eyes, through a pair of reading glasses, fixed with intense concentration on his work. A beam of warm afternoon sunlight slants in through the high windows, casting a visible shaft of light in the air, illuminating the dancing dust particles. He wears a washed-out blue cotton work suit, leaning slightly forward, his calloused yet remarkably steady hands meticulously working: his right hand tightly grips a carving knife, its handle smooth and warm from years of use, while his left hand carefully supports the elegantly shaped pearwood body of a pipa. With steady and deliberate force from his wrist, the sharp tip of the knife glided across the delicate, luminous wood of the pipa's back. Slender wood petals curled and gently lifted, gradually revealing the intricate and lifelike outline of a peony. The sunlight precisely illuminated the pipa and his deft hands, making every carved grain visible. A striking contrast of light and shadow formed against the warm wood tone, and the entire scene was permeated with an atmosphere of profound tranquility, almost sacred, and the spirit of craftsmanship.", + "seed": 14731, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_15/index_717_seed_14731_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1024230808019638, + "vlm_logprob": -0.1024230808019638, + "vlm_yes_no_margin": 3.049999952316284, + "vlm_token_credit": -0.1024230808019638, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up filmed from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elderly craftsman with white hair wearing a washed-out blue cotton work suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman carving a pearwood pipa with a carving knife to reveal a peony design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a shaft of light through dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17549137771129608 + ], + "opposite_logprob_tokens": [ + -4.3004913330078125 + ], + "logprob_reward": -0.17549137771129608, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.17549137771129608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17527857422828674, + 0.12601818144321442, + 0.07906530052423477, + 0.060867294669151306, + 0.06665989011526108, + 0.056847453117370605, + 0.1383652687072754, + 0.13676412403583527, + 0.09055555611848831, + 0.0695783793926239 + ], + "credit_map": [ + 0.01870085299015045, + 0.03936786949634552, + 0.025027956813573837, + 0.02150113508105278, + 0.022134652361273766, + 0.025967076420783997, + 0.01108785904943943, + 0.01149115338921547, + 0.014567387290298939, + 0.03314988687634468, + 0.015347122214734554, + 0.010518942028284073, + 0.019510867074131966, + 0.013941878452897072, + 0.010230187326669693, + 0.008751904591917992, + 0.006985767744481564, + 0.011355533264577389, + 0.009002060629427433, + 0.0041274926625192165, + 0.009205920621752739, + 0.014329776167869568, + 0.010036218911409378, + 0.014022531919181347, + 0.009986520744860172, + 0.008565138094127178, + 0.00813589058816433, + 0.005890975706279278, + 0.006550640799105167, + 0.01181748416274786, + 0.004967756103724241, + 0.004952885676175356, + 0.008139844983816147, + 0.010398521088063717, + 0.007842748425900936, + 0.00486984895542264, + 0.007172140292823315, + 0.01327878050506115, + 0.005794945638626814, + 0.00916306208819151, + 0.007814796641469002, + 0.0076498002745211124, + 0.005620914977043867, + 0.004631831776350737, + 0.006058768834918737, + 0.009597511030733585, + 0.010608464479446411, + 0.0048653604462742805, + 0.014203421771526337, + 0.029329167678952217, + 0.018004341050982475, + 0.014733673073351383, + 0.018291277810931206, + 0.01698841154575348, + 0.01169812772423029, + 0.01511684525758028, + 0.010912343859672546, + 0.0174312312155962, + 0.030057145282626152, + 0.013889512047171593, + 0.014630102552473545, + 0.012507621198892593, + 0.02716316655278206, + 0.01017299946397543, + 0.008440956473350525, + 0.009384623728692532, + 0.01442812755703926, + 0.010945837944746017, + 0.010409876704216003, + 0.008544610813260078, + 0.017222842201590538, + 0.011178676038980484, + 0.008473551832139492, + 0.011497734114527702, + 0.006627874448895454, + 0.008418853394687176, + 0.010127308778464794, + 0.008445530198514462, + 0.006697529926896095, + 0.009289990179240704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.17549137771129608 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up filmed from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elderly craftsman with white hair wearing a washed-out blue cotton work suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman carving a pearwood pipa with a carving knife to reveal a peony design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a shaft of light through dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1799401491880417 + ], + "opposite_logprob_tokens": [ + -1.804940104484558 + ], + "logprob_reward": -0.1799401491880417, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1799401491880417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09222300350666046, + 0.059810180217027664, + 0.06558737903833389, + 0.0551188625395298, + 0.07493274658918381, + 0.06671169400215149, + 0.2473924607038498, + 0.09520016610622406, + 0.10934391617774963, + 0.13367967307567596 + ], + "credit_map": [ + 0.011596357449889183, + 0.019256575033068657, + 0.015652257949113846, + 0.006336126010864973, + 0.010415008291602135, + 0.014321396127343178, + 0.005447349976748228, + 0.009197917766869068, + 0.008636197075247765, + 0.009555855765938759, + 0.006837930530309677, + 0.006674614734947681, + 0.007907507941126823, + 0.008507659658789635, + 0.006807644385844469, + 0.00488276407122612, + 0.009766648523509502, + 0.010765980929136276, + 0.009887272492051125, + 0.005339723080396652, + 0.007857770659029484, + 0.011815345846116543, + 0.0052416520193219185, + 0.004912978038191795, + 0.011163162998855114, + 0.007985767908394337, + 0.009037211537361145, + 0.004108465276658535, + 0.0054208505898714066, + 0.005956966895610094, + 0.005165407434105873, + 0.006281024776399136, + 0.00893451552838087, + 0.00905047357082367, + 0.00935698114335537, + 0.008972784504294395, + 0.0077727073803544044, + 0.01385189313441515, + 0.009715006686747074, + 0.0072783720679581165, + 0.008976224809885025, + 0.008035951294004917, + 0.011234256438910961, + 0.009706581011414528, + 0.005750901997089386, + 0.008485551923513412, + 0.010208867490291595, + 0.004313352517783642, + 0.02819780632853508, + 0.05339605733752251, + 0.030313968658447266, + 0.026502735912799835, + 0.044026341289281845, + 0.02774924598634243, + 0.017114825546741486, + 0.020091433078050613, + 0.01373105775564909, + 0.017759723588824272, + 0.018105367198586464, + 0.010362702421844006, + 0.009261325933039188, + 0.008203999139368534, + 0.01121631357818842, + 0.006559666711837053, + 0.015704039484262466, + 0.02310444787144661, + 0.01377306878566742, + 0.012389253824949265, + 0.01127120666205883, + 0.0088205486536026, + 0.013526245951652527, + 0.010755087248980999, + 0.018002158030867577, + 0.025131449103355408, + 0.010111776180565357, + 0.029650036245584488, + 0.013804396614432335, + 0.013073339127004147, + 0.014062282629311085, + 0.009844214655458927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.1799401491880417 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up filmed from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elderly craftsman with white hair wearing a washed-out blue cotton work suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman carving a pearwood pipa with a carving knife to reveal a peony design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a shaft of light through dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018207203596830368 + ], + "opposite_logprob_tokens": [ + -4.01820707321167 + ], + "logprob_reward": -0.018207203596830368, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018207203596830368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42719361186027527, + 0.09613873809576035, + 0.08173863589763641, + 0.029610207304358482, + 0.05446649715304375, + 0.06862741708755493, + 0.08805327862501144, + 0.06284234672784805, + 0.04150298237800598, + 0.0498262420296669 + ], + "credit_map": [ + 0.04910700395703316, + 0.12273626029491425, + 0.07625114172697067, + 0.03329843655228615, + 0.057238634675741196, + 0.04403986781835556, + 0.020309116691350937, + 0.024213146418333054, + 0.010247793979942799, + 0.02296983264386654, + 0.014000154100358486, + 0.010407115332782269, + 0.013136387802660465, + 0.011645765043795109, + 0.007168642710894346, + 0.006563038565218449, + 0.009532651863992214, + 0.008694523945450783, + 0.012668445706367493, + 0.0045102303847670555, + 0.010017303749918938, + 0.02229061722755432, + 0.006311883218586445, + 0.007712980732321739, + 0.005330787971615791, + 0.0038194817025214434, + 0.004504989366978407, + 0.0030250349082052708, + 0.0028143664821982384, + 0.004619731567800045, + 0.0025464375503361225, + 0.002949376357719302, + 0.010465788654983044, + 0.011757474392652512, + 0.007125504780560732, + 0.004764865152537823, + 0.006692023016512394, + 0.006872230675071478, + 0.0028504293877631426, + 0.0039381845854222775, + 0.014033223502337933, + 0.011096090078353882, + 0.01009112223982811, + 0.007552044000476599, + 0.006181421689689159, + 0.008704778738319874, + 0.006282186135649681, + 0.00468654977157712, + 0.009939136914908886, + 0.018519632518291473, + 0.010037658736109734, + 0.009857382625341415, + 0.015606836415827274, + 0.009995480068027973, + 0.006277070846408606, + 0.00782008282840252, + 0.008180673234164715, + 0.009522661566734314, + 0.010464939288794994, + 0.007857395336031914, + 0.006787474267184734, + 0.007330704014748335, + 0.007866722531616688, + 0.00483177974820137, + 0.0057601239532232285, + 0.009158394299447536, + 0.005558236967772245, + 0.0053452737629413605, + 0.0036770363803952932, + 0.0033192867413163185, + 0.004989147186279297, + 0.0036954821553081274, + 0.006255306303501129, + 0.010334464721381664, + 0.0036840522661805153, + 0.0066961864940822124, + 0.005248615518212318, + 0.0038921840023249388, + 0.009839769452810287, + 0.003875661175698042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.018207203596830368 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up filmed from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elderly craftsman with white hair wearing a washed-out blue cotton work suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman carving a pearwood pipa with a carving knife to reveal a peony design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a shaft of light through dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10038407891988754 + ], + "opposite_logprob_tokens": [ + -2.350383996963501 + ], + "logprob_reward": -0.10038407891988754, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10038407891988754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.271218866109848, + 0.06301452964544296, + 0.03083154745399952, + 0.06003522500395775, + 0.14653150737285614, + 0.0741671621799469, + 0.11384078860282898, + 0.09603136777877808, + 0.07594706863164902, + 0.0683818981051445 + ], + "credit_map": [ + 0.022055653855204582, + 0.04014262184500694, + 0.04428187012672424, + 0.030435698106884956, + 0.02338656596839428, + 0.03128291293978691, + 0.037066008895635605, + 0.04256753250956535, + 0.0061419387347996235, + 0.00835639052093029, + 0.00792402122169733, + 0.00945198256522417, + 0.006108837202191353, + 0.007281414233148098, + 0.007331373635679483, + 0.010418571531772614, + 0.0034482735209167004, + 0.005793158430606127, + 0.004902203567326069, + 0.0034980629570782185, + 0.0038785049691796303, + 0.003666982986032963, + 0.0026683895848691463, + 0.00297597236931324, + 0.01109431590884924, + 0.00782371498644352, + 0.012349450029432774, + 0.004062711726874113, + 0.005982589442282915, + 0.008418981917202473, + 0.003623291617259383, + 0.006680171005427837, + 0.01227523759007454, + 0.013223029673099518, + 0.017748622223734856, + 0.01780043914914131, + 0.016238048672676086, + 0.029537765309214592, + 0.021545685827732086, + 0.01816268265247345, + 0.006877785082906485, + 0.007835145108401775, + 0.00696071982383728, + 0.00670330785214901, + 0.008688118308782578, + 0.015719914808869362, + 0.013603204861283302, + 0.007778964936733246, + 0.01385423168540001, + 0.020161038264632225, + 0.013838785700500011, + 0.012626572512090206, + 0.016891034319996834, + 0.012766542844474316, + 0.009466411545872688, + 0.014236174523830414, + 0.011157902888953686, + 0.014264745637774467, + 0.01946585811674595, + 0.010504498146474361, + 0.012291870079934597, + 0.012459008023142815, + 0.008808453567326069, + 0.007079034112393856, + 0.00755305727943778, + 0.010407189838588238, + 0.02150481566786766, + 0.00857082474976778, + 0.005113636609166861, + 0.005051620304584503, + 0.0072615728713572025, + 0.010484348051249981, + 0.005972289014607668, + 0.006933171767741442, + 0.006303239613771439, + 0.02010948397219181, + 0.005248126108199358, + 0.007393276318907738, + 0.0100800646468997, + 0.00634224945679307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.10038407891988754 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up filmed from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elderly craftsman with white hair wearing a washed-out blue cotton work suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman carving a pearwood pipa with a carving knife to reveal a peony design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight visible casting a shaft of light through dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809259459376335 + ], + "opposite_logprob_tokens": [ + -3.288092613220215 + ], + "logprob_reward": -0.03809259459376335, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809259459376335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1885358840227127, + 0.030953122302889824, + 0.01959245093166828, + 0.04206187278032303, + 0.03813154995441437, + 0.03317376971244812, + 0.04870278388261795, + 0.4421856999397278, + 0.03968515992164612, + 0.11697778850793839 + ], + "credit_map": [ + 0.026720911264419556, + 0.02300531603395939, + 0.028819069266319275, + 0.012182642705738544, + 0.02329069748520851, + 0.04029263183474541, + 0.015744131058454514, + 0.018480440601706505, + 0.004564028698951006, + 0.006285176612436771, + 0.00417330302298069, + 0.003143104026094079, + 0.004435138776898384, + 0.0037454324774444103, + 0.00202124728821218, + 0.002585687441751361, + 0.002382157603278756, + 0.0030795056372880936, + 0.0024631391279399395, + 0.002195945242419839, + 0.0024820291437208652, + 0.003654658794403076, + 0.0016895695589482784, + 0.0016454412834718823, + 0.01093947235494852, + 0.005448160693049431, + 0.006971473339945078, + 0.0033647159580141306, + 0.004307277500629425, + 0.004177021328359842, + 0.0031032233964651823, + 0.003750517964363098, + 0.003558387281373143, + 0.0045283157378435135, + 0.005207960493862629, + 0.004349900409579277, + 0.003938398789614439, + 0.006374815478920937, + 0.005855611525475979, + 0.004318150691688061, + 0.002899361541494727, + 0.004879766143858433, + 0.0037308363243937492, + 0.004117815755307674, + 0.0032358220778405666, + 0.004493018612265587, + 0.006313567981123924, + 0.003503573127090931, + 0.005352823995053768, + 0.007277014199644327, + 0.0060447514988482, + 0.005788476672023535, + 0.006872061174362898, + 0.0073114181868731976, + 0.004819643218070269, + 0.00523658562451601, + 0.03252783790230751, + 0.05079815909266472, + 0.06760847568511963, + 0.05429434776306152, + 0.05422288924455643, + 0.05696140229701996, + 0.08621919900178909, + 0.03955334797501564, + 0.004538052715361118, + 0.005782122258096933, + 0.007045393344014883, + 0.004114637617021799, + 0.004585026763379574, + 0.00366107071749866, + 0.0063493032939732075, + 0.0036095448303967714, + 0.011278685182332993, + 0.0176401324570179, + 0.00822750199586153, + 0.026303917169570923, + 0.013104886747896671, + 0.012576399371027946, + 0.020146571099758148, + 0.007699677255004644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.03809259459376335 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_2/index_tensor([3698, 3698])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_2/index_tensor([3698, 3698])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c60b65fce107a2f076a7d6b82a3b56190ebe88e1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_2/index_tensor([3698, 3698])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a strikingly low-angle, static foreground shot, a massive Amazonian centipede dominates the frame. Its bizarrely colored carapace reflects an eerie metallic sheen in the strong light, a blend of deep purple and ominous electric blue. Its body arches back suddenly, each chitinous plate tense, while countless pairs of sharp, bright yellow-tipped legs grip the ground, a gesture of poised tension. The camera focuses on its massive, fully extended mandibles, massive and sharp like obsidian sickles, extending outward with a powerful gesture as if in a silent roar, their tips dripping with a sticky, shimmering venom. Behind it, the materialized Japanese onomatopoeia \"ゴゴゴゴ\" (gogogogo) hovers in the dim background, a massive, rough form seemingly carved from energy or rock, each letter emanating an unsteady, pulsating halo of light. The entire scene is illuminated by an unnatural high-contrast light source, casting sharp and deep shadows, making the outline of the centipede even more hideous, full of JOJO-style wild movement and suffocating threat, freezing the creature's primitive aggressiveness in a moment full of visual impact.", + "seed": 1481, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_2/index_3698_seed_1481_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10129406601190567, + "vlm_logprob": -0.10129406601190567, + "vlm_yes_no_margin": 2.45, + "vlm_token_credit": -0.10129406601190567, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with the centipede in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede's carapace colored with a blend of deep purple and electric blue with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs bright yellow-tipped and gripping the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the mandibles massive, sharp like obsidian sickles, and dripping with sticky, shimmering venom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Japanese onomatopoeia 'ゴゴゴゴ' visible in the background as a massive, rough form with a pulsating halo of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04347437992691994 + ], + "opposite_logprob_tokens": [ + -3.4184744358062744 + ], + "logprob_reward": -0.04347437992691994, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04347437992691994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22904524207115173, + 0.05832826346158981, + 0.05769244581460953, + 0.058728352189064026, + 0.04213733226060867, + 0.1786005049943924, + 0.1596260964870453, + 0.08314847946166992, + 0.07137573510408401, + 0.061317577958106995 + ], + "credit_map": [ + 0.030415702611207962, + 0.03290008008480072, + 0.02154131419956684, + 0.02469661831855774, + 0.04059990122914314, + 0.028211496770381927, + 0.028513792902231216, + 0.022166313603520393, + 0.007798424921929836, + 0.009214029647409916, + 0.0047594658099114895, + 0.005184464622288942, + 0.010539046488702297, + 0.006308397743850946, + 0.008543561212718487, + 0.005980868823826313, + 0.005119622219353914, + 0.006108155939728022, + 0.005945865530520678, + 0.00603427505120635, + 0.007227667141705751, + 0.010841622948646545, + 0.007027181796729565, + 0.009388050064444542, + 0.009139678440988064, + 0.00642102025449276, + 0.007619950454682112, + 0.007505745626986027, + 0.007234806194901466, + 0.006600397638976574, + 0.007958547212183475, + 0.006248204503208399, + 0.005031066946685314, + 0.005071664694696665, + 0.005262475460767746, + 0.005578923039138317, + 0.005115682724863291, + 0.005707263946533203, + 0.00476115383207798, + 0.005609099753201008, + 0.01705772988498211, + 0.011741356924176216, + 0.014344647526741028, + 0.0121121471747756, + 0.020362861454486847, + 0.047780126333236694, + 0.03650853782892227, + 0.01869308389723301, + 0.04239383339881897, + 0.020054705440998077, + 0.015693867579102516, + 0.014224506914615631, + 0.019565774127840996, + 0.018007660284638405, + 0.014824519865214825, + 0.014861222356557846, + 0.013714947737753391, + 0.011585628613829613, + 0.009347520768642426, + 0.011880042031407356, + 0.011444426141679287, + 0.0085734399035573, + 0.007414994761347771, + 0.009187474846839905, + 0.008822440169751644, + 0.006371487397700548, + 0.0071667954325675964, + 0.010555559769272804, + 0.013605891726911068, + 0.010537565685808659, + 0.006204486358910799, + 0.008111507631838322, + 0.01122036948800087, + 0.006391533184796572, + 0.005539475008845329, + 0.005436100531369448, + 0.01260306965559721, + 0.009470956400036812, + 0.004863919224590063, + 0.0057921516709029675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.04347437992691994 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with the centipede in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede's carapace colored with a blend of deep purple and electric blue with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs bright yellow-tipped and gripping the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the mandibles massive, sharp like obsidian sickles, and dripping with sticky, shimmering venom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Japanese onomatopoeia 'ゴゴゴゴ' visible in the background as a massive, rough form with a pulsating halo of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430963970720768 + ], + "opposite_logprob_tokens": [ + -3.1680963039398193 + ], + "logprob_reward": -0.0430963970720768, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430963970720768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09348494559526443, + 0.034908246248960495, + 0.06692907214164734, + 0.051422540098428726, + 0.1533515453338623, + 0.2641177773475647, + 0.07256922125816345, + 0.05917135253548622, + 0.08389610052108765, + 0.12014923244714737 + ], + "credit_map": [ + 0.011031213216483593, + 0.009993885643780231, + 0.014470885507762432, + 0.011984615586698055, + 0.012038118205964565, + 0.010480598546564579, + 0.013587186112999916, + 0.009898433461785316, + 0.005227331072092056, + 0.004445318598300219, + 0.0032581589184701443, + 0.0036669201217591763, + 0.006620053667575121, + 0.0037777870893478394, + 0.0043422868475317955, + 0.003570388536900282, + 0.006310102064162493, + 0.003957261331379414, + 0.005337158218026161, + 0.005750261712819338, + 0.0049142250791192055, + 0.01642661727964878, + 0.00630848528817296, + 0.0179249607026577, + 0.006719158962368965, + 0.005523966159671545, + 0.00575751019641757, + 0.006639271974563599, + 0.005294266622513533, + 0.00562298484146595, + 0.009031686000525951, + 0.006833695340901613, + 0.021327221766114235, + 0.022833185270428658, + 0.012409438379108906, + 0.017593709751963615, + 0.021753789857029915, + 0.0240815207362175, + 0.014792515896260738, + 0.018560146912932396, + 0.019254466518759727, + 0.013720034621655941, + 0.013653820380568504, + 0.014001766219735146, + 0.02457323670387268, + 0.06582242995500565, + 0.07313042134046555, + 0.03996158763766289, + 0.009477050974965096, + 0.00837161485105753, + 0.00797094777226448, + 0.010380463674664497, + 0.01181779708713293, + 0.008464885875582695, + 0.008056584745645523, + 0.008029874414205551, + 0.011162721551954746, + 0.005109915044158697, + 0.005737838800996542, + 0.006416152697056532, + 0.009502221830189228, + 0.008755265735089779, + 0.006358820479363203, + 0.006128409411758184, + 0.009169926866889, + 0.006901264656335115, + 0.009543221443891525, + 0.013482457026839256, + 0.016012364998459816, + 0.008789068087935448, + 0.009455827064812183, + 0.010541961528360844, + 0.012878176756203175, + 0.013637220486998558, + 0.009286637417972088, + 0.010784194804728031, + 0.026140479370951653, + 0.02690020389854908, + 0.009935928508639336, + 0.010586380958557129 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 9 + ], + "selected_reward": -0.0430963970720768 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with the centipede in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede's carapace colored with a blend of deep purple and electric blue with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs bright yellow-tipped and gripping the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the mandibles massive, sharp like obsidian sickles, and dripping with sticky, shimmering venom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Japanese onomatopoeia 'ゴゴゴゴ' visible in the background as a massive, rough form with a pulsating halo of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270543485879898 + ], + "opposite_logprob_tokens": [ + -2.127054452896118 + ], + "logprob_reward": -0.1270543485879898, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270543485879898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1432119905948639, + 0.05982040613889694, + 0.058278121054172516, + 0.0521356500685215, + 0.06671278923749924, + 0.20308664441108704, + 0.05556575208902359, + 0.08993850648403168, + 0.15748989582061768, + 0.11376026272773743 + ], + "credit_map": [ + 0.02015332505106926, + 0.01994434930384159, + 0.019329523667693138, + 0.017944421619176865, + 0.0254356749355793, + 0.01615038886666298, + 0.012112867087125778, + 0.012141449376940727, + 0.011424466967582703, + 0.008299645036458969, + 0.005445134826004505, + 0.0051444703713059425, + 0.008822238072752953, + 0.007305259350687265, + 0.006231482606381178, + 0.007147707510739565, + 0.008917147293686867, + 0.004745353013277054, + 0.004182707984000444, + 0.004234140273183584, + 0.004730943590402603, + 0.014642393216490746, + 0.005019401665776968, + 0.011806035414338112, + 0.010363496840000153, + 0.005854855291545391, + 0.004388557747006416, + 0.005230134353041649, + 0.005372749641537666, + 0.005930866580456495, + 0.007701007649302483, + 0.00729397963732481, + 0.014161632396280766, + 0.008907529525458813, + 0.007352782413363457, + 0.00627699913457036, + 0.0065284906886518, + 0.008900254964828491, + 0.006479709409177303, + 0.008105388842523098, + 0.015365594998002052, + 0.015298423357307911, + 0.023897899314761162, + 0.01828560046851635, + 0.032836657017469406, + 0.024757638573646545, + 0.03167768567800522, + 0.04096712917089462, + 0.008738357573747635, + 0.006568148732185364, + 0.005931898020207882, + 0.006838970817625523, + 0.007647781632840633, + 0.007458237931132317, + 0.0067482623271644115, + 0.0056340936571359634, + 0.018163641914725304, + 0.00982190202921629, + 0.009301084093749523, + 0.010062357410788536, + 0.009666532278060913, + 0.01360862236469984, + 0.00809059850871563, + 0.011223766952753067, + 0.023108569905161858, + 0.0164297167211771, + 0.021584169939160347, + 0.018388355150818825, + 0.031096940860152245, + 0.016454095020890236, + 0.014909518882632256, + 0.015518524684011936, + 0.011784030124545097, + 0.01965346187353134, + 0.013357125222682953, + 0.013976626098155975, + 0.015042114071547985, + 0.013893445022404194, + 0.010569860227406025, + 0.015483594499528408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.1270543485879898 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with the centipede in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede's carapace colored with a blend of deep purple and electric blue with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs bright yellow-tipped and gripping the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the mandibles massive, sharp like obsidian sickles, and dripping with sticky, shimmering venom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Japanese onomatopoeia 'ゴゴゴゴ' visible in the background as a massive, rough form with a pulsating halo of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17991754412651062 + ], + "opposite_logprob_tokens": [ + -1.804917573928833 + ], + "logprob_reward": -0.17991754412651062, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17991754412651062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1542282998561859, + 0.07194622606039047, + 0.07533622533082962, + 0.062437910586595535, + 0.07469936460256577, + 0.22434861958026886, + 0.08049718290567398, + 0.09011882543563843, + 0.08744356036186218, + 0.07894380390644073 + ], + "credit_map": [ + 0.018518054857850075, + 0.01765291951596737, + 0.022878777235746384, + 0.01916912943124771, + 0.019728802144527435, + 0.022206200286746025, + 0.016952211037278175, + 0.017122209072113037, + 0.009361742995679379, + 0.01179687399417162, + 0.007593176793307066, + 0.007763891480863094, + 0.011137736029922962, + 0.008628920651972294, + 0.008477005176246166, + 0.007186880335211754, + 0.007023111917078495, + 0.008525639772415161, + 0.008825013414025307, + 0.008026345632970333, + 0.010698213241994381, + 0.009777518920600414, + 0.009613393805921078, + 0.01284699235111475, + 0.010379889979958534, + 0.0070012942887842655, + 0.007689448073506355, + 0.006092460360378027, + 0.006880254950374365, + 0.006152795162051916, + 0.009359639137983322, + 0.008882135152816772, + 0.014321468770503998, + 0.010812257416546345, + 0.008481363765895367, + 0.005735432263463736, + 0.010621489025652409, + 0.009335294365882874, + 0.005081394221633673, + 0.010310675948858261, + 0.017652099952101707, + 0.01348605751991272, + 0.01740836724638939, + 0.01623462326824665, + 0.025158878415822983, + 0.037156373262405396, + 0.06441027671098709, + 0.032841961830854416, + 0.01185715664178133, + 0.00916656106710434, + 0.00954136811196804, + 0.011123538948595524, + 0.01283347886055708, + 0.0088758934289217, + 0.00917601678520441, + 0.007923183031380177, + 0.016719641163945198, + 0.00959306675940752, + 0.012899719178676605, + 0.008757444098591805, + 0.007830864749848843, + 0.010609740391373634, + 0.00896292831748724, + 0.014745427295565605, + 0.011711080558598042, + 0.0076459357514977455, + 0.008770066313445568, + 0.008971158415079117, + 0.02009649947285652, + 0.011354832910001278, + 0.008971227332949638, + 0.009922772645950317, + 0.009577902033925056, + 0.00878998264670372, + 0.006620401982218027, + 0.0073728542774915695, + 0.015904640778899193, + 0.01487429253757, + 0.007026000879704952, + 0.008777742274105549 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.17991754412651062 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with the centipede in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede's carapace colored with a blend of deep purple and electric blue with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs bright yellow-tipped and gripping the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the mandibles massive, sharp like obsidian sickles, and dripping with sticky, shimmering venom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Japanese onomatopoeia 'ゴゴゴゴ' visible in the background as a massive, rough form with a pulsating halo of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292766034603119 + ], + "opposite_logprob_tokens": [ + -2.2379276752471924 + ], + "logprob_reward": -0.11292766034603119, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292766034603119, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1908297836780548, + 0.10770232230424881, + 0.07071411609649658, + 0.05690903961658478, + 0.07485591620206833, + 0.03444557264447212, + 0.1998066008090973, + 0.13861820101737976, + 0.026295559480786324, + 0.09982287138700485 + ], + "credit_map": [ + 0.03056998737156391, + 0.03339379280805588, + 0.015094349160790443, + 0.017952371388673782, + 0.03918769210577011, + 0.024869279935956, + 0.0168868787586689, + 0.012875454500317574, + 0.015051428228616714, + 0.016506807878613472, + 0.008207855746150017, + 0.00837738998234272, + 0.02174954302608967, + 0.01201969850808382, + 0.014161071740090847, + 0.01162853091955185, + 0.006809961050748825, + 0.004597250837832689, + 0.004917978774756193, + 0.0046504284255206585, + 0.005202504340559244, + 0.018190698698163033, + 0.005564725026488304, + 0.02078057825565338, + 0.010247834026813507, + 0.005819639656692743, + 0.0038834528531879187, + 0.005760884843766689, + 0.0094812186434865, + 0.006487978156656027, + 0.007626392878592014, + 0.007601642515510321, + 0.020012304186820984, + 0.010396257042884827, + 0.007418825291097164, + 0.004848123528063297, + 0.007146912626922131, + 0.009968429803848267, + 0.005696985404938459, + 0.009368079714477062, + 0.004862332716584206, + 0.0033547773491591215, + 0.0033332526218146086, + 0.0024994825944304466, + 0.005573740694671869, + 0.0058778696693480015, + 0.005072579253464937, + 0.00387153890915215, + 0.06471219658851624, + 0.018520142883062363, + 0.017764072865247726, + 0.013992827385663986, + 0.019407296553254128, + 0.02780270017683506, + 0.020581873133778572, + 0.01702546887099743, + 0.026485556736588478, + 0.018626095727086067, + 0.011728753335773945, + 0.015713952481746674, + 0.01472023967653513, + 0.01549046952277422, + 0.013142416253685951, + 0.022710714489221573, + 0.003924562595784664, + 0.001971875550225377, + 0.0033828257583081722, + 0.0034696778748184443, + 0.004982566460967064, + 0.0029420380014926195, + 0.0032518208026885986, + 0.002370191738009453, + 0.00931969378143549, + 0.008743155747652054, + 0.005962247494608164, + 0.005759957246482372, + 0.029480386525392532, + 0.02220727875828743, + 0.008437156677246094, + 0.009912999346852303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.11292766034603119 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a strikingly low-angle, static foreground shot, a massive Amazonian centipede dominates the frame. Its bizarrely colored carapace reflects an eerie metallic sheen in the strong light, a blend of deep purple and ominous electric blue. Its body arches back suddenly, each chitinous plate tense, while countless pairs of sharp, bright yellow-tipped legs grip the ground, a gesture of poised tension. The camera focuses on its massive, fully extended mandibles, massive and sharp like obsidian sickles, extending outward with a powerful gesture as if in a silent roar, their tips dripping with a sticky, shimmering venom. Behind it, the materialized Japanese onomatopoeia \"ゴゴゴゴ\" (gogogogo) hovers in the dim background, a massive, rough form seemingly carved from energy or rock, each letter emanating an unsteady, pulsating halo of light. The entire scene is illuminated by an unnatural high-contrast light source, casting sharp and deep shadows, making the outline of the centipede even more hideous, full of JOJO-style wild movement and suffocating threat, freezing the creature's primitive aggressiveness in a moment full of visual impact.", + "seed": 1482, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_2/index_3698_seed_1482_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1474974375218153, + "vlm_logprob": -0.1474974375218153, + "vlm_yes_no_margin": 2.3, + "vlm_token_credit": -0.1474974375218153, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with the centipede in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede's carapace colored with a blend of deep purple and electric blue with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs bright yellow-tipped and gripping the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the mandibles massive, sharp like obsidian sickles, and dripping with sticky, shimmering venom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Japanese onomatopoeia 'ゴゴゴゴ' visible in the background as a massive, rough form with a pulsating halo of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014439260587096214 + ], + "opposite_logprob_tokens": [ + -4.639439105987549 + ], + "logprob_reward": -0.014439260587096214, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.014439260587096214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21103566884994507, + 0.2731073796749115, + 0.08555583655834198, + 0.046580735594034195, + 0.041792258620262146, + 0.1295587718486786, + 0.0343705490231514, + 0.05461909621953964, + 0.0644589513540268, + 0.05892077833414078 + ], + "credit_map": [ + 0.03686967119574547, + 0.019168825820088387, + 0.037650927901268005, + 0.02661486342549324, + 0.019071411341428757, + 0.02232666313648224, + 0.02548559568822384, + 0.023847708478569984, + 0.023105336353182793, + 0.03179235756397247, + 0.0320822149515152, + 0.03702917695045471, + 0.043526582419872284, + 0.025790685787796974, + 0.04493318870663643, + 0.03484783694148064, + 0.00835807528346777, + 0.011119461618363857, + 0.00789086613804102, + 0.016093969345092773, + 0.011185300536453724, + 0.006840110756456852, + 0.014092797413468361, + 0.009975261054933071, + 0.004026886075735092, + 0.004124152008444071, + 0.005908633582293987, + 0.008798541501164436, + 0.004940606653690338, + 0.005113127175718546, + 0.007414310704916716, + 0.006254471372812986, + 0.004129523411393166, + 0.003982975147664547, + 0.005320793949067593, + 0.0056650531478226185, + 0.0063789295963943005, + 0.005221537314355373, + 0.006208829581737518, + 0.004884613677859306, + 0.013878462836146355, + 0.012426567263901234, + 0.014084206894040108, + 0.013885041698813438, + 0.015173610299825668, + 0.03355272859334946, + 0.012603242881596088, + 0.013954912312328815, + 0.0042919619008898735, + 0.004455737303942442, + 0.0033138561993837357, + 0.0032836159225553274, + 0.007055931258946657, + 0.005665535572916269, + 0.002996420953422785, + 0.0033074882812798023, + 0.0068340119905769825, + 0.006226790137588978, + 0.006471818313002586, + 0.008496036753058434, + 0.006183619610965252, + 0.0076434919610619545, + 0.006413592025637627, + 0.00634973356500268, + 0.004377526696771383, + 0.00894667487591505, + 0.009566646069288254, + 0.00816225353628397, + 0.011566072702407837, + 0.007049106061458588, + 0.0074087330140173435, + 0.007381940260529518, + 0.006018970627337694, + 0.009355772286653519, + 0.007751192431896925, + 0.0060967556200921535, + 0.008451425470411777, + 0.00720325019210577, + 0.006503777112811804, + 0.007539630401879549 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.014439260587096214 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with the centipede in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede's carapace colored with a blend of deep purple and electric blue with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs bright yellow-tipped and gripping the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the mandibles massive, sharp like obsidian sickles, and dripping with sticky, shimmering venom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Japanese onomatopoeia 'ゴゴゴゴ' visible in the background as a massive, rough form with a pulsating halo of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906160295009613 + ], + "opposite_logprob_tokens": [ + -2.4640614986419678 + ], + "logprob_reward": -0.08906160295009613, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906160295009613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2844006419181824, + 0.21365728974342346, + 0.06778465956449509, + 0.017997946590185165, + 0.03809218853712082, + 0.12163033336400986, + 0.050175316631793976, + 0.07124888896942139, + 0.06622643768787384, + 0.06878631561994553 + ], + "credit_map": [ + 0.049295153468847275, + 0.030301634222269058, + 0.0434085838496685, + 0.0327925980091095, + 0.03199107572436333, + 0.029498016461730003, + 0.03421561047434807, + 0.03289797529578209, + 0.0193742997944355, + 0.021277712658047676, + 0.02746276929974556, + 0.03193044662475586, + 0.031011130660772324, + 0.02076263353228569, + 0.034216850996017456, + 0.027621448040008545, + 0.007160155102610588, + 0.007886381819844246, + 0.006677846424281597, + 0.009177657775580883, + 0.013590178452432156, + 0.005680756643414497, + 0.009330961853265762, + 0.008280723355710506, + 0.0025219654198735952, + 0.0024797599762678146, + 0.0017547508468851447, + 0.002225505653768778, + 0.0018713477766141295, + 0.002748606028035283, + 0.002373718423768878, + 0.002022292697802186, + 0.0045697432942688465, + 0.0031135466415435076, + 0.002948418725281954, + 0.003959459252655506, + 0.0048913718201220036, + 0.012112623080611229, + 0.003914526663720608, + 0.002582498360425234, + 0.01382347010076046, + 0.010985703207552433, + 0.011511791497468948, + 0.018806500360369682, + 0.013589284382760525, + 0.023203860968351364, + 0.015293871983885765, + 0.014415848068892956, + 0.005015726666897535, + 0.007383996620774269, + 0.0048010870814323425, + 0.0035824596416205168, + 0.016118580475449562, + 0.006097546312958002, + 0.0040840813890099525, + 0.0030918431002646685, + 0.008820747956633568, + 0.00964006595313549, + 0.006431254558265209, + 0.010300268419086933, + 0.011521779000759125, + 0.007807641290128231, + 0.008848435245454311, + 0.007878697477281094, + 0.006039753090590239, + 0.00826928112655878, + 0.011982826516032219, + 0.0077278222888708115, + 0.007402345072478056, + 0.006975137162953615, + 0.005428795237094164, + 0.012400471605360508, + 0.00620881887152791, + 0.009838066063821316, + 0.00820319913327694, + 0.0067861685529351234, + 0.009847030974924564, + 0.008182586170732975, + 0.009965802542865276, + 0.009754648432135582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.08906160295009613 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with the centipede in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede's carapace colored with a blend of deep purple and electric blue with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs bright yellow-tipped and gripping the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the mandibles massive, sharp like obsidian sickles, and dripping with sticky, shimmering venom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Japanese onomatopoeia 'ゴゴゴゴ' visible in the background as a massive, rough form with a pulsating halo of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22558407485485077 + ], + "opposite_logprob_tokens": [ + -1.6005840301513672 + ], + "logprob_reward": -0.22558407485485077, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22558407485485077, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08534222096204758, + 0.2575329840183258, + 0.22169913351535797, + 0.07517719268798828, + 0.07120631635189056, + 0.07567271590232849, + 0.03200671076774597, + 0.03654588386416435, + 0.08534900844097137, + 0.059467803686857224 + ], + "credit_map": [ + 0.014092820696532726, + 0.007846525870263577, + 0.014465023763477802, + 0.010566777549684048, + 0.009687187150120735, + 0.00925508327782154, + 0.010184522718191147, + 0.009244276210665703, + 0.023603815585374832, + 0.027685176581144333, + 0.03376329317688942, + 0.03512139618396759, + 0.038828663527965546, + 0.02458062581717968, + 0.04098384082317352, + 0.032966189086437225, + 0.018959954380989075, + 0.026278603821992874, + 0.022498302161693573, + 0.03376605734229088, + 0.03149530291557312, + 0.01760466955602169, + 0.04300903156399727, + 0.028087208047509193, + 0.005940630100667477, + 0.013166571967303753, + 0.010463306680321693, + 0.010327952913939953, + 0.010991482995450497, + 0.008657272905111313, + 0.0081715676933527, + 0.007458406034857035, + 0.007989432662725449, + 0.009090716950595379, + 0.0065659224055707455, + 0.004411973524838686, + 0.01664603315293789, + 0.012248010374605656, + 0.006788081955164671, + 0.007466139271855354, + 0.008297236636281013, + 0.006829591002315283, + 0.007715762592852116, + 0.009597048163414001, + 0.008893902413547039, + 0.01715116575360298, + 0.008697188459336758, + 0.008490823209285736, + 0.003121444722637534, + 0.0037543291691690683, + 0.0034712208434939384, + 0.002718554111197591, + 0.007431717589497566, + 0.005038735456764698, + 0.003450642805546522, + 0.0030200656037777662, + 0.004887726623564959, + 0.0041206213645637035, + 0.004230355378240347, + 0.005150837358087301, + 0.0033443253487348557, + 0.0057123270817101, + 0.004888343624770641, + 0.00421134801581502, + 0.00810214877128601, + 0.007937696762382984, + 0.017138414084911346, + 0.0101009551435709, + 0.009805549867451191, + 0.00951316300779581, + 0.009605403058230877, + 0.013145672157406807, + 0.004846630617976189, + 0.009975423105061054, + 0.007084391545504332, + 0.006455165334045887, + 0.005907900631427765, + 0.008065701462328434, + 0.008745419792830944, + 0.008387167938053608 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 8 + ], + "selected_reward": -0.22558407485485077 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with the centipede in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede's carapace colored with a blend of deep purple and electric blue with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs bright yellow-tipped and gripping the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the mandibles massive, sharp like obsidian sickles, and dripping with sticky, shimmering venom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Japanese onomatopoeia 'ゴゴゴゴ' visible in the background as a massive, rough form with a pulsating halo of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28135472536087036 + ], + "opposite_logprob_tokens": [ + -1.4063546657562256 + ], + "logprob_reward": -0.28135472536087036, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28135472536087036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23852647840976715, + 0.13755139708518982, + 0.16769976913928986, + 0.04171857237815857, + 0.03093479946255684, + 0.047686491161584854, + 0.03951692953705788, + 0.19300100207328796, + 0.06326064467430115, + 0.040103983134031296 + ], + "credit_map": [ + 0.04152299463748932, + 0.025659998878836632, + 0.03654889762401581, + 0.027953511103987694, + 0.026355691254138947, + 0.024231797084212303, + 0.0289111640304327, + 0.027342427521944046, + 0.012193207629024982, + 0.013728895224630833, + 0.01734468713402748, + 0.021746860817074776, + 0.019840480759739876, + 0.013524437323212624, + 0.02159193344414234, + 0.017580891028046608, + 0.015733828768134117, + 0.014609558507800102, + 0.017938416451215744, + 0.036486927419900894, + 0.02190890535712242, + 0.014514802023768425, + 0.027094770222902298, + 0.019412554800510406, + 0.004541302565485239, + 0.005533238407224417, + 0.003844267688691616, + 0.008348038420081139, + 0.004735481925308704, + 0.005887313280254602, + 0.005321040749549866, + 0.0035078884102404118, + 0.00414707325398922, + 0.0035653093364089727, + 0.0027011805213987827, + 0.002392548369243741, + 0.005593229550868273, + 0.007044075056910515, + 0.0029955068603157997, + 0.0024958751164376736, + 0.005208831746131182, + 0.003715728409588337, + 0.005990033037960529, + 0.0067202067002654076, + 0.005185883026570082, + 0.010941920801997185, + 0.0049207862466573715, + 0.005003101192414761, + 0.004095320124179125, + 0.00406290777027607, + 0.005521767772734165, + 0.0038515205960720778, + 0.006701256614178419, + 0.0042524440214037895, + 0.00487774470821023, + 0.006153967697173357, + 0.019082969054579735, + 0.018682146444916725, + 0.018534621223807335, + 0.03353167697787285, + 0.022175436839461327, + 0.027333585545420647, + 0.02818155661225319, + 0.025478998199105263, + 0.005688004195690155, + 0.0059893191792070866, + 0.012751695699989796, + 0.007660750299692154, + 0.006444221828132868, + 0.006389977410435677, + 0.006841450463980436, + 0.011495227925479412, + 0.003973761107772589, + 0.005716546438634396, + 0.004787063226103783, + 0.0036342402454465628, + 0.004234557971358299, + 0.008468538522720337, + 0.00468777259811759, + 0.004601499997079372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.28135472536087036 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with the centipede in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede's carapace colored with a blend of deep purple and electric blue with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs bright yellow-tipped and gripping the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the mandibles massive, sharp like obsidian sickles, and dripping with sticky, shimmering venom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Japanese onomatopoeia 'ゴゴゴゴ' visible in the background as a massive, rough form with a pulsating halo of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704752385616302 + ], + "opposite_logprob_tokens": [ + -2.127047538757324 + ], + "logprob_reward": -0.12704752385616302, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704752385616302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05348096787929535, + 0.10070385038852692, + 0.4709987938404083, + 0.02048027329146862, + 0.030658364295959473, + 0.03212730213999748, + 0.07503069192171097, + 0.13266390562057495, + 0.024758046492934227, + 0.05909787490963936 + ], + "credit_map": [ + 0.010458644479513168, + 0.008029635064303875, + 0.006460410077124834, + 0.005429790820926428, + 0.005688381381332874, + 0.0066277338191866875, + 0.005540504585951567, + 0.005245865788310766, + 0.010038664564490318, + 0.009205789305269718, + 0.013056514784693718, + 0.02622886188328266, + 0.0118649210780859, + 0.009503299370408058, + 0.011510079726576805, + 0.009295712225139141, + 0.03906011953949928, + 0.06320347636938095, + 0.04812812805175781, + 0.0555289126932621, + 0.07029485702514648, + 0.03554709255695343, + 0.09413913637399673, + 0.06509704887866974, + 0.002375646261498332, + 0.0024434905499219894, + 0.002002003602683544, + 0.0033176010474562645, + 0.002200711751356721, + 0.0035272990353405476, + 0.0021869505289942026, + 0.0024265681859105825, + 0.0031547946855425835, + 0.003003818914294243, + 0.003098894841969013, + 0.0031745336018502712, + 0.004849749151617289, + 0.00721115292981267, + 0.0036169555969536304, + 0.002548462012782693, + 0.003084725933149457, + 0.004063117783516645, + 0.003329925239086151, + 0.004490321967750788, + 0.005333151668310165, + 0.0036178550217300653, + 0.004091046284884214, + 0.004117155447602272, + 0.006821167189627886, + 0.0076675256714224815, + 0.009727791883051395, + 0.0076791890896856785, + 0.015194445848464966, + 0.008083137683570385, + 0.009599040262401104, + 0.0102583859115839, + 0.013313381001353264, + 0.019435826689004898, + 0.014762909151613712, + 0.019262973219156265, + 0.014911239966750145, + 0.018424909561872482, + 0.01567940227687359, + 0.01687324419617653, + 0.001899836934171617, + 0.003383154748007655, + 0.0037453121040016413, + 0.0027671819552779198, + 0.0033229105174541473, + 0.003068285994231701, + 0.0028954760637134314, + 0.003675886895507574, + 0.005324874073266983, + 0.011526376940310001, + 0.007323415484279394, + 0.0063553787767887115, + 0.007752588950097561, + 0.0072730411775410175, + 0.006146387662738562, + 0.007395806722342968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 1 + ], + "selected_reward": -0.12704752385616302 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a strikingly low-angle, static foreground shot, a massive Amazonian centipede dominates the frame. Its bizarrely colored carapace reflects an eerie metallic sheen in the strong light, a blend of deep purple and ominous electric blue. Its body arches back suddenly, each chitinous plate tense, while countless pairs of sharp, bright yellow-tipped legs grip the ground, a gesture of poised tension. The camera focuses on its massive, fully extended mandibles, massive and sharp like obsidian sickles, extending outward with a powerful gesture as if in a silent roar, their tips dripping with a sticky, shimmering venom. Behind it, the materialized Japanese onomatopoeia \"ゴゴゴゴ\" (gogogogo) hovers in the dim background, a massive, rough form seemingly carved from energy or rock, each letter emanating an unsteady, pulsating halo of light. The entire scene is illuminated by an unnatural high-contrast light source, casting sharp and deep shadows, making the outline of the centipede even more hideous, full of JOJO-style wild movement and suffocating threat, freezing the creature's primitive aggressiveness in a moment full of visual impact.", + "seed": 1483, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_2/index_3698_seed_1483_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1248345948755741, + "vlm_logprob": -0.1248345948755741, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.1248345948755741, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with the centipede in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede's carapace colored with a blend of deep purple and electric blue with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs bright yellow-tipped and gripping the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the mandibles massive, sharp like obsidian sickles, and dripping with sticky, shimmering venom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Japanese onomatopoeia 'ゴゴゴゴ' visible in the background as a massive, rough form with a pulsating halo of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02921578660607338 + ], + "opposite_logprob_tokens": [ + -3.9042158126831055 + ], + "logprob_reward": -0.02921578660607338, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02921578660607338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07978911697864532, + 0.18591423332691193, + 0.030701687559485435, + 0.037480756640434265, + 0.023472420871257782, + 0.07808743417263031, + 0.027059009298682213, + 0.47249653935432434, + 0.04454822093248367, + 0.02045060135424137 + ], + "credit_map": [ + 0.0068357218988239765, + 0.005554163828492165, + 0.008388428948819637, + 0.009939426556229591, + 0.009880010038614273, + 0.013266566209495068, + 0.014574453234672546, + 0.011350348591804504, + 0.027758080512285233, + 0.007249620743095875, + 0.018596261739730835, + 0.017520004883408546, + 0.008478946052491665, + 0.024932080879807472, + 0.05995175614953041, + 0.021427500993013382, + 0.003927930258214474, + 0.0028923973441123962, + 0.004617985337972641, + 0.003978798631578684, + 0.00406411848962307, + 0.003760790918022394, + 0.004469729959964752, + 0.0029899408109486103, + 0.005095313768833876, + 0.0036091350484639406, + 0.006375669967383146, + 0.004030659794807434, + 0.004009867087006569, + 0.004424840211868286, + 0.005063725635409355, + 0.004871549550443888, + 0.003061875933781266, + 0.0031435606069862843, + 0.0026348684914410114, + 0.002913733944296837, + 0.0033832499757409096, + 0.0028025938663631678, + 0.003231398994103074, + 0.0023011404555290937, + 0.00586702348664403, + 0.0038418290205299854, + 0.004565672483295202, + 0.005176158621907234, + 0.0039630127139389515, + 0.024311799556016922, + 0.024572378024458885, + 0.0057895611971616745, + 0.0026674452237784863, + 0.0023056904319673777, + 0.002545538591220975, + 0.004547808784991503, + 0.001974892569705844, + 0.0037015541456639767, + 0.003599000396206975, + 0.00571708008646965, + 0.026772651821374893, + 0.009909051470458508, + 0.021469393745064735, + 0.02288001962006092, + 0.017825838178396225, + 0.03129473328590393, + 0.1724012941122055, + 0.16994355618953705, + 0.00327664939686656, + 0.0016860621981322765, + 0.002394258975982666, + 0.0025586900301277637, + 0.002118394011631608, + 0.004279118496924639, + 0.01695580594241619, + 0.011279240250587463, + 0.003208821639418602, + 0.002341612009331584, + 0.002016282407566905, + 0.0018275148468092084, + 0.0022657744120806456, + 0.0024601048789918423, + 0.003163502784445882, + 0.003166990587487817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.02921578660607338 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with the centipede in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede's carapace colored with a blend of deep purple and electric blue with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs bright yellow-tipped and gripping the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the mandibles massive, sharp like obsidian sickles, and dripping with sticky, shimmering venom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Japanese onomatopoeia 'ゴゴゴゴ' visible in the background as a massive, rough form with a pulsating halo of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048683732748031616 + ], + "opposite_logprob_tokens": [ + -3.0486836433410645 + ], + "logprob_reward": -0.048683732748031616, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048683732748031616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0655941516160965, + 0.3051831126213074, + 0.0435900017619133, + 0.043208908289670944, + 0.011294707655906677, + 0.015949303284287453, + 0.05048603191971779, + 0.41093698143959045, + 0.03614308312535286, + 0.017613684758543968 + ], + "credit_map": [ + 0.006964591797441244, + 0.005338259041309357, + 0.007390518672764301, + 0.009242698550224304, + 0.0070802681148052216, + 0.008090617135167122, + 0.013204464688897133, + 0.008282739669084549, + 0.03763922303915024, + 0.010029390454292297, + 0.030408333986997604, + 0.03340277075767517, + 0.013838632963597775, + 0.03492431342601776, + 0.12200408428907394, + 0.022936362773180008, + 0.004877220373600721, + 0.0038329854141920805, + 0.006229925435036421, + 0.007332094945013523, + 0.004010497126728296, + 0.006002645008265972, + 0.006220994982868433, + 0.005083635915070772, + 0.004317276179790497, + 0.0034820681903511286, + 0.008497742004692554, + 0.006419062148779631, + 0.003345176577568054, + 0.0046801138669252396, + 0.006813704967498779, + 0.005653764121234417, + 0.0013397476868703961, + 0.0013141147792339325, + 0.0012788529274985194, + 0.0013807722134515643, + 0.00160426483489573, + 0.0013527215924113989, + 0.0012784779537469149, + 0.0017457547364756465, + 0.0015078731812536716, + 0.001637943903915584, + 0.0015113606350496411, + 0.001056921319104731, + 0.0010900201741605997, + 0.004012326709926128, + 0.0029739458113908768, + 0.002158910734578967, + 0.004518874920904636, + 0.003919686656445265, + 0.006250015925616026, + 0.011749562807381153, + 0.003284843871369958, + 0.0055160908959805965, + 0.008547179400920868, + 0.006699778605252504, + 0.023987237364053726, + 0.009490227326750755, + 0.019212869927287102, + 0.02069837413728237, + 0.01600799337029457, + 0.026946229860186577, + 0.14883869886398315, + 0.1457553505897522, + 0.00221787067130208, + 0.0007047852850519121, + 0.0014149746857583523, + 0.0017726076766848564, + 0.0014030244201421738, + 0.003529256209731102, + 0.014745672233402729, + 0.01035489235073328, + 0.0024169611278921366, + 0.0014745317166671157, + 0.0021785658318549395, + 0.002677153330296278, + 0.0011575239477679133, + 0.0023735936265438795, + 0.002773710060864687, + 0.0025616460479795933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.048683732748031616 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with the centipede in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede's carapace colored with a blend of deep purple and electric blue with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs bright yellow-tipped and gripping the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the mandibles massive, sharp like obsidian sickles, and dripping with sticky, shimmering venom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Japanese onomatopoeia 'ゴゴゴゴ' visible in the background as a massive, rough form with a pulsating halo of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22557055950164795 + ], + "opposite_logprob_tokens": [ + -1.600570559501648 + ], + "logprob_reward": -0.22557055950164795, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22557055950164795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05656382068991661, + 0.20270609855651855, + 0.0443987213075161, + 0.017474062740802765, + 0.017921440303325653, + 0.04275774955749512, + 0.017662718892097473, + 0.516066312789917, + 0.03009369783103466, + 0.05435538291931152 + ], + "credit_map": [ + 0.00769251212477684, + 0.0067006852477788925, + 0.005583517719060183, + 0.006100787781178951, + 0.007583112455904484, + 0.007758994586765766, + 0.008642465807497501, + 0.006501742172986269, + 0.028421694412827492, + 0.007116563152521849, + 0.019902024418115616, + 0.02079862728714943, + 0.010074655525386333, + 0.025458497926592827, + 0.07274506241083145, + 0.01818898320198059, + 0.004248438868671656, + 0.003188190283253789, + 0.0060469722375273705, + 0.008974092081189156, + 0.0038286824710667133, + 0.0060543701983988285, + 0.0058632283471524715, + 0.006194746121764183, + 0.0017075976356863976, + 0.0013620286481454968, + 0.003136721672490239, + 0.0021749441511929035, + 0.0017025070264935493, + 0.003594657639041543, + 0.002093171002343297, + 0.0017024335684254766, + 0.0018484393367543817, + 0.0017363959923386574, + 0.0022406752686947584, + 0.002378286560997367, + 0.0024515611585229635, + 0.002161875832825899, + 0.0024343677796423435, + 0.0026698389556258917, + 0.003832900431007147, + 0.0037324100267142057, + 0.003919559996575117, + 0.0027195648290216923, + 0.0032107161823660135, + 0.011050932109355927, + 0.008571196347475052, + 0.0057204719632864, + 0.001842365600168705, + 0.001743459957651794, + 0.0030607092194259167, + 0.0021979697048664093, + 0.001557183568365872, + 0.0022394550032913685, + 0.0028403890319168568, + 0.00218118610791862, + 0.029650138691067696, + 0.011142592877149582, + 0.024491913616657257, + 0.0259716734290123, + 0.019427884370088577, + 0.033788882195949554, + 0.1863575428724289, + 0.1852356642484665, + 0.0019655455835163593, + 0.000656988238915801, + 0.0013930826680734754, + 0.001660288660787046, + 0.0010995273478329182, + 0.002438010647892952, + 0.01163793820887804, + 0.009242314845323563, + 0.0054808491840958595, + 0.0031265090219676495, + 0.006257548462599516, + 0.008890731260180473, + 0.002975969575345516, + 0.006502701435238123, + 0.009863057173788548, + 0.011258013546466827 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.22557055950164795 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with the centipede in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede's carapace colored with a blend of deep purple and electric blue with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs bright yellow-tipped and gripping the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the mandibles massive, sharp like obsidian sickles, and dripping with sticky, shimmering venom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Japanese onomatopoeia 'ゴゴゴゴ' visible in the background as a massive, rough form with a pulsating halo of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.160362109541893 + ], + "opposite_logprob_tokens": [ + -1.9103621244430542 + ], + "logprob_reward": -0.160362109541893, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.160362109541893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18878696858882904, + 0.3023393750190735, + 0.07588434964418411, + 0.06396492570638657, + 0.024285651743412018, + 0.07436291128396988, + 0.04273919388651848, + 0.16238461434841156, + 0.031994953751564026, + 0.033257074654102325 + ], + "credit_map": [ + 0.021342335268855095, + 0.016185563057661057, + 0.019627366214990616, + 0.024735338985919952, + 0.029618550091981888, + 0.024289121851325035, + 0.02854219824075699, + 0.024446498602628708, + 0.04315057769417763, + 0.011374332010746002, + 0.029920566827058792, + 0.03138541057705879, + 0.014045917429029942, + 0.036676596850156784, + 0.10503865033388138, + 0.03074730560183525, + 0.007607556413859129, + 0.005754783749580383, + 0.011457287706434727, + 0.013109685853123665, + 0.005801825318485498, + 0.009700560942292213, + 0.014077837578952312, + 0.00837481115013361, + 0.006143426522612572, + 0.0052442653104662895, + 0.012504158541560173, + 0.009380063973367214, + 0.005507458932697773, + 0.0069700269959867, + 0.009793300181627274, + 0.008422226645052433, + 0.002760546049103141, + 0.0027727121487259865, + 0.00306374067440629, + 0.002665797946974635, + 0.004180326592177153, + 0.0032790559343993664, + 0.0029161537531763315, + 0.002647318411618471, + 0.00467499066144228, + 0.002335528377443552, + 0.004227602854371071, + 0.004949274938553572, + 0.0032473248429596424, + 0.022067828103899956, + 0.02704496867954731, + 0.005815393757075071, + 0.0037695281207561493, + 0.00351398391649127, + 0.004894474521279335, + 0.009753048419952393, + 0.0031170505098998547, + 0.005286291241645813, + 0.007036228198558092, + 0.005368589889258146, + 0.010257185436785221, + 0.005303477868437767, + 0.009120365604758263, + 0.009412902407348156, + 0.006946972571313381, + 0.011092317290604115, + 0.05505896732211113, + 0.05519242212176323, + 0.0019839692395180464, + 0.0009425826137885451, + 0.0014776946045458317, + 0.001771391718648374, + 0.0015017078258097172, + 0.0034052010159939528, + 0.012444846332073212, + 0.008467557840049267, + 0.0047016264870762825, + 0.0028460444882512093, + 0.002716179471462965, + 0.003053282154724002, + 0.003239769721403718, + 0.004606813192367554, + 0.008001072332262993, + 0.004092290066182613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.160362109541893 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with the centipede in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede's carapace colored with a blend of deep purple and electric blue with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs bright yellow-tipped and gripping the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the mandibles massive, sharp like obsidian sickles, and dripping with sticky, shimmering venom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Japanese onomatopoeia 'ゴゴゴゴ' visible in the background as a massive, rough form with a pulsating halo of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603407859802246 + ], + "opposite_logprob_tokens": [ + -1.9103407859802246 + ], + "logprob_reward": -0.1603407859802246, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603407859802246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11856777966022491, + 0.13833314180374146, + 0.05564872547984123, + 0.036646854132413864, + 0.024188751354813576, + 0.0509442500770092, + 0.047309134155511856, + 0.4848562777042389, + 0.029901200905442238, + 0.013603911735117435 + ], + "credit_map": [ + 0.014383357018232346, + 0.013339270837605, + 0.010688714683055878, + 0.01408722810447216, + 0.015018190257251263, + 0.024448169395327568, + 0.016335371881723404, + 0.01026748213917017, + 0.012819355353713036, + 0.004621436353772879, + 0.012524156831204891, + 0.01340596191585064, + 0.007371579296886921, + 0.014144036918878555, + 0.05710362643003464, + 0.01634298823773861, + 0.005421773064881563, + 0.005624115001410246, + 0.005619278643280268, + 0.005788540001958609, + 0.004977251403033733, + 0.010892197489738464, + 0.01284701470285654, + 0.004478554241359234, + 0.004042801447212696, + 0.007238931953907013, + 0.00395252276211977, + 0.003134301397949457, + 0.004853524267673492, + 0.007035619579255581, + 0.0036897414829581976, + 0.0026994114741683006, + 0.0028302425052970648, + 0.0030744983814656734, + 0.002814414445310831, + 0.0024760598316788673, + 0.005074824206531048, + 0.0033044246956706047, + 0.0025779488496482372, + 0.002036337275058031, + 0.0028148300480097532, + 0.001849641790613532, + 0.0038470178842544556, + 0.003793353447690606, + 0.002137318719178438, + 0.013531534932553768, + 0.018237318843603134, + 0.0047332365065813065, + 0.005703589413315058, + 0.003705010749399662, + 0.007090107072144747, + 0.008609567768871784, + 0.0031347577460110188, + 0.004987758584320545, + 0.008117154240608215, + 0.005961186718195677, + 0.026027686893939972, + 0.010855146683752537, + 0.02296464331448078, + 0.02513156831264496, + 0.01846800558269024, + 0.032856862992048264, + 0.17547480762004852, + 0.1730775535106659, + 0.0020966585725545883, + 0.0009729975718073547, + 0.0013928404077887535, + 0.0017515416257083416, + 0.0014844840625301003, + 0.0028140456415712833, + 0.011196494102478027, + 0.00819213967770338, + 0.0014796778559684753, + 0.0010309702483937144, + 0.0015897006960585713, + 0.002366479253396392, + 0.0009670585859566927, + 0.0017001210944727063, + 0.0025274090003222227, + 0.0019424943020567298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.1603407859802246 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a strikingly low-angle, static foreground shot, a massive Amazonian centipede dominates the frame. Its bizarrely colored carapace reflects an eerie metallic sheen in the strong light, a blend of deep purple and ominous electric blue. Its body arches back suddenly, each chitinous plate tense, while countless pairs of sharp, bright yellow-tipped legs grip the ground, a gesture of poised tension. The camera focuses on its massive, fully extended mandibles, massive and sharp like obsidian sickles, extending outward with a powerful gesture as if in a silent roar, their tips dripping with a sticky, shimmering venom. Behind it, the materialized Japanese onomatopoeia \"ゴゴゴゴ\" (gogogogo) hovers in the dim background, a massive, rough form seemingly carved from energy or rock, each letter emanating an unsteady, pulsating halo of light. The entire scene is illuminated by an unnatural high-contrast light source, casting sharp and deep shadows, making the outline of the centipede even more hideous, full of JOJO-style wild movement and suffocating threat, freezing the creature's primitive aggressiveness in a moment full of visual impact.", + "seed": 1484, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_2/index_3698_seed_1484_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0726839292794466, + "vlm_logprob": -0.0726839292794466, + "vlm_yes_no_margin": 2.899999976158142, + "vlm_token_credit": -0.0726839292794466, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with the centipede in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede's carapace colored with a blend of deep purple and electric blue with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs bright yellow-tipped and gripping the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the mandibles massive, sharp like obsidian sickles, and dripping with sticky, shimmering venom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Japanese onomatopoeia 'ゴゴゴゴ' visible in the background as a massive, rough form with a pulsating halo of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023708192631602287 + ], + "opposite_logprob_tokens": [ + -4.148708343505859 + ], + "logprob_reward": -0.023708192631602287, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.023708192631602287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19132700562477112, + 0.10463898628950119, + 0.12207937985658646, + 0.1880062222480774, + 0.039843324571847916, + 0.08405033499002457, + 0.06203414872288704, + 0.03436511754989624, + 0.06736870855093002, + 0.10628674924373627 + ], + "credit_map": [ + 0.017675179988145828, + 0.018050651997327805, + 0.020165294408798218, + 0.02491411752998829, + 0.03072606399655342, + 0.021540852263569832, + 0.030539998784661293, + 0.027714865282177925, + 0.010420260950922966, + 0.01222892478108406, + 0.015682648867368698, + 0.01413026638329029, + 0.01127139013260603, + 0.006733992137014866, + 0.020130233839154243, + 0.014041286893188953, + 0.02046777866780758, + 0.013827276416122913, + 0.011073964647948742, + 0.018323592841625214, + 0.011241339147090912, + 0.00791978184133768, + 0.017123624682426453, + 0.022102024406194687, + 0.0388818085193634, + 0.013448906131088734, + 0.020173704251646996, + 0.013577976264059544, + 0.0173783116042614, + 0.010024175979197025, + 0.01792452484369278, + 0.05659683048725128, + 0.0064019011333584785, + 0.004374251235276461, + 0.004161193035542965, + 0.005445840768516064, + 0.005398299545049667, + 0.0044508399441838264, + 0.004891891963779926, + 0.0047191111370921135, + 0.011103260330855846, + 0.010851576924324036, + 0.013937657698988914, + 0.01055450551211834, + 0.009183979593217373, + 0.010380689054727554, + 0.010687430389225483, + 0.007351241540163755, + 0.005951971746981144, + 0.006275641266256571, + 0.007579753175377846, + 0.009864765219390392, + 0.005522345658391714, + 0.005469310563057661, + 0.014545900747179985, + 0.0068244729191064835, + 0.0033006907906383276, + 0.002753661945462227, + 0.0037653630133718252, + 0.005309572443366051, + 0.00363244884647429, + 0.006534272339195013, + 0.0039910986088216305, + 0.005078012123703957, + 0.007906789891421795, + 0.0069883414544165134, + 0.007637885399162769, + 0.008501150645315647, + 0.007641048636287451, + 0.010589114390313625, + 0.010855409316718578, + 0.007248971611261368, + 0.014165520668029785, + 0.01020823884755373, + 0.014159840531647205, + 0.011940030381083488, + 0.009408385492861271, + 0.00843031331896782, + 0.014715461060404778, + 0.023258967325091362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.023708192631602287 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with the centipede in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede's carapace colored with a blend of deep purple and electric blue with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs bright yellow-tipped and gripping the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the mandibles massive, sharp like obsidian sickles, and dripping with sticky, shimmering venom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Japanese onomatopoeia 'ゴゴゴゴ' visible in the background as a massive, rough form with a pulsating halo of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023318354040384293 + ], + "opposite_logprob_tokens": [ + -3.773318290710449 + ], + "logprob_reward": -0.023318354040384293, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023318354040384293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22663894295692444, + 0.14048880338668823, + 0.04898214712738991, + 0.1167960911989212, + 0.04477331042289734, + 0.09245451539754868, + 0.12745797634124756, + 0.03273391351103783, + 0.06513872742652893, + 0.10453556478023529 + ], + "credit_map": [ + 0.019681056961417198, + 0.02041855826973915, + 0.023609865456819534, + 0.02598441392183304, + 0.03615330532193184, + 0.03522840887308121, + 0.029586346819996834, + 0.035976964980363846, + 0.012590625323355198, + 0.013164747506380081, + 0.01564970798790455, + 0.010119978338479996, + 0.0233331099152565, + 0.025540094822645187, + 0.021781306713819504, + 0.018309228122234344, + 0.005396051332354546, + 0.004602937027812004, + 0.005722502246499062, + 0.005974990781396627, + 0.007089980412274599, + 0.004851905163377523, + 0.006297793239355087, + 0.009045989252626896, + 0.01941552944481373, + 0.009096847847104073, + 0.0159351397305727, + 0.010256616398692131, + 0.010342655703425407, + 0.008089544251561165, + 0.013243445195257664, + 0.03041631355881691, + 0.006333441473543644, + 0.0034337283577769995, + 0.005070515442639589, + 0.005090332590043545, + 0.00447437958791852, + 0.005147635005414486, + 0.011115578934550285, + 0.004107697866857052, + 0.01253527496010065, + 0.00858375709503889, + 0.00992869958281517, + 0.013097522780299187, + 0.007042033597826958, + 0.010527702048420906, + 0.019342267885804176, + 0.011397255584597588, + 0.011157400906085968, + 0.010418512858450413, + 0.014922228641808033, + 0.021131422370672226, + 0.007170007564127445, + 0.012570571154356003, + 0.0377453938126564, + 0.012342443689703941, + 0.0026297648437321186, + 0.0032794310245662928, + 0.003117149695754051, + 0.004271690268069506, + 0.0037858549039810896, + 0.007297174539417028, + 0.0046110376715660095, + 0.0037418126594275236, + 0.009570718742907047, + 0.011227196082472801, + 0.005618643946945667, + 0.005531577859073877, + 0.008128533139824867, + 0.005955332424491644, + 0.011019822210073471, + 0.008086904883384705, + 0.015909135341644287, + 0.011879763565957546, + 0.011571642942726612, + 0.010327237658202648, + 0.0163420382887125, + 0.013629093766212463, + 0.009372973814606667, + 0.015503685921430588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.023318354040384293 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with the centipede in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede's carapace colored with a blend of deep purple and electric blue with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs bright yellow-tipped and gripping the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the mandibles massive, sharp like obsidian sickles, and dripping with sticky, shimmering venom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Japanese onomatopoeia 'ゴゴゴゴ' visible in the background as a massive, rough form with a pulsating halo of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903706073760986 + ], + "opposite_logprob_tokens": [ + -2.4640369415283203 + ], + "logprob_reward": -0.08903706073760986, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903706073760986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09797070920467377, + 0.0920044556260109, + 0.18530337512493134, + 0.12146060913801193, + 0.03289724513888359, + 0.040116701275110245, + 0.2882006764411926, + 0.026698116213083267, + 0.04207444190979004, + 0.07327370345592499 + ], + "credit_map": [ + 0.008264858275651932, + 0.011582964099943638, + 0.010404947213828564, + 0.013135835528373718, + 0.01299044955521822, + 0.013204461894929409, + 0.014763860031962395, + 0.01362332422286272, + 0.011350294575095177, + 0.007220923900604248, + 0.013220040127635002, + 0.008826342411339283, + 0.009244252927601337, + 0.01388589758425951, + 0.014589826576411724, + 0.01366686262190342, + 0.05048736184835434, + 0.014321059919893742, + 0.01708300970494747, + 0.013838620856404305, + 0.014222737401723862, + 0.009109963662922382, + 0.01675061509013176, + 0.049489982426166534, + 0.023399628698825836, + 0.008690791204571724, + 0.01420041173696518, + 0.010199595242738724, + 0.01130108255892992, + 0.007242856081575155, + 0.013311021961271763, + 0.033115215599536896, + 0.0038905609399080276, + 0.0023695521522313356, + 0.003957810346037149, + 0.0038906389381736517, + 0.0027749568689614534, + 0.005746501963585615, + 0.005834285635501146, + 0.004432935267686844, + 0.004698366392403841, + 0.005598054267466068, + 0.005117285531014204, + 0.004518405068665743, + 0.0040402538143098354, + 0.00404335418716073, + 0.007673847954720259, + 0.0044271317310631275, + 0.028705507516860962, + 0.022631000727415085, + 0.03846263140439987, + 0.046941906213760376, + 0.01776607520878315, + 0.02465428225696087, + 0.07979246228933334, + 0.029246771708130836, + 0.0023841257207095623, + 0.0021089313086122274, + 0.002708467422053218, + 0.003702486166730523, + 0.002786485245451331, + 0.005216112360358238, + 0.00446769455447793, + 0.0033238076139241457, + 0.005462829023599625, + 0.004219356458634138, + 0.004582940600812435, + 0.003970228135585785, + 0.005027096252888441, + 0.0039268494583666325, + 0.01011048536747694, + 0.004774654749780893, + 0.0110589899122715, + 0.007954910397529602, + 0.008112187497317791, + 0.00805517565459013, + 0.006773747503757477, + 0.005499191116541624, + 0.01045220997184515, + 0.015367276966571808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 3 + ], + "selected_reward": -0.08903706073760986 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with the centipede in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede's carapace colored with a blend of deep purple and electric blue with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs bright yellow-tipped and gripping the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the mandibles massive, sharp like obsidian sickles, and dripping with sticky, shimmering venom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Japanese onomatopoeia 'ゴゴゴゴ' visible in the background as a massive, rough form with a pulsating halo of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031700879335403 + ], + "opposite_logprob_tokens": [ + -2.3503170013427734 + ], + "logprob_reward": -0.10031700879335403, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031700879335403, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19506308436393738, + 0.11447382718324661, + 0.15192000567913055, + 0.1525217443704605, + 0.03179130703210831, + 0.1026710718870163, + 0.08372501283884048, + 0.031013479456305504, + 0.04951375350356102, + 0.08730669319629669 + ], + "credit_map": [ + 0.017704427242279053, + 0.019556988030672073, + 0.021403225138783455, + 0.029852140694856644, + 0.02327532321214676, + 0.02189461886882782, + 0.031686022877693176, + 0.029690338298678398, + 0.01030002161860466, + 0.008922466076910496, + 0.013512075878679752, + 0.007596537005156279, + 0.014346298761665821, + 0.02480439655482769, + 0.02127709984779358, + 0.013714943081140518, + 0.0424681082367897, + 0.009862885810434818, + 0.011800902895629406, + 0.010478087700903416, + 0.008216400630772114, + 0.007101478520780802, + 0.01569831557571888, + 0.04629383236169815, + 0.027538008987903595, + 0.012824255973100662, + 0.016845734789967537, + 0.011798134073615074, + 0.014907893724739552, + 0.00860540196299553, + 0.017186559736728668, + 0.04281575605273247, + 0.004126434214413166, + 0.002776946872472763, + 0.003187670838087797, + 0.004117754753679037, + 0.003501876723021269, + 0.003826760919764638, + 0.00787388440221548, + 0.0023799787741154432, + 0.011671834625303745, + 0.009849009104073048, + 0.013994499109685421, + 0.011957161128520966, + 0.00983571819961071, + 0.019114932045340538, + 0.017090901732444763, + 0.009157017804682255, + 0.009646170772612095, + 0.0072254217229783535, + 0.011539188213646412, + 0.012493596412241459, + 0.007280619349330664, + 0.008640549145638943, + 0.019354436546564102, + 0.007545027416199446, + 0.002772960579022765, + 0.002526308177039027, + 0.0023817885667085648, + 0.004224487114697695, + 0.002518162364140153, + 0.005903743207454681, + 0.006046977825462818, + 0.0046390509232878685, + 0.007317728362977505, + 0.005511614494025707, + 0.004218036308884621, + 0.005013000685721636, + 0.005028272047638893, + 0.003933512605726719, + 0.012355702929198742, + 0.0061358860693871975, + 0.0071890149265527725, + 0.007263421081006527, + 0.009213539771735668, + 0.00942971371114254, + 0.01560957171022892, + 0.011830332688987255, + 0.01242845319211483, + 0.014342637732625008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.10031700879335403 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with the centipede in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede's carapace colored with a blend of deep purple and electric blue with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs bright yellow-tipped and gripping the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the mandibles massive, sharp like obsidian sickles, and dripping with sticky, shimmering venom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Japanese onomatopoeia 'ゴゴゴゴ' visible in the background as a massive, rough form with a pulsating halo of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703903019428253 + ], + "opposite_logprob_tokens": [ + -2.1270389556884766 + ], + "logprob_reward": -0.12703903019428253, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12703903019428253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.171045184135437, + 0.04764930158853531, + 0.11890297383069992, + 0.3725157380104065, + 0.052269745618104935, + 0.09967965632677078, + 0.045651040971279144, + 0.04628961905837059, + 0.02240106090903282, + 0.023595700040459633 + ], + "credit_map": [ + 0.01666007563471794, + 0.015214898623526096, + 0.014671909622848034, + 0.014472538605332375, + 0.04449174180626869, + 0.01863767020404339, + 0.022367944940924644, + 0.02452841028571129, + 0.005124485585838556, + 0.005329164210706949, + 0.006032757461071014, + 0.005210483446717262, + 0.006851749029010534, + 0.005176454316824675, + 0.008836569264531136, + 0.005087637808173895, + 0.01840941421687603, + 0.011479669250547886, + 0.010080850683152676, + 0.00999955739825964, + 0.016042441129684448, + 0.018514428287744522, + 0.017406480386853218, + 0.01697012223303318, + 0.04746415838599205, + 0.03127383813261986, + 0.04221845790743828, + 0.019705459475517273, + 0.07840954512357712, + 0.027590112760663033, + 0.037556201219558716, + 0.08829794824123383, + 0.006018946412950754, + 0.004624160472303629, + 0.00680069625377655, + 0.005280328914523125, + 0.007420967798680067, + 0.005039628595113754, + 0.010782487690448761, + 0.006302531808614731, + 0.015179944224655628, + 0.009416822344064713, + 0.014423588290810585, + 0.01087134052067995, + 0.011764085851609707, + 0.015117181465029716, + 0.014265185222029686, + 0.008641509339213371, + 0.0051726894453167915, + 0.004592372570186853, + 0.00760185020044446, + 0.004497678484767675, + 0.003989480901509523, + 0.006393848452717066, + 0.005174022167921066, + 0.008229098282754421, + 0.004472521599382162, + 0.003632806707173586, + 0.0044416217133402824, + 0.004432501271367073, + 0.004421467427164316, + 0.01095667015761137, + 0.007054776884615421, + 0.0068772523663938046, + 0.0026588398031890392, + 0.0029199167620390654, + 0.0027314615435898304, + 0.003085289616137743, + 0.0025528341066092253, + 0.0029925715643912554, + 0.0030407928861677647, + 0.002419355558231473, + 0.0035342886112630367, + 0.002594289369881153, + 0.002780197188258171, + 0.002979226876050234, + 0.001980522647500038, + 0.004261189606040716, + 0.002266457537189126, + 0.003199527272954583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.12703903019428253 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_3/index_tensor([2841, 2841])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_3/index_tensor([2841, 2841])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..81a7329cc7de0ef0aec061d127ec3e26c6fac5ea --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_3/index_tensor([2841, 2841])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "cinematic close-up slowly zooms in, capturing a magnificent parrot gracefully perched on a dense, gnarled jungle vine. The parrot's plumage is a breathtaking, surreal mosaic of color: its head and neck a deep sapphire blue, its breast a vibrant sun yellow, its wings a rich emerald green, each feather rendered in crisp detail. Its intelligent black eyes gleam, and its rapid, quirky turns of the head cause a subtle rustling of feathers. The vine it clings to is gnarled and ancient, its brown, richly textured bark clinging to a mat of damp green moss. The background is a vibrant, soft-focus blur of lush, dewy foliage, with large, shimmering tropical leaves. Dramatic, warm sunlight filters down through the dense canopy, penetrating the humid air, illuminating floating particles and creating a dynamic play of light and shadow across the scene. The overall atmosphere is vibrant and energized, a vibrant and dynamic snapshot of the pulsating heart of the rainforest.", + "seed": 6714, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_3/index_2841_seed_6714_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05417161183431744, + "vlm_logprob": -0.05417161183431744, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.05417161183431744, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic close-up of a parrot perched on a vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the parrot's head and neck a deep sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the parrot's breast a vibrant sun yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parrot's wings a rich emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017049528658390045 + ], + "opposite_logprob_tokens": [ + -4.517049312591553 + ], + "logprob_reward": -0.017049528658390045, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.017049528658390045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08718616515398026, + 0.027966145426034927, + 0.027343258261680603, + 0.052111607044935226, + 0.22744834423065186, + 0.1460466831922531, + 0.10802360624074936, + 0.1398727297782898, + 0.07463312149047852, + 0.10936831682920456 + ], + "credit_map": [ + 0.011729616671800613, + 0.009202638640999794, + 0.010392139665782452, + 0.010318021290004253, + 0.010974766686558723, + 0.011688493192195892, + 0.01229088380932808, + 0.01058960985392332, + 0.0038783166091889143, + 0.002904633292928338, + 0.003688737051561475, + 0.0035197329707443714, + 0.005081478971987963, + 0.0027134798001497984, + 0.0031751247588545084, + 0.0030046438332647085, + 0.002219229470938444, + 0.004907399881631136, + 0.004802416078746319, + 0.004731133580207825, + 0.002928343368694186, + 0.002271510660648346, + 0.0028004071209579706, + 0.002682818565517664, + 0.006514179520308971, + 0.007265442982316017, + 0.005871863104403019, + 0.0065514883026480675, + 0.007006923668086529, + 0.0050217448733747005, + 0.006728836335241795, + 0.00715113477781415, + 0.019312506541609764, + 0.009760439395904541, + 0.08015617728233337, + 0.014686549082398415, + 0.010393114760518074, + 0.0741475373506546, + 0.009060872718691826, + 0.009931167587637901, + 0.008980468846857548, + 0.006503021344542503, + 0.009322365745902061, + 0.004932475741952658, + 0.0556325763463974, + 0.04467550292611122, + 0.005205011926591396, + 0.010795271024107933, + 0.009515931829810143, + 0.007123814895749092, + 0.011193769983947277, + 0.007526860106736422, + 0.028356028720736504, + 0.018189694732427597, + 0.009339144453406334, + 0.016778362914919853, + 0.020712800323963165, + 0.012357918545603752, + 0.029744522646069527, + 0.018075518310070038, + 0.01689896173775196, + 0.008612261153757572, + 0.013819453306496143, + 0.019651297479867935, + 0.007027226034551859, + 0.008761526085436344, + 0.007650688756257296, + 0.00587487081065774, + 0.021530359983444214, + 0.012401417829096317, + 0.004892708268016577, + 0.006494330242276192, + 0.01500178687274456, + 0.01422975119203329, + 0.01555210817605257, + 0.014446046203374863, + 0.013225717470049858, + 0.006652183830738068, + 0.01343897171318531, + 0.01682175137102604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.017049528658390045 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic close-up of a parrot perched on a vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the parrot's head and neck a deep sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the parrot's breast a vibrant sun yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parrot's wings a rich emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16036272048950195 + ], + "opposite_logprob_tokens": [ + -1.910362720489502 + ], + "logprob_reward": -0.16036272048950195, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16036272048950195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0710168108344078, + 0.02361610345542431, + 0.030505886301398277, + 0.06600065529346466, + 0.05463170260190964, + 0.22182753682136536, + 0.07437478750944138, + 0.046059027314186096, + 0.03153572231531143, + 0.38043177127838135 + ], + "credit_map": [ + 0.010783007368445396, + 0.010503120720386505, + 0.007577456999570131, + 0.010020353831350803, + 0.009155510924756527, + 0.008809586986899376, + 0.006947603076696396, + 0.007220177911221981, + 0.0024654639419168234, + 0.004600043408572674, + 0.003475737990811467, + 0.003404970047995448, + 0.0030922202859073877, + 0.0023616678081452847, + 0.0023457477800548077, + 0.0018702541710808873, + 0.0030090759973973036, + 0.004449155647307634, + 0.0041298773139715195, + 0.005067471880465746, + 0.004267279524356127, + 0.003545785555616021, + 0.003390108933672309, + 0.0026471312157809734, + 0.006167948246002197, + 0.011345536448061466, + 0.00623624213039875, + 0.012441324070096016, + 0.011827233247458935, + 0.005134277045726776, + 0.005700834561139345, + 0.007147267926484346, + 0.004333107266575098, + 0.006966143846511841, + 0.007462460547685623, + 0.005885721184313297, + 0.008259573951363564, + 0.009332314133644104, + 0.007061624899506569, + 0.005330762360244989, + 0.013172874227166176, + 0.013282873667776585, + 0.013925186358392239, + 0.009147058241069317, + 0.08393391221761703, + 0.06650687009096146, + 0.008770709857344627, + 0.013088052161037922, + 0.007279008161276579, + 0.008088178001344204, + 0.005292731802910566, + 0.007065061945468187, + 0.021560417488217354, + 0.014630342833697796, + 0.0049994755536317825, + 0.005459584761410952, + 0.005125676281750202, + 0.00697846245020628, + 0.007176114246249199, + 0.006689026020467281, + 0.006914670113474131, + 0.0037174089811742306, + 0.004643050022423267, + 0.0048146238550543785, + 0.003843157086521387, + 0.004893295932561159, + 0.004583955276757479, + 0.004566957708448172, + 0.00339902238920331, + 0.0026774343568831682, + 0.003631369210779667, + 0.0039405361749231815, + 0.05505238100886345, + 0.04975443333387375, + 0.05183912441134453, + 0.051625072956085205, + 0.04150259867310524, + 0.022257614880800247, + 0.04766317829489708, + 0.06073739007115364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.16036272048950195 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic close-up of a parrot perched on a vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the parrot's head and neck a deep sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the parrot's breast a vibrant sun yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parrot's wings a rich emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005256047938019037 + ], + "opposite_logprob_tokens": [ + -5.255256175994873 + ], + "logprob_reward": -0.005256047938019037, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005256047938019037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07492804527282715, + 0.060819875448942184, + 0.054608359932899475, + 0.11348053067922592, + 0.1470901221036911, + 0.12390530854463577, + 0.07219655811786652, + 0.17934593558311462, + 0.07133549451828003, + 0.10228978097438812 + ], + "credit_map": [ + 0.009207290597259998, + 0.012403298169374466, + 0.010002796538174152, + 0.011933384463191032, + 0.008335324004292488, + 0.008609236218035221, + 0.005929001607000828, + 0.008507718332111835, + 0.0060916258953511715, + 0.014410682022571564, + 0.007059419061988592, + 0.008667398244142532, + 0.008550357073545456, + 0.0057363370433449745, + 0.005755098070949316, + 0.004548956174403429, + 0.0044844853691756725, + 0.011110282503068447, + 0.008347530849277973, + 0.008551976643502712, + 0.00735184783115983, + 0.005473353434354067, + 0.004209705162793398, + 0.0050791786052286625, + 0.009868379682302475, + 0.024010570719838142, + 0.009879428893327713, + 0.019028816372156143, + 0.015259425155818462, + 0.011625774204730988, + 0.015391902066767216, + 0.008416234515607357, + 0.011937622912228107, + 0.013997928239405155, + 0.038122598081827164, + 0.014358780346810818, + 0.011356422677636147, + 0.03884592652320862, + 0.009448982775211334, + 0.009021857753396034, + 0.007953465916216373, + 0.005677275825291872, + 0.008247430436313152, + 0.005077415145933628, + 0.047093238681554794, + 0.03595655411481857, + 0.005610170308500528, + 0.008289755322039127, + 0.010570873506367207, + 0.007020452991127968, + 0.008804220706224442, + 0.0064378343522548676, + 0.008732728660106659, + 0.0060273148119449615, + 0.007801102474331856, + 0.01680203154683113, + 0.023793602362275124, + 0.021180829033255577, + 0.023243948817253113, + 0.028386959806084633, + 0.02288532257080078, + 0.01687454991042614, + 0.01984848454594612, + 0.023132242262363434, + 0.011052882298827171, + 0.010744886472821236, + 0.010686943307518959, + 0.010593890212476254, + 0.00854562595486641, + 0.005503035616129637, + 0.007522637490183115, + 0.006685595493763685, + 0.014380122534930706, + 0.010523844510316849, + 0.016120685264468193, + 0.011290650814771652, + 0.010955576784908772, + 0.006258626468479633, + 0.01384806726127863, + 0.01891220733523369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 5 + ], + "selected_reward": -0.005256047938019037 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic close-up of a parrot perched on a vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the parrot's head and neck a deep sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the parrot's breast a vibrant sun yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parrot's wings a rich emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018178056925535202 + ], + "opposite_logprob_tokens": [ + -4.0181779861450195 + ], + "logprob_reward": -0.018178056925535202, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018178056925535202, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03135678172111511, + 0.030308077111840248, + 0.02150432951748371, + 0.04351363703608513, + 0.508061408996582, + 0.1362326592206955, + 0.03101346828043461, + 0.07386485487222672, + 0.0428948849439621, + 0.08124992996454239 + ], + "credit_map": [ + 0.004567598924040794, + 0.004204913042485714, + 0.003507474670186639, + 0.005719783250242472, + 0.003219485515728593, + 0.0030810311436653137, + 0.003481541993096471, + 0.0035749529488384724, + 0.0024694588501006365, + 0.004560223780572414, + 0.0043982844799757, + 0.003019695868715644, + 0.005775593686848879, + 0.0032270660158246756, + 0.004545577336102724, + 0.0023121777921915054, + 0.0024466661270707846, + 0.00278132944367826, + 0.002930303802713752, + 0.0037814381066709757, + 0.002174344379454851, + 0.0025943114887923002, + 0.001894774497486651, + 0.00290116248652339, + 0.005132821388542652, + 0.006048140581697226, + 0.005351708736270666, + 0.00728087080642581, + 0.004845384508371353, + 0.0038318480364978313, + 0.007087287958711386, + 0.003935575485229492, + 0.04408629238605499, + 0.02186453528702259, + 0.18303735554218292, + 0.03405675292015076, + 0.020244862884283066, + 0.16472183167934418, + 0.01738150231540203, + 0.022668272256851196, + 0.008706752210855484, + 0.00519914086908102, + 0.00923109333962202, + 0.0041065216064453125, + 0.05282432958483696, + 0.041286349296569824, + 0.005052095744758844, + 0.009826370514929295, + 0.0033946505282074213, + 0.00410043727606535, + 0.0035575474612414837, + 0.00392926437780261, + 0.003453959012404084, + 0.0032019270583987236, + 0.004026543349027634, + 0.005349142011255026, + 0.009707382880151272, + 0.009873814880847931, + 0.010629863478243351, + 0.012091086246073246, + 0.007425688207149506, + 0.004920950625091791, + 0.009044031612575054, + 0.010172038339078426, + 0.004019460640847683, + 0.005624965764582157, + 0.003666252363473177, + 0.0038279977161437273, + 0.012262795120477676, + 0.00743868900462985, + 0.0027002887800335884, + 0.0033544343896210194, + 0.014072313904762268, + 0.009545677341520786, + 0.01567680761218071, + 0.00984211079776287, + 0.006841565482318401, + 0.005862011108547449, + 0.008482273668050766, + 0.010927168652415276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 9 + ], + "selected_reward": -0.018178056925535202 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic close-up of a parrot perched on a vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the parrot's head and neck a deep sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the parrot's breast a vibrant sun yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parrot's wings a rich emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001170516014099 + ], + "opposite_logprob_tokens": [ + -2.695011615753174 + ], + "logprob_reward": -0.07001170516014099, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001170516014099, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.035943854600191116, + 0.03144882991909981, + 0.031487930566072464, + 0.03825009986758232, + 0.2666642367839813, + 0.038014404475688934, + 0.05339757725596428, + 0.260013610124588, + 0.08353367447853088, + 0.16124583780765533 + ], + "credit_map": [ + 0.004898164886981249, + 0.005615512374788523, + 0.004350780509412289, + 0.006604959722608328, + 0.003775437595322728, + 0.0032240687869489193, + 0.0031404306646436453, + 0.004334498196840286, + 0.0026342535857111216, + 0.00569012900814414, + 0.004101430997252464, + 0.004071438685059547, + 0.004716896452009678, + 0.003274446353316307, + 0.004451697692275047, + 0.0025085355155169964, + 0.0023341800551861525, + 0.0046593076549470425, + 0.006258031353354454, + 0.007058050483465195, + 0.0027213934808969498, + 0.0030154313426464796, + 0.0025983364321291447, + 0.0028431997634470463, + 0.0054169208742678165, + 0.007007298059761524, + 0.003982932772487402, + 0.0057344879023730755, + 0.005699491128325462, + 0.002993847243487835, + 0.0033434892538934946, + 0.004071628209203482, + 0.02290777675807476, + 0.011805982328951359, + 0.09554813802242279, + 0.018605323508381844, + 0.012682602740824223, + 0.08466114848852158, + 0.009038056246936321, + 0.01141518633812666, + 0.0030765188857913017, + 0.004611883778125048, + 0.0023468006402254105, + 0.0022190758027136326, + 0.012165690772235394, + 0.0084513695910573, + 0.002071601804345846, + 0.0030714578460901976, + 0.008242671377956867, + 0.00846156757324934, + 0.005840437952429056, + 0.006867602467536926, + 0.005095769185572863, + 0.0032694702968001366, + 0.004587185103446245, + 0.011032870970666409, + 0.031434476375579834, + 0.031860366463661194, + 0.029450412839651108, + 0.03866245597600937, + 0.041600119322538376, + 0.0259742122143507, + 0.026939351111650467, + 0.03409220278263092, + 0.011410011909902096, + 0.011006229557096958, + 0.012255254201591015, + 0.012267252430319786, + 0.011619683355093002, + 0.009250479750335217, + 0.008051134645938873, + 0.007673623505979776, + 0.023895680904388428, + 0.01766922138631344, + 0.025788580998778343, + 0.019307633861899376, + 0.018956879153847694, + 0.010333895683288574, + 0.015864618122577667, + 0.029429316520690918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 9 + ], + "selected_reward": -0.07001170516014099 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic close-up slowly zooms in, capturing a magnificent parrot gracefully perched on a dense, gnarled jungle vine. The parrot's plumage is a breathtaking, surreal mosaic of color: its head and neck a deep sapphire blue, its breast a vibrant sun yellow, its wings a rich emerald green, each feather rendered in crisp detail. Its intelligent black eyes gleam, and its rapid, quirky turns of the head cause a subtle rustling of feathers. The vine it clings to is gnarled and ancient, its brown, richly textured bark clinging to a mat of damp green moss. The background is a vibrant, soft-focus blur of lush, dewy foliage, with large, shimmering tropical leaves. Dramatic, warm sunlight filters down through the dense canopy, penetrating the humid air, illuminating floating particles and creating a dynamic play of light and shadow across the scene. The overall atmosphere is vibrant and energized, a vibrant and dynamic snapshot of the pulsating heart of the rainforest.", + "seed": 6715, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_3/index_2841_seed_6715_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.10368521120399236, + "vlm_logprob": -0.10368521120399236, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.10368521120399236, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic close-up of a parrot perched on a vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the parrot's head and neck a deep sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the parrot's breast a vibrant sun yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parrot's wings a rich emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024686425924301147 + ], + "opposite_logprob_tokens": [ + -3.899686336517334 + ], + "logprob_reward": -0.024686425924301147, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.024686425924301147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11076074838638306, + 0.06696809083223343, + 0.04764222353696823, + 0.1291518211364746, + 0.02988603338599205, + 0.08035767078399658, + 0.19328300654888153, + 0.15327095985412598, + 0.1575346440076828, + 0.03114481456577778 + ], + "credit_map": [ + 0.010744263418018818, + 0.020220905542373657, + 0.01018125843256712, + 0.012259313836693764, + 0.012496137991547585, + 0.00786313135176897, + 0.02303478866815567, + 0.01396095659583807, + 0.006669096183031797, + 0.020203683525323868, + 0.005551012232899666, + 0.0071844663470983505, + 0.006844630464911461, + 0.00400901585817337, + 0.00749430526047945, + 0.009011883288621902, + 0.005180623382329941, + 0.009782468900084496, + 0.005423842929303646, + 0.008159803226590157, + 0.0035302042961120605, + 0.0031947435345500708, + 0.0071502807550132275, + 0.0052202558144927025, + 0.010955984704196453, + 0.017679363489151, + 0.017496854066848755, + 0.03502832353115082, + 0.009036835283041, + 0.007212326396256685, + 0.017348047345876694, + 0.014394084922969341, + 0.002620234154164791, + 0.0033198855817317963, + 0.0027531252708286047, + 0.004047533962875605, + 0.003779242979362607, + 0.004127694759517908, + 0.004895832855254412, + 0.004342482890933752, + 0.007109588012099266, + 0.010223033837974072, + 0.009621058590710163, + 0.01394900493323803, + 0.01050981879234314, + 0.009986556135118008, + 0.008652936667203903, + 0.010305671952664852, + 0.02613963559269905, + 0.0233976598829031, + 0.02361265942454338, + 0.0385090634226799, + 0.022798992693424225, + 0.014930391684174538, + 0.023898180574178696, + 0.019996412098407745, + 0.019089289009571075, + 0.015501735731959343, + 0.015012992545962334, + 0.02647036500275135, + 0.0201248899102211, + 0.012582327239215374, + 0.019352789968252182, + 0.02513657882809639, + 0.011315703392028809, + 0.027406679466366768, + 0.014420104213058949, + 0.022130465134978294, + 0.01593826152384281, + 0.01992226578295231, + 0.025427669286727905, + 0.02097349800169468, + 0.003199464175850153, + 0.003431211691349745, + 0.005110724363476038, + 0.0033740163780748844, + 0.002986184787005186, + 0.0048196325078606606, + 0.004510577768087387, + 0.0037130010314285755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.024686425924301147 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic close-up of a parrot perched on a vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the parrot's head and neck a deep sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the parrot's breast a vibrant sun yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parrot's wings a rich emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47429031133651733 + ], + "opposite_logprob_tokens": [ + -0.9742903113365173 + ], + "logprob_reward": -0.47429031133651733, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47429031133651733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1416654884815216, + 0.08688206225633621, + 0.028180817142128944, + 0.049761492758989334, + 0.07495369762182236, + 0.04691358283162117, + 0.14771956205368042, + 0.1840362250804901, + 0.20395825803279877, + 0.03592881187796593 + ], + "credit_map": [ + 0.016828063875436783, + 0.02305825799703598, + 0.014325067400932312, + 0.012837505899369717, + 0.02077791653573513, + 0.01319549698382616, + 0.022955670952796936, + 0.017687499523162842, + 0.007565983571112156, + 0.02450459636747837, + 0.007432702928781509, + 0.01012447476387024, + 0.008837953209877014, + 0.007753604091703892, + 0.010183000937104225, + 0.010479745455086231, + 0.002939818426966667, + 0.0040738401003181934, + 0.0034827603958547115, + 0.005042831879109144, + 0.003074542386457324, + 0.002502213465049863, + 0.0041049993596971035, + 0.0029598078690469265, + 0.00448443740606308, + 0.007332762237638235, + 0.005845129489898682, + 0.01487946417182684, + 0.0034745975863188505, + 0.0026989446487277746, + 0.005183875095099211, + 0.005862282123416662, + 0.007043198682367802, + 0.011463556438684464, + 0.008558203466236591, + 0.02405085787177086, + 0.005372604355216026, + 0.003687528660520911, + 0.00588661152869463, + 0.008891135454177856, + 0.005689695011824369, + 0.007995987311005592, + 0.006800060160458088, + 0.006676125805824995, + 0.005389675963670015, + 0.004149045329540968, + 0.005721862427890301, + 0.00449112756177783, + 0.019779345020651817, + 0.016882142052054405, + 0.018961995840072632, + 0.03354660049080849, + 0.016355492174625397, + 0.010971535928547382, + 0.016147322952747345, + 0.015075130388140678, + 0.02360936999320984, + 0.021205734461545944, + 0.017993489280343056, + 0.03462855517864227, + 0.022345811128616333, + 0.011817652732133865, + 0.023339364677667618, + 0.02909623645246029, + 0.022267308086156845, + 0.02881377749145031, + 0.02958120033144951, + 0.02995310164988041, + 0.02176859974861145, + 0.014559829607605934, + 0.024198444560170174, + 0.03281598910689354, + 0.0031280298717319965, + 0.00469351839274168, + 0.005665986333042383, + 0.0056428187526762486, + 0.0035123417619615793, + 0.003379930043593049, + 0.00392415514215827, + 0.005982028786092997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.47429031133651733 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic close-up of a parrot perched on a vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the parrot's head and neck a deep sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the parrot's breast a vibrant sun yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parrot's wings a rich emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002824010793119669 + ], + "opposite_logprob_tokens": [ + -5.877823829650879 + ], + "logprob_reward": -0.002824010793119669, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002824010793119669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07097356766462326, + 0.07181213796138763, + 0.04716777428984642, + 0.09044189751148224, + 0.060092464089393616, + 0.11186585575342178, + 0.17673896253108978, + 0.13483211398124695, + 0.19557735323905945, + 0.04049788415431976 + ], + "credit_map": [ + 0.007708852645009756, + 0.011574242264032364, + 0.007186216302216053, + 0.009571443311870098, + 0.009107345715165138, + 0.0057546463795006275, + 0.010814973153173923, + 0.009255856275558472, + 0.00336730619892478, + 0.020437754690647125, + 0.003964316565543413, + 0.009973260574042797, + 0.009231777861714363, + 0.005409540608525276, + 0.01048420462757349, + 0.00894397683441639, + 0.0034319255501031876, + 0.006287428084760904, + 0.0040564993396401405, + 0.012124677188694477, + 0.004457516595721245, + 0.006659611128270626, + 0.005222187843173742, + 0.0049279313534498215, + 0.009825400076806545, + 0.009620297700166702, + 0.01493012998253107, + 0.026143943890929222, + 0.006121295969933271, + 0.004456477705389261, + 0.010180504992604256, + 0.009163856506347656, + 0.0062821838073432446, + 0.01049126498401165, + 0.006378487683832645, + 0.009559080936014652, + 0.0053484514355659485, + 0.0048929667100310326, + 0.011341551318764687, + 0.005798482336103916, + 0.016192203387618065, + 0.012449842877686024, + 0.015327260829508305, + 0.016204895451664925, + 0.01262142974883318, + 0.012652029283344746, + 0.013166765682399273, + 0.013251439668238163, + 0.02321293205022812, + 0.017598679289221764, + 0.029820658266544342, + 0.032695967704057693, + 0.018653858453035355, + 0.012871546670794487, + 0.020662952214479446, + 0.021222369745373726, + 0.017473680898547173, + 0.012976483441889286, + 0.012241302989423275, + 0.02450193092226982, + 0.016892505809664726, + 0.010770739987492561, + 0.01856677420437336, + 0.02140870876610279, + 0.01993551105260849, + 0.023807041347026825, + 0.02766242064535618, + 0.02794763259589672, + 0.021665003150701523, + 0.020665820688009262, + 0.023892931640148163, + 0.03000100515782833, + 0.00444478914141655, + 0.0062268623150885105, + 0.0064500318840146065, + 0.00423665763810277, + 0.0035909933503717184, + 0.00653855362907052, + 0.004068647976964712, + 0.00494134658947587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.002824010793119669 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic close-up of a parrot perched on a vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the parrot's head and neck a deep sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the parrot's breast a vibrant sun yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parrot's wings a rich emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012531350366771221 + ], + "opposite_logprob_tokens": [ + -4.387531280517578 + ], + "logprob_reward": -0.012531350366771221, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012531350366771221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07596148550510406, + 0.04511880502104759, + 0.04535992071032524, + 0.14280575513839722, + 0.28431436419487, + 0.06288642436265945, + 0.18338584899902344, + 0.06821900606155396, + 0.06410889327526093, + 0.027839478105306625 + ], + "credit_map": [ + 0.008052191697061062, + 0.010136262513697147, + 0.006874237209558487, + 0.012680952437222004, + 0.008834421634674072, + 0.006318960338830948, + 0.012224432080984116, + 0.010840027593076229, + 0.003441356820985675, + 0.007907955907285213, + 0.004639767110347748, + 0.00816405564546585, + 0.0048956791870296, + 0.003659036010503769, + 0.0070241037756204605, + 0.00538684893399477, + 0.006107877939939499, + 0.008268815465271473, + 0.007105051539838314, + 0.005676380358636379, + 0.0037505857180804014, + 0.003840383607894182, + 0.005687455181032419, + 0.004923371132463217, + 0.018496306613087654, + 0.02330625057220459, + 0.02452515996992588, + 0.0327472910284996, + 0.007488158997148275, + 0.006432914640754461, + 0.013573826290667057, + 0.016235850751399994, + 0.02928748168051243, + 0.024776814505457878, + 0.034776825457811356, + 0.03763199970126152, + 0.03392970561981201, + 0.03649768978357315, + 0.04421398788690567, + 0.04319986328482628, + 0.009141897782683372, + 0.008987180888652802, + 0.010056788101792336, + 0.010972087271511555, + 0.006196597591042519, + 0.003774292767047882, + 0.008508156053721905, + 0.0052494267001748085, + 0.019775385037064552, + 0.01991841569542885, + 0.027632176876068115, + 0.04172495752573013, + 0.020971376448869705, + 0.012720414437353611, + 0.01978779025375843, + 0.02085532620549202, + 0.007731543853878975, + 0.005855115596204996, + 0.006359825376421213, + 0.01252830121666193, + 0.008967400528490543, + 0.00571823213249445, + 0.010561040602624416, + 0.01049755048006773, + 0.006219826638698578, + 0.012022117152810097, + 0.009354392066597939, + 0.011201431043446064, + 0.005956794135272503, + 0.003813211340457201, + 0.005625268444418907, + 0.009915852919220924, + 0.0030839431565254927, + 0.002914965618401766, + 0.005177795421332121, + 0.004704684019088745, + 0.0022232725750654936, + 0.0018018238479271531, + 0.004022604785859585, + 0.003910387400537729 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 3 + ], + "selected_reward": -0.012531350366771221 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic close-up of a parrot perched on a vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the parrot's head and neck a deep sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the parrot's breast a vibrant sun yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parrot's wings a rich emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004093957599252462 + ], + "opposite_logprob_tokens": [ + -5.504094123840332 + ], + "logprob_reward": -0.004093957599252462, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004093957599252462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08754459023475647, + 0.0643744021654129, + 0.06177924945950508, + 0.09252893924713135, + 0.11969062685966492, + 0.05997280403971672, + 0.11150230467319489, + 0.16050846874713898, + 0.21777579188346863, + 0.024322811514139175 + ], + "credit_map": [ + 0.01462737750262022, + 0.011201203800737858, + 0.015550785697996616, + 0.011140983551740646, + 0.008659780956804752, + 0.007356238551437855, + 0.009141805581748486, + 0.009866412729024887, + 0.004855838604271412, + 0.01623113639652729, + 0.006083239801228046, + 0.010645967908203602, + 0.006949717178940773, + 0.004273686092346907, + 0.007700151763856411, + 0.0076346672140061855, + 0.006623944733291864, + 0.009086553007364273, + 0.007304136175662279, + 0.011138711124658585, + 0.004857385065406561, + 0.005454982165247202, + 0.010040387511253357, + 0.007273152936249971, + 0.008413395844399929, + 0.012940103188157082, + 0.011075137183070183, + 0.01987818442285061, + 0.008851694874465466, + 0.006168186664581299, + 0.013963074423372746, + 0.011239156126976013, + 0.0082778325304389, + 0.010049106553196907, + 0.016545800492167473, + 0.021779296919703484, + 0.015232665464282036, + 0.012672081589698792, + 0.021252598613500595, + 0.013881251215934753, + 0.006852526217699051, + 0.010566268116235733, + 0.00701518077403307, + 0.01177372969686985, + 0.0065026856027543545, + 0.0051122913137078285, + 0.005897499620914459, + 0.006252621300518513, + 0.014451134018599987, + 0.010319541208446026, + 0.015877271071076393, + 0.021562697365880013, + 0.013718071393668652, + 0.010068456642329693, + 0.012353111989796162, + 0.013152028433978558, + 0.022067023441195488, + 0.018669558688998222, + 0.01622561365365982, + 0.031749434769153595, + 0.016674503684043884, + 0.010067174211144447, + 0.019373398274183273, + 0.02568175457417965, + 0.023120321333408356, + 0.02983090840280056, + 0.031034179031848907, + 0.03264433890581131, + 0.02347576804459095, + 0.017737427726387978, + 0.025404509156942368, + 0.034528348594903946, + 0.002035332378000021, + 0.003272255416959524, + 0.005145857110619545, + 0.004601833410561085, + 0.0019558293279260397, + 0.0013557100901380181, + 0.002531161066144705, + 0.003424832597374916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 4 + ], + "selected_reward": -0.004093957599252462 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic close-up slowly zooms in, capturing a magnificent parrot gracefully perched on a dense, gnarled jungle vine. The parrot's plumage is a breathtaking, surreal mosaic of color: its head and neck a deep sapphire blue, its breast a vibrant sun yellow, its wings a rich emerald green, each feather rendered in crisp detail. Its intelligent black eyes gleam, and its rapid, quirky turns of the head cause a subtle rustling of feathers. The vine it clings to is gnarled and ancient, its brown, richly textured bark clinging to a mat of damp green moss. The background is a vibrant, soft-focus blur of lush, dewy foliage, with large, shimmering tropical leaves. Dramatic, warm sunlight filters down through the dense canopy, penetrating the humid air, illuminating floating particles and creating a dynamic play of light and shadow across the scene. The overall atmosphere is vibrant and energized, a vibrant and dynamic snapshot of the pulsating heart of the rainforest.", + "seed": 6716, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_3/index_2841_seed_6716_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08570848349481822, + "vlm_logprob": -0.08570848349481822, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.08570848349481822, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic close-up of a parrot perched on a vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the parrot's head and neck a deep sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the parrot's breast a vibrant sun yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parrot's wings a rich emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023460427299141884 + ], + "opposite_logprob_tokens": [ + -4.023460388183594 + ], + "logprob_reward": -0.023460427299141884, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.023460427299141884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1096988245844841, + 0.0304185152053833, + 0.043311405926942825, + 0.15101519227027893, + 0.037840552628040314, + 0.10810527205467224, + 0.08832178264856339, + 0.24666672945022583, + 0.10856883227825165, + 0.07605286687612534 + ], + "credit_map": [ + 0.011677371338009834, + 0.012188184075057507, + 0.01432251837104559, + 0.015833226963877678, + 0.014375807717442513, + 0.012633133679628372, + 0.014450914226472378, + 0.014217670075595379, + 0.002491318853572011, + 0.002823880175128579, + 0.005989676341414452, + 0.004476859234273434, + 0.002723480574786663, + 0.004057447891682386, + 0.003871393855661154, + 0.003984458744525909, + 0.003965078387409449, + 0.003855098970234394, + 0.006244610995054245, + 0.005867115221917629, + 0.005069654900580645, + 0.0055342731066048145, + 0.0071234446950256824, + 0.005652130581438541, + 0.015210499987006187, + 0.011631880886852741, + 0.04754382744431496, + 0.012394120916724205, + 0.013931003399193287, + 0.010122830979526043, + 0.011108734644949436, + 0.029072297737002373, + 0.004232834558933973, + 0.0036988502833992243, + 0.00618391577154398, + 0.002856723265722394, + 0.005044950637966394, + 0.004956664517521858, + 0.004253280349075794, + 0.0066133346408605576, + 0.009471188299357891, + 0.012070235796272755, + 0.022376999258995056, + 0.014459380879998207, + 0.013728966005146503, + 0.015406672842800617, + 0.01074896939098835, + 0.009842856787145138, + 0.00555171025916934, + 0.005138318054378033, + 0.010557159781455994, + 0.005464634858071804, + 0.007822924293577671, + 0.021516069769859314, + 0.005802890285849571, + 0.026468081399798393, + 0.018782416358590126, + 0.01106945425271988, + 0.01179088931530714, + 0.014809147454798222, + 0.07230827957391739, + 0.08626391738653183, + 0.013750326819717884, + 0.017892304807901382, + 0.009551934897899628, + 0.005723482929170132, + 0.005630686413496733, + 0.006598224397748709, + 0.03428203612565994, + 0.031431011855602264, + 0.0062238541431725025, + 0.009127608500421047, + 0.003505818545818329, + 0.0044799065217375755, + 0.009722902439534664, + 0.0027509864885360003, + 0.006452742964029312, + 0.017379099503159523, + 0.0033653995487838984, + 0.028396010398864746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.023460427299141884 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic close-up of a parrot perched on a vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the parrot's head and neck a deep sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the parrot's breast a vibrant sun yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parrot's wings a rich emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.281332790851593 + ], + "opposite_logprob_tokens": [ + -1.4063327312469482 + ], + "logprob_reward": -0.281332790851593, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.281332790851593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06267803907394409, + 0.024390321224927902, + 0.03215675801038742, + 0.08233243227005005, + 0.04354526475071907, + 0.05047629401087761, + 0.048566464334726334, + 0.2874625325202942, + 0.34063422679901123, + 0.02775765210390091 + ], + "credit_map": [ + 0.009239626117050648, + 0.008917702361941338, + 0.009114604443311691, + 0.007373669184744358, + 0.008411332964897156, + 0.005063075106590986, + 0.007581367623060942, + 0.006976666394621134, + 0.001518816570751369, + 0.0027316787745803595, + 0.003478046040982008, + 0.003311265492811799, + 0.0023456253111362457, + 0.002476384863257408, + 0.0038551597390323877, + 0.004673345480114222, + 0.003454176476225257, + 0.0037149142008274794, + 0.004526238888502121, + 0.004838604014366865, + 0.003701139008626342, + 0.0026414443273097277, + 0.004105981905013323, + 0.005174263380467892, + 0.008419179357588291, + 0.004691827576607466, + 0.028687098994851112, + 0.006362559739500284, + 0.008228139020502567, + 0.004951938986778259, + 0.0065897488966584206, + 0.014401945285499096, + 0.004018764477223158, + 0.004937794525176287, + 0.005306121427565813, + 0.005184749606996775, + 0.008398599922657013, + 0.003991414792835712, + 0.007103040348738432, + 0.004604785703122616, + 0.005982999224215746, + 0.007239805534482002, + 0.004457778763025999, + 0.004736659117043018, + 0.011898305267095566, + 0.005608428735285997, + 0.005977816414088011, + 0.00457450607791543, + 0.002671498106792569, + 0.0032011575531214476, + 0.002503893803805113, + 0.003847816027700901, + 0.0039102863520383835, + 0.014603127725422382, + 0.0027331330347806215, + 0.015095550566911697, + 0.019048558548092842, + 0.011985337361693382, + 0.01418768148869276, + 0.014981427229940891, + 0.08895568549633026, + 0.10300090163946152, + 0.016676995903253555, + 0.01862596906721592, + 0.016531232744455338, + 0.013748157769441605, + 0.015417232178151608, + 0.009141096845269203, + 0.1358329802751541, + 0.11103605479001999, + 0.01160904299467802, + 0.02731843665242195, + 0.0018160013714805245, + 0.0017528256867080927, + 0.00801975466310978, + 0.0018249739659950137, + 0.0033491291105747223, + 0.00200617965310812, + 0.0018531247042119503, + 0.007135664578527212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -0.281332790851593 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic close-up of a parrot perched on a vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the parrot's head and neck a deep sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the parrot's breast a vibrant sun yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parrot's wings a rich emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080636642873287 + ], + "opposite_logprob_tokens": [ + -4.511080741882324 + ], + "logprob_reward": -0.011080636642873287, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080636642873287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10037266463041306, + 0.04750342294573784, + 0.05563678964972496, + 0.08884871006011963, + 0.07457523792982101, + 0.0669797882437706, + 0.06255743652582169, + 0.30258607864379883, + 0.1798143982887268, + 0.02112538181245327 + ], + "credit_map": [ + 0.013238941319286823, + 0.017463799566030502, + 0.013696727342903614, + 0.01682085543870926, + 0.009780623018741608, + 0.007021394092589617, + 0.012987136840820312, + 0.009363193064928055, + 0.0035989070311188698, + 0.006332710850983858, + 0.004469645209610462, + 0.007520970888435841, + 0.0061312485486269, + 0.004208954516798258, + 0.008808491751551628, + 0.006432500202208757, + 0.004193369299173355, + 0.007734541781246662, + 0.007550001610070467, + 0.00850833859294653, + 0.0058885966427624226, + 0.005113308317959309, + 0.007610568776726723, + 0.009038073010742664, + 0.007997467182576656, + 0.01521753054112196, + 0.01751658320426941, + 0.01094499509781599, + 0.01055445522069931, + 0.006504662800580263, + 0.008561879396438599, + 0.011551134288311005, + 0.00852497573941946, + 0.007394855841994286, + 0.008569613099098206, + 0.008617187850177288, + 0.012786605395376682, + 0.007422184571623802, + 0.00987879280000925, + 0.011381025426089764, + 0.012058225460350513, + 0.006103040184825659, + 0.00849870778620243, + 0.00754172820597887, + 0.010052378289401531, + 0.008832228370010853, + 0.007266780827194452, + 0.006626703776419163, + 0.005230583716183901, + 0.006012970581650734, + 0.006087493151426315, + 0.006656383164227009, + 0.007779581006616354, + 0.009608888998627663, + 0.0054337382316589355, + 0.015747802332043648, + 0.019516397267580032, + 0.028289170935750008, + 0.016024934127926826, + 0.038087647408246994, + 0.06965772062540054, + 0.07407175749540329, + 0.030603395774960518, + 0.02633506804704666, + 0.0079717505723238, + 0.006553069222718477, + 0.006583178881555796, + 0.005851680412888527, + 0.0719301775097847, + 0.06062547490000725, + 0.006172269582748413, + 0.014126830734312534, + 0.002261595567688346, + 0.0012218381743878126, + 0.002189196879044175, + 0.0012253796448931098, + 0.003317833412438631, + 0.00395966274663806, + 0.0017816567560657859, + 0.005168219096958637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.011080636642873287 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic close-up of a parrot perched on a vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the parrot's head and neck a deep sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the parrot's breast a vibrant sun yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parrot's wings a rich emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368484601378441 + ], + "opposite_logprob_tokens": [ + -3.408684730529785 + ], + "logprob_reward": -0.03368484601378441, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368484601378441, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08148820698261261, + 0.03070458583533764, + 0.049273762851953506, + 0.21097417175769806, + 0.16673849523067474, + 0.12397806346416473, + 0.04713701084256172, + 0.10647235065698624, + 0.08364291489124298, + 0.0995904803276062 + ], + "credit_map": [ + 0.009408123791217804, + 0.014582549221813679, + 0.011450102552771568, + 0.012241249904036522, + 0.008930504322052002, + 0.007134189363569021, + 0.009861703962087631, + 0.00787978246808052, + 0.0021924006287008524, + 0.0054872590117156506, + 0.0030759084038436413, + 0.004973108414560556, + 0.003086462849751115, + 0.003292068373411894, + 0.00404604896903038, + 0.004551329184323549, + 0.003562888829037547, + 0.007211561780422926, + 0.006924766581505537, + 0.007553418166935444, + 0.004970361478626728, + 0.004980173893272877, + 0.007277378812432289, + 0.0067932154051959515, + 0.017977718263864517, + 0.022224057465791702, + 0.05069796368479729, + 0.019315695390105247, + 0.03226403892040253, + 0.014355858787894249, + 0.02086191624403, + 0.03327692300081253, + 0.013544558547437191, + 0.018889624625444412, + 0.021351097151637077, + 0.021969228982925415, + 0.026490002870559692, + 0.016500037163496017, + 0.025936633348464966, + 0.0220573078840971, + 0.026847247034311295, + 0.012375148944556713, + 0.01691460981965065, + 0.013827850110828876, + 0.014845247380435467, + 0.010752702131867409, + 0.015282266773283482, + 0.013132985681295395, + 0.004469498060643673, + 0.003187904367223382, + 0.007867787964642048, + 0.0034318778198212385, + 0.009102658368647099, + 0.008240852504968643, + 0.004163785837590694, + 0.006672642193734646, + 0.02444295957684517, + 0.008207617327570915, + 0.009255348704755306, + 0.006835779175162315, + 0.019080234691500664, + 0.0192539244890213, + 0.0072392746806144714, + 0.012157212011516094, + 0.004298543091863394, + 0.005073957145214081, + 0.006271690595895052, + 0.005988001823425293, + 0.025087030604481697, + 0.02452310547232628, + 0.006067507900297642, + 0.006333073601126671, + 0.00872764177620411, + 0.0031282100826501846, + 0.011980993673205376, + 0.0033619331661611795, + 0.008448759093880653, + 0.025559594854712486, + 0.005244364496320486, + 0.03313898295164108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 5 + ], + "selected_reward": -0.03368484601378441 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic close-up of a parrot perched on a vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the parrot's head and neck a deep sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the parrot's breast a vibrant sun yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parrot's wings a rich emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898371666669846 + ], + "opposite_logprob_tokens": [ + -2.578983783721924 + ], + "logprob_reward": -0.07898371666669846, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898371666669846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059858791530132294, + 0.053524360060691833, + 0.03859561309218407, + 0.0995323657989502, + 0.11386831849813461, + 0.026934757828712463, + 0.1179051548242569, + 0.08058174699544907, + 0.3836936950683594, + 0.02550516463816166 + ], + "credit_map": [ + 0.005818621255457401, + 0.010038938373327255, + 0.006368396338075399, + 0.014012356288731098, + 0.006475526839494705, + 0.004218992777168751, + 0.006882940884679556, + 0.006043016444891691, + 0.004660154227167368, + 0.004818584769964218, + 0.008434219285845757, + 0.007615128997713327, + 0.006459699012339115, + 0.005486596375703812, + 0.00813583005219698, + 0.007914149202406406, + 0.0038935523480176926, + 0.005321904085576534, + 0.005111376754939556, + 0.005812402814626694, + 0.0036469839978963137, + 0.003750942414626479, + 0.005519988015294075, + 0.005538460798561573, + 0.009800530970096588, + 0.018736252561211586, + 0.01586124300956726, + 0.0137141402810812, + 0.011386578902602196, + 0.007341755088418722, + 0.010176344774663448, + 0.012515515089035034, + 0.011073267087340355, + 0.015059449709951878, + 0.011156330816447735, + 0.014971231110394001, + 0.01572255976498127, + 0.01022334210574627, + 0.01863081380724907, + 0.01703132502734661, + 0.003547076368704438, + 0.0030468483455479145, + 0.003920755814760923, + 0.004528461489826441, + 0.003313880180940032, + 0.002079957164824009, + 0.0035107028670608997, + 0.0029870762955397367, + 0.004690764006227255, + 0.00730934040620923, + 0.006605958566069603, + 0.008580184541642666, + 0.017315249890089035, + 0.02823548950254917, + 0.009741618297994137, + 0.035426560789346695, + 0.01441203523427248, + 0.005628164857625961, + 0.005930286832153797, + 0.007450955454260111, + 0.016272317618131638, + 0.016907498240470886, + 0.006054028403013945, + 0.007926465943455696, + 0.01901998370885849, + 0.017787836492061615, + 0.02091878652572632, + 0.017229439690709114, + 0.14855007827281952, + 0.11540068686008453, + 0.015137916430830956, + 0.02964896522462368, + 0.0023092718329280615, + 0.0014509992906823754, + 0.0026439225766807795, + 0.0011803758097812533, + 0.002638191217556596, + 0.00611467519775033, + 0.0020552226342260838, + 0.007112504914402962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 4 + ], + "selected_reward": -0.07898371666669846 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic close-up slowly zooms in, capturing a magnificent parrot gracefully perched on a dense, gnarled jungle vine. The parrot's plumage is a breathtaking, surreal mosaic of color: its head and neck a deep sapphire blue, its breast a vibrant sun yellow, its wings a rich emerald green, each feather rendered in crisp detail. Its intelligent black eyes gleam, and its rapid, quirky turns of the head cause a subtle rustling of feathers. The vine it clings to is gnarled and ancient, its brown, richly textured bark clinging to a mat of damp green moss. The background is a vibrant, soft-focus blur of lush, dewy foliage, with large, shimmering tropical leaves. Dramatic, warm sunlight filters down through the dense canopy, penetrating the humid air, illuminating floating particles and creating a dynamic play of light and shadow across the scene. The overall atmosphere is vibrant and energized, a vibrant and dynamic snapshot of the pulsating heart of the rainforest.", + "seed": 6717, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_3/index_2841_seed_6717_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07857820279896259, + "vlm_logprob": -0.07857820279896259, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.07857820279896259, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic close-up of a parrot perched on a vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the parrot's head and neck a deep sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the parrot's breast a vibrant sun yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parrot's wings a rich emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022045783698558807 + ], + "opposite_logprob_tokens": [ + -4.147045612335205 + ], + "logprob_reward": -0.022045783698558807, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.022045783698558807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19853723049163818, + 0.029505852609872818, + 0.013662979938089848, + 0.17280539870262146, + 0.41335347294807434, + 0.025718532502651215, + 0.04023034870624542, + 0.01213789265602827, + 0.07650942355394363, + 0.017538830637931824 + ], + "credit_map": [ + 0.013210605829954147, + 0.0739419087767601, + 0.007502518594264984, + 0.007296817377209663, + 0.00874337274581194, + 0.011793381534516811, + 0.006615183781832457, + 0.06943343579769135, + 0.002916515339165926, + 0.0071594384498894215, + 0.002685015322640538, + 0.0018630065023899078, + 0.0018720177467912436, + 0.002529824385419488, + 0.002386908745393157, + 0.008093126118183136, + 0.0012329431483522058, + 0.003717331914231181, + 0.0021501900628209114, + 0.0007941526710055768, + 0.001184621942229569, + 0.0013098755152896047, + 0.0011949643958359957, + 0.0020789001137018204, + 0.01504612062126398, + 0.007599291391670704, + 0.06512875854969025, + 0.01152580976486206, + 0.0026146036107093096, + 0.011337146162986755, + 0.004709561355412006, + 0.054844100028276443, + 0.04514053836464882, + 0.01700785756111145, + 0.17048431932926178, + 0.009924063459038734, + 0.007073802873492241, + 0.018453801050782204, + 0.012356369756162167, + 0.13291271030902863, + 0.00366584793664515, + 0.002348525682464242, + 0.007580723613500595, + 0.002015521517023444, + 0.0021468179766088724, + 0.0014072388876229525, + 0.00233660452067852, + 0.004217253997921944, + 0.004783794283866882, + 0.002095230855047703, + 0.01151304505765438, + 0.002071095397695899, + 0.0023080208338797092, + 0.0023141002748161554, + 0.0025272639468312263, + 0.012617796659469604, + 0.0032017037738114595, + 0.0014773993752896786, + 0.00197205669246614, + 0.0010125822154805064, + 0.0007518931524828076, + 0.0008908477611839771, + 0.0013043462531641126, + 0.0015270636649802327, + 0.028656087815761566, + 0.003077809000387788, + 0.030518192797899246, + 0.003508741268888116, + 0.0018761473475024104, + 0.002223346382379532, + 0.0019276401726529002, + 0.004721460398286581, + 0.0023505501449108124, + 0.0017499012174084783, + 0.0037019250448793173, + 0.0016953934682533145, + 0.0018790285103023052, + 0.0015296214260160923, + 0.0019923378713428974, + 0.0026400743518024683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.022045783698558807 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic close-up of a parrot perched on a vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the parrot's head and neck a deep sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the parrot's breast a vibrant sun yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parrot's wings a rich emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28132450580596924 + ], + "opposite_logprob_tokens": [ + -1.4063245058059692 + ], + "logprob_reward": -0.28132450580596924, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28132450580596924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0748296007514, + 0.02009219489991665, + 0.03476202115416527, + 0.1415332704782486, + 0.5075863003730774, + 0.015960007905960083, + 0.04916426166892052, + 0.013634169474244118, + 0.1353408545255661, + 0.007097295485436916 + ], + "credit_map": [ + 0.004900677595287561, + 0.027491625398397446, + 0.0031738143879920244, + 0.0033553082030266523, + 0.003493386786431074, + 0.004147744737565517, + 0.002706085564568639, + 0.02556096576154232, + 0.0022077620960772038, + 0.0020168712362647057, + 0.0036289142444729805, + 0.0017931086476892233, + 0.0018866852624341846, + 0.0020113494247198105, + 0.003600973868742585, + 0.0029465309344232082, + 0.002899909159168601, + 0.011589680798351765, + 0.0036637892480939627, + 0.001777262194082141, + 0.0015809014439582825, + 0.0022481975611299276, + 0.002027059206739068, + 0.008975224569439888, + 0.01286549773067236, + 0.007977696135640144, + 0.050162456929683685, + 0.009029681794345379, + 0.004745835438370705, + 0.00873111467808485, + 0.00543862022459507, + 0.042582377791404724, + 0.05646892264485359, + 0.020745640620589256, + 0.21037471294403076, + 0.01190072949975729, + 0.008396243676543236, + 0.022422756999731064, + 0.015029148198664188, + 0.1622481793165207, + 0.0017713119741529226, + 0.0010962061351165175, + 0.005370358936488628, + 0.001105481875129044, + 0.0007433487917296588, + 0.0009554122225381434, + 0.001200418220832944, + 0.0037174716126173735, + 0.009356866590678692, + 0.002832193858921528, + 0.01033910270780325, + 0.002167210914194584, + 0.0022150243166834116, + 0.003099653637036681, + 0.002578513929620385, + 0.01657569594681263, + 0.0024538415018469095, + 0.0008772409055382013, + 0.004182828590273857, + 0.0010068576084449887, + 0.0011526077287271619, + 0.0007437054300680757, + 0.0012265382101759315, + 0.0019905506633222103, + 0.05126427859067917, + 0.005581438075751066, + 0.05404914915561676, + 0.005845176987349987, + 0.0031898976303637028, + 0.004182863049209118, + 0.0031472358386963606, + 0.008080818690359592, + 0.0008234751876443624, + 0.0008057568920776248, + 0.000962209131103009, + 0.0006067117210477591, + 0.0014106078306213021, + 0.0007013384019955993, + 0.0011318789329379797, + 0.0006553177954629064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 8 + ], + "selected_reward": -0.28132450580596924 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic close-up of a parrot perched on a vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the parrot's head and neck a deep sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the parrot's breast a vibrant sun yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parrot's wings a rich emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865194112062454 + ], + "opposite_logprob_tokens": [ + -3.048651933670044 + ], + "logprob_reward": -0.04865194112062454, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865194112062454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07928290218114853, + 0.021224718540906906, + 0.07001034170389175, + 0.1329449862241745, + 0.4615439474582672, + 0.02901645191013813, + 0.015949208289384842, + 0.018440978601574898, + 0.15864351391792297, + 0.012942987494170666 + ], + "credit_map": [ + 0.005033985245972872, + 0.028939640149474144, + 0.0031240067910403013, + 0.004101690836250782, + 0.0034652170725166798, + 0.0042549073696136475, + 0.0039177630096673965, + 0.026445694267749786, + 0.0013880410697311163, + 0.006278838962316513, + 0.0013501821085810661, + 0.0012279570801183581, + 0.001441253931261599, + 0.001539785647764802, + 0.0014439495280385017, + 0.00655471021309495, + 0.005571340210735798, + 0.02561294659972191, + 0.0044220127165317535, + 0.002907005837187171, + 0.003640095703303814, + 0.0033797668293118477, + 0.0027740998193621635, + 0.021703073754906654, + 0.011280223727226257, + 0.006438917014747858, + 0.04834967479109764, + 0.008688309229910374, + 0.003844509134069085, + 0.009391249157488346, + 0.004041386768221855, + 0.040910713374614716, + 0.05162845179438591, + 0.018413616344332695, + 0.19155605137348175, + 0.010678810067474842, + 0.007055954076349735, + 0.020639467984437943, + 0.013270655646920204, + 0.1483009308576584, + 0.0033773959148675203, + 0.0015704361721873283, + 0.00900533888489008, + 0.0012140055187046528, + 0.0014042959082871675, + 0.0016105909598991275, + 0.0016321904258802533, + 0.009202197194099426, + 0.0039225127547979355, + 0.0013824658235535026, + 0.0020742290653288364, + 0.0010190076427534223, + 0.001534322276711464, + 0.001234763185493648, + 0.0013906963868066669, + 0.0033912109211087227, + 0.0024523334577679634, + 0.0017288995441049337, + 0.0029679362196475267, + 0.0018977305153384805, + 0.0028983433730900288, + 0.0018430527998134494, + 0.0026910235174000263, + 0.0019616594072431326, + 0.060374870896339417, + 0.006343506742268801, + 0.0639476552605629, + 0.006987293250858784, + 0.0031657880172133446, + 0.004698210395872593, + 0.003491863841190934, + 0.00963432528078556, + 0.001957993023097515, + 0.0012736310018226504, + 0.0029328796081244946, + 0.0011362091172486544, + 0.0009724870324134827, + 0.0012442613951861858, + 0.0012418435653671622, + 0.002183683216571808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 3 + ], + "selected_reward": -0.04865194112062454 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic close-up of a parrot perched on a vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the parrot's head and neck a deep sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the parrot's breast a vibrant sun yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parrot's wings a rich emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011068610474467278 + ], + "opposite_logprob_tokens": [ + -4.511068820953369 + ], + "logprob_reward": -0.011068610474467278, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011068610474467278, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08805995434522629, + 0.03252458572387695, + 0.04470989853143692, + 0.040687479078769684, + 0.11127885431051254, + 0.03695373237133026, + 0.05194511264562607, + 0.0433536134660244, + 0.46093350648880005, + 0.08955322951078415 + ], + "credit_map": [ + 0.005635252222418785, + 0.02905300445854664, + 0.004491711501032114, + 0.005355391185730696, + 0.004287929739803076, + 0.005417265929281712, + 0.005695931147783995, + 0.028123468160629272, + 0.002776683308184147, + 0.006659398786723614, + 0.003250310430303216, + 0.003238470060750842, + 0.003665896598249674, + 0.0027434967923909426, + 0.0043980455957353115, + 0.00579228438436985, + 0.003148491494357586, + 0.012018673121929169, + 0.00427475618198514, + 0.003265111707150936, + 0.004206505604088306, + 0.0031727603636682034, + 0.0032076810020953417, + 0.011415920220315456, + 0.00476812431588769, + 0.0050835683941841125, + 0.009120733477175236, + 0.004508948419243097, + 0.0032545856665819883, + 0.0032623049337416887, + 0.004014861304312944, + 0.00667435210198164, + 0.009686561301350594, + 0.00822699349373579, + 0.036885954439640045, + 0.00627192622050643, + 0.006750299129635096, + 0.006103401072323322, + 0.008302190341055393, + 0.029051532968878746, + 0.004306357819586992, + 0.004144915845245123, + 0.0041431766003370285, + 0.004034968558698893, + 0.005505059380084276, + 0.003561189165338874, + 0.006919809617102146, + 0.004338255617767572, + 0.012440908700227737, + 0.0033108340576291084, + 0.010003739967942238, + 0.0036018078681081533, + 0.005558923818171024, + 0.0037261464167386293, + 0.005036333110183477, + 0.008266418240964413, + 0.009529203176498413, + 0.0036167434882372618, + 0.011477340012788773, + 0.0031534209847450256, + 0.004477644804865122, + 0.0029711804818361998, + 0.0036063888110220432, + 0.00452168844640255, + 0.17477887868881226, + 0.018309054896235466, + 0.18547989428043365, + 0.020855672657489777, + 0.008942990563809872, + 0.013566072098910809, + 0.010635504499077797, + 0.02836540713906288, + 0.014260840602219105, + 0.008264021947979927, + 0.022226011380553246, + 0.010693419724702835, + 0.007265900727361441, + 0.006350948475301266, + 0.009418154135346413, + 0.011073931120336056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.011068610474467278 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic close-up of a parrot perched on a vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the parrot's head and neck a deep sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the parrot's breast a vibrant sun yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parrot's wings a rich emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0298001728951931 + ], + "opposite_logprob_tokens": [ + -3.5298001766204834 + ], + "logprob_reward": -0.0298001728951931, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0298001728951931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04850070923566818, + 0.005017824936658144, + 0.016390301287174225, + 0.0186479315161705, + 0.4735524356365204, + 0.022247038781642914, + 0.006273748818784952, + 0.012977269478142262, + 0.38615259528160095, + 0.01024019718170166 + ], + "credit_map": [ + 0.0031838275026530027, + 0.017923230305314064, + 0.001989006996154785, + 0.0020329675171524286, + 0.002133421367034316, + 0.0027364774141460657, + 0.0019407847430557013, + 0.0165609922260046, + 0.00033256656024605036, + 0.0007221479318104684, + 0.0005978333647362888, + 0.0006256127962842584, + 0.0005157856503501534, + 0.00033167889341712, + 0.0008771470747888088, + 0.0010150523157790303, + 0.0012158970348536968, + 0.005365010350942612, + 0.001434944337233901, + 0.0010452084243297577, + 0.0012356051011011004, + 0.0009063653997145593, + 0.0008488933090120554, + 0.00433837715536356, + 0.0015906976768746972, + 0.0010937891202047467, + 0.006320266984403133, + 0.0013452512212097645, + 0.0008414466283284128, + 0.0010268499609082937, + 0.0009714784100651741, + 0.005458149127662182, + 0.053263481706380844, + 0.01906532794237137, + 0.19654135406017303, + 0.011006719432771206, + 0.007507523521780968, + 0.020754680037498474, + 0.013706905767321587, + 0.15170642733573914, + 0.002698468742892146, + 0.001478915335610509, + 0.00513416389003396, + 0.001566912978887558, + 0.0015614076983183622, + 0.001591865555383265, + 0.0018356984946876764, + 0.0063796015456318855, + 0.0018026254838332534, + 0.00041579888784326613, + 0.0013624008279293776, + 0.00035330525133758783, + 0.0004254925006534904, + 0.0004152916662860662, + 0.0003917608119081706, + 0.0011070731561630964, + 0.0033030021004378796, + 0.0010282736038789153, + 0.004219209309667349, + 0.0006672986783087254, + 0.0009373986395075917, + 0.0008678393787704408, + 0.0006822351133450866, + 0.0012720109662041068, + 0.1464853137731552, + 0.015412108972668648, + 0.15579967200756073, + 0.017372993752360344, + 0.007681221701204777, + 0.011382024735212326, + 0.008368405513465405, + 0.023650843650102615, + 0.001361860311590135, + 0.001197259989567101, + 0.001957949949428439, + 0.0012978189624845982, + 0.000999158015474677, + 0.0008595646941103041, + 0.0012412710348144174, + 0.0013253145152702928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.0298001728951931 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_4/index_tensor([8086, 8086])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_4/index_tensor([8086, 8086])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6a6caf59a52a58ade8eb1cb1758cf3365a16a681 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_4/index_tensor([8086, 8086])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "From a first-person perspective, an uninterrupted gliding shot moves smoothly forward along a park path, all rendered in a clean, minimalist art style defined by elegant, soft gradients. The air is thick with a dense, gentle mass of cherry blossom petals, which slowly fall like pink snow, carpeting the stone-paved path with a delicate, textured carpet. In the background, the soft-focus bokeh of countless other cherry trees creates a bright, all-encompassing canopy of vibrant pink and white. Diffuse, ethereal light filters through the blossoms, casting no harsh shadows and bathing the entire scene in a warm, tranquil glow. As the camera advances, a slender, pale-skinned hand enters the frame from the lower right, its form defined by smooth tonal transitions rather than sharp lines. Moving with deliberate grace, the hand extends forward until it reaches its center frame, offering the viewer a single, perfectly formed branch of cherry blossoms. The branch itself is a dark, smooth wood from which sprout several beautifully shaped flowers, with five delicate petals perfectly blending soft white and pale pink gradients, creating a peaceful, intimate and dreamy atmosphere.", + "seed": 8771, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_4/index_8086_seed_8771_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.008541455771774054, + "vlm_logprob": -0.008541455771774054, + "vlm_yes_no_margin": 5.15, + "vlm_token_credit": -0.008541455771774054, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective moving forward along a park path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style clean and minimalist with soft gradients and no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, pale-skinned hand entering from the lower right and extending to the center frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand offering a single branch with dark smooth wood and flowers having five petals in soft white and pale pink gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus bokeh of cherry trees creating a canopy of vibrant pink and white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014428450725972652 + ], + "opposite_logprob_tokens": [ + -4.639428615570068 + ], + "logprob_reward": -0.014428450725972652, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.014428450725972652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10624129325151443, + 0.09889921545982361, + 0.026832666248083115, + 0.013196020387113094, + 0.052071284502744675, + 0.4431924819946289, + 0.06710769236087799, + 0.023039305582642555, + 0.11449889838695526, + 0.054921116679906845 + ], + "credit_map": [ + 0.011497212573885918, + 0.01636948250234127, + 0.009915377013385296, + 0.010922924615442753, + 0.01987026445567608, + 0.014775150455534458, + 0.0076921600848436356, + 0.015198716893792152, + 0.01215924322605133, + 0.014398962259292603, + 0.01167527586221695, + 0.008566304109990597, + 0.0154543062672019, + 0.0140689043328166, + 0.014490518718957901, + 0.008085698820650578, + 0.002902466803789139, + 0.002826882293447852, + 0.0034779333509504795, + 0.0023585332091897726, + 0.005251920782029629, + 0.0031762300059199333, + 0.0038979523815214634, + 0.002940748119726777, + 0.0019571997690945864, + 0.0017054971540346742, + 0.0013881054474040866, + 0.0011386544210836291, + 0.0016266132006421685, + 0.0025320749264210463, + 0.0017946865409612656, + 0.001053188811056316, + 0.005561784841120243, + 0.0071592689491808414, + 0.005241820588707924, + 0.004523982759565115, + 0.005624089390039444, + 0.013080950826406479, + 0.0045513673685491085, + 0.0063280221074819565, + 0.03959894925355911, + 0.12018308788537979, + 0.01751146838068962, + 0.029363717883825302, + 0.10099629312753677, + 0.08352569490671158, + 0.01976596936583519, + 0.03224729001522064, + 0.007536588702350855, + 0.011982937343418598, + 0.008506576530635357, + 0.01023428700864315, + 0.008811279200017452, + 0.008835284039378166, + 0.0060514179058372974, + 0.0051493258215487, + 0.002120572840794921, + 0.0033057802356779575, + 0.0021931969095021486, + 0.0027776139322668314, + 0.002451553475111723, + 0.0029949534218758345, + 0.004466360434889793, + 0.0027292759623378515, + 0.027200421318411827, + 0.011541750282049179, + 0.010303459130227566, + 0.014302127063274384, + 0.013317275792360306, + 0.012743178755044937, + 0.011141764931380749, + 0.01394891832023859, + 0.00594264967367053, + 0.006069169845432043, + 0.004693952389061451, + 0.00734941428527236, + 0.007660895586013794, + 0.007411807309836149, + 0.009188932366669178, + 0.006604295689612627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.014428450725972652 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective moving forward along a park path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style clean and minimalist with soft gradients and no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, pale-skinned hand entering from the lower right and extending to the center frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand offering a single branch with dark smooth wood and flowers having five petals in soft white and pale pink gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus bokeh of cherry trees creating a canopy of vibrant pink and white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014213516376912594 + ], + "opposite_logprob_tokens": [ + -4.264213562011719 + ], + "logprob_reward": -0.014213516376912594, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014213516376912594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07132477313280106, + 0.15292571485042572, + 0.03423778712749481, + 0.06376662850379944, + 0.17849569022655487, + 0.2299853265285492, + 0.022243240848183632, + 0.018163710832595825, + 0.1349697709083557, + 0.0938873365521431 + ], + "credit_map": [ + 0.0070456406101584435, + 0.011580446735024452, + 0.009217606857419014, + 0.008532347157597542, + 0.008038738742470741, + 0.011835303157567978, + 0.006243518088012934, + 0.008831177838146687, + 0.02473323978483677, + 0.028976257890462875, + 0.012995646335184574, + 0.021904204040765762, + 0.018493685871362686, + 0.019618241116404533, + 0.01083580031991005, + 0.01536865159869194, + 0.005493836011737585, + 0.00471270689740777, + 0.004071150440722704, + 0.0025599466171115637, + 0.004806166049093008, + 0.004908936098217964, + 0.0032125916332006454, + 0.004472452215850353, + 0.009155665524303913, + 0.008007003925740719, + 0.00713264150545001, + 0.006099845748394728, + 0.01295696571469307, + 0.009602289646863937, + 0.006674860138446093, + 0.004137362819164991, + 0.018291553482413292, + 0.03286012262105942, + 0.015967724844813347, + 0.017825303599238396, + 0.021771317347884178, + 0.04324640333652496, + 0.011302332393825054, + 0.017230944707989693, + 0.01940583996474743, + 0.06550123542547226, + 0.008653760887682438, + 0.013972925022244453, + 0.058683332055807114, + 0.03621973469853401, + 0.010905059054493904, + 0.016643453389406204, + 0.0024425932206213474, + 0.0045463149435818195, + 0.002771412953734398, + 0.002340833656489849, + 0.003241532715037465, + 0.003071719082072377, + 0.0014374973252415657, + 0.0023913385812193155, + 0.0015647674445062876, + 0.0019336303230375051, + 0.002438467927277088, + 0.002206905512139201, + 0.002133334754034877, + 0.0026691833045333624, + 0.0031463608611375093, + 0.0020710607059299946, + 0.014572454616427422, + 0.01488799974322319, + 0.011953640729188919, + 0.022791769355535507, + 0.013472502119839191, + 0.012490502558648586, + 0.024597352370619774, + 0.020203547552227974, + 0.010052175261080265, + 0.017964527010917664, + 0.008530596271157265, + 0.009407365694642067, + 0.010882914066314697, + 0.01005369983613491, + 0.011073577217757702, + 0.015922483056783676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 1 + ], + "selected_reward": -0.014213516376912594 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective moving forward along a park path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style clean and minimalist with soft gradients and no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, pale-skinned hand entering from the lower right and extending to the center frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand offering a single branch with dark smooth wood and flowers having five petals in soft white and pale pink gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus bokeh of cherry trees creating a canopy of vibrant pink and white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004093601368367672 + ], + "opposite_logprob_tokens": [ + -5.504093647003174 + ], + "logprob_reward": -0.004093601368367672, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004093601368367672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05827823653817177, + 0.03691466525197029, + 0.016981346532702446, + 0.028766034170985222, + 0.033517707139253616, + 0.2948507070541382, + 0.02722625620663166, + 0.023556943982839584, + 0.10677879303693771, + 0.373129278421402 + ], + "credit_map": [ + 0.007782258093357086, + 0.00869800616055727, + 0.0061225565150380135, + 0.006138152442872524, + 0.009596606716513634, + 0.007684655021876097, + 0.004889636766165495, + 0.007366362027823925, + 0.005678222049027681, + 0.006542721297591925, + 0.0038310475647449493, + 0.0044655571691691875, + 0.005517336539924145, + 0.004755155649036169, + 0.003163970308378339, + 0.0029606528114527464, + 0.002726192818954587, + 0.0022246839944273233, + 0.002404356375336647, + 0.001552567700855434, + 0.0022048186510801315, + 0.002394731855019927, + 0.0017905335407704115, + 0.0016834619455039501, + 0.004130987916141748, + 0.0029627045150846243, + 0.003244261723011732, + 0.002975254785269499, + 0.005123573821038008, + 0.0041683935560286045, + 0.004102727863937616, + 0.002058130456134677, + 0.004142028745263815, + 0.0047950162552297115, + 0.004373369738459587, + 0.003466269699856639, + 0.004200160037726164, + 0.00565127981826663, + 0.0030890346970409155, + 0.0038005488459020853, + 0.025757689028978348, + 0.08485884219408035, + 0.010739986784756184, + 0.016418248414993286, + 0.07513360679149628, + 0.048542268574237823, + 0.012818640097975731, + 0.0205814391374588, + 0.0035746437497437, + 0.0030155775602906942, + 0.00209380523301661, + 0.0026164401788264513, + 0.0032505656126886606, + 0.0050989179871976376, + 0.004603415261954069, + 0.0029728917870670557, + 0.002481302246451378, + 0.004327035043388605, + 0.0037586763501167297, + 0.00357043300755322, + 0.002195364097133279, + 0.0029662686865776777, + 0.002312007825821638, + 0.0019458559108898044, + 0.022542130202054977, + 0.010489514097571373, + 0.008376060053706169, + 0.01224212720990181, + 0.013002289459109306, + 0.012465199455618858, + 0.013319878838956356, + 0.014341595582664013, + 0.03275943547487259, + 0.0631483718752861, + 0.02293725125491619, + 0.02444501966238022, + 0.0502903014421463, + 0.05295746773481369, + 0.04822167009115219, + 0.07836975157260895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.004093601368367672 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective moving forward along a park path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style clean and minimalist with soft gradients and no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, pale-skinned hand entering from the lower right and extending to the center frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand offering a single branch with dark smooth wood and flowers having five petals in soft white and pale pink gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus bokeh of cherry trees creating a canopy of vibrant pink and white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636504411697388 + ], + "opposite_logprob_tokens": [ + -4.758636474609375 + ], + "logprob_reward": -0.008636504411697388, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636504411697388, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054350320249795914, + 0.105238176882267, + 0.03796416148543358, + 0.038753725588321686, + 0.0240738857537508, + 0.4128960967063904, + 0.08538434654474258, + 0.029820263385772705, + 0.06384571641683578, + 0.14767338335514069 + ], + "credit_map": [ + 0.008103553205728531, + 0.00875181332230568, + 0.0050925398245453835, + 0.005868124775588512, + 0.006610326003283262, + 0.010232701897621155, + 0.00494140712544322, + 0.004749847110360861, + 0.016045359894633293, + 0.019194981083273888, + 0.00860624574124813, + 0.0144509207457304, + 0.011353936046361923, + 0.014291804283857346, + 0.009698131121695042, + 0.011596783995628357, + 0.00511355884373188, + 0.0056373621337115765, + 0.004642112646251917, + 0.0034030310343950987, + 0.00635415967553854, + 0.005571592133492231, + 0.0035646993201225996, + 0.00367764406837523, + 0.00481392489746213, + 0.004699463490396738, + 0.005110822152346373, + 0.003835059702396393, + 0.007425528950989246, + 0.005399636458605528, + 0.004346392583101988, + 0.0031228901352733374, + 0.003322909353300929, + 0.0039810361340641975, + 0.0014791827416047454, + 0.001753646065481007, + 0.003277479438111186, + 0.006366139277815819, + 0.0018549973610788584, + 0.002038494450971484, + 0.03474884480237961, + 0.12127598375082016, + 0.014579126611351967, + 0.022865287959575653, + 0.10961935669183731, + 0.06342169642448425, + 0.01865512505173683, + 0.02773064747452736, + 0.00843129400163889, + 0.0198709424585104, + 0.010933863930404186, + 0.00981086865067482, + 0.011394040659070015, + 0.01224230695515871, + 0.0067582279443740845, + 0.005942796356976032, + 0.0025807153433561325, + 0.0049784062430262566, + 0.002232953906059265, + 0.0035108125302940607, + 0.004171918611973524, + 0.005305554252117872, + 0.004518521949648857, + 0.0025213807821273804, + 0.008456455543637276, + 0.008211346343159676, + 0.005418778862804174, + 0.011604908853769302, + 0.006462442688643932, + 0.007967781275510788, + 0.006996214855462313, + 0.008727779611945152, + 0.01604638621211052, + 0.021356606855988503, + 0.012258266098797321, + 0.011967500671744347, + 0.024390745908021927, + 0.02055268920958042, + 0.020113853737711906, + 0.020987315103411674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 1 + ], + "selected_reward": -0.008636504411697388 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective moving forward along a park path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style clean and minimalist with soft gradients and no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, pale-skinned hand entering from the lower right and extending to the center frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand offering a single branch with dark smooth wood and flowers having five petals in soft white and pale pink gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus bokeh of cherry trees creating a canopy of vibrant pink and white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001335205975919962 + ], + "opposite_logprob_tokens": [ + -6.626335144042969 + ], + "logprob_reward": -0.001335205975919962, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001335205975919962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08596032112836838, + 0.09798789769411087, + 0.053224749863147736, + 0.05800473317503929, + 0.04908901825547218, + 0.338128924369812, + 0.04579390212893486, + 0.06719827651977539, + 0.08399304002523422, + 0.12061911076307297 + ], + "credit_map": [ + 0.013371563516557217, + 0.010236470028758049, + 0.011153863742947578, + 0.007980911992490292, + 0.011692068539559841, + 0.012556229718029499, + 0.010294230654835701, + 0.00867499690502882, + 0.015752334147691727, + 0.019096726551651955, + 0.008200489915907383, + 0.012795518152415752, + 0.01119242887943983, + 0.013023138977587223, + 0.00831393338739872, + 0.009613336995244026, + 0.007976839318871498, + 0.008531206287443638, + 0.006535302381962538, + 0.004964349791407585, + 0.007506603840738535, + 0.00725893909111619, + 0.0054843854159116745, + 0.00496712839230895, + 0.008668030612170696, + 0.0067194425500929356, + 0.0055943820625543594, + 0.005414373241364956, + 0.01084146834909916, + 0.009245731867849827, + 0.007856299169361591, + 0.0036650088150054216, + 0.005934899672865868, + 0.008239979855716228, + 0.005644472315907478, + 0.005278072319924831, + 0.005588345229625702, + 0.010173424147069454, + 0.003440580330789089, + 0.004789246246218681, + 0.0289171002805233, + 0.09716253727674484, + 0.011922609061002731, + 0.01882118731737137, + 0.08733686059713364, + 0.05515691265463829, + 0.015174835920333862, + 0.023636890575289726, + 0.004869478289037943, + 0.007143874187022448, + 0.006249372381716967, + 0.005983009003102779, + 0.00526426499709487, + 0.00461568171158433, + 0.006181074306368828, + 0.005487154237926006, + 0.0074325748719275, + 0.01401908602565527, + 0.0076938048005104065, + 0.008126798085868359, + 0.00709345331415534, + 0.00815585721284151, + 0.0078083970583975315, + 0.006868305150419474, + 0.009274332784116268, + 0.009714454412460327, + 0.012776237912476063, + 0.012228595092892647, + 0.007794278208166361, + 0.0117723373696208, + 0.010127868503332138, + 0.010304935276508331, + 0.011290461756289005, + 0.01821022853255272, + 0.010128401219844818, + 0.015644319355487823, + 0.010789326392114162, + 0.013254379853606224, + 0.01632234826683998, + 0.02497965283691883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 1 + ], + "selected_reward": -0.001335205975919962 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a first-person perspective, an uninterrupted gliding shot moves smoothly forward along a park path, all rendered in a clean, minimalist art style defined by elegant, soft gradients. The air is thick with a dense, gentle mass of cherry blossom petals, which slowly fall like pink snow, carpeting the stone-paved path with a delicate, textured carpet. In the background, the soft-focus bokeh of countless other cherry trees creates a bright, all-encompassing canopy of vibrant pink and white. Diffuse, ethereal light filters through the blossoms, casting no harsh shadows and bathing the entire scene in a warm, tranquil glow. As the camera advances, a slender, pale-skinned hand enters the frame from the lower right, its form defined by smooth tonal transitions rather than sharp lines. Moving with deliberate grace, the hand extends forward until it reaches its center frame, offering the viewer a single, perfectly formed branch of cherry blossoms. The branch itself is a dark, smooth wood from which sprout several beautifully shaped flowers, with five delicate petals perfectly blending soft white and pale pink gradients, creating a peaceful, intimate and dreamy atmosphere.", + "seed": 8772, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_4/index_8086_seed_8772_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.011318620806559921, + "vlm_logprob": -0.011318620806559921, + "vlm_yes_no_margin": 5.075, + "vlm_token_credit": -0.011318620806559921, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective moving forward along a park path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style clean and minimalist with soft gradients and no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, pale-skinned hand entering from the lower right and extending to the center frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand offering a single branch with dark smooth wood and flowers having five petals in soft white and pale pink gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus bokeh of cherry trees creating a canopy of vibrant pink and white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011382521130144596 + ], + "opposite_logprob_tokens": [ + -5.386382579803467 + ], + "logprob_reward": -0.011382521130144596, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.011382521130144596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0964767262339592, + 0.043606869876384735, + 0.062394846230745316, + 0.1072985976934433, + 0.08197075128555298, + 0.05851175636053085, + 0.09373446553945541, + 0.07840253412723541, + 0.30803239345550537, + 0.06957107037305832 + ], + "credit_map": [ + 0.010565808042883873, + 0.020157799124717712, + 0.011015133932232857, + 0.010734910145401955, + 0.014713567681610584, + 0.01004102360457182, + 0.012579366564750671, + 0.006669113412499428, + 0.004696703050285578, + 0.007971950806677341, + 0.00593888945877552, + 0.005569806322455406, + 0.006027757655829191, + 0.0044954800978302956, + 0.005235536489635706, + 0.0036707406397908926, + 0.006915553938597441, + 0.008131999522447586, + 0.012229209765791893, + 0.008937589824199677, + 0.006199480500072241, + 0.007352340966463089, + 0.005708801560103893, + 0.006919867359101772, + 0.013602971099317074, + 0.014195511117577553, + 0.011144126765429974, + 0.011901957914233208, + 0.012596573680639267, + 0.023799337446689606, + 0.010577055625617504, + 0.00948106124997139, + 0.007385662756860256, + 0.011685462668538094, + 0.00999479554593563, + 0.008803225122392178, + 0.013525682501494884, + 0.015275533311069012, + 0.008457884192466736, + 0.00684249447658658, + 0.005546906031668186, + 0.006645205896347761, + 0.008546562865376472, + 0.006176326889544725, + 0.008063707500696182, + 0.009202911518514156, + 0.007080377079546452, + 0.007249752059578896, + 0.012682253494858742, + 0.010873166844248772, + 0.007728589233011007, + 0.010397158563137054, + 0.012897414155304432, + 0.013729996979236603, + 0.01251937821507454, + 0.012906498275697231, + 0.008177665062248707, + 0.010695342905819416, + 0.005914583802223206, + 0.008175879716873169, + 0.008194874972105026, + 0.01101510040462017, + 0.0156569704413414, + 0.010572110302746296, + 0.02833830751478672, + 0.032301925122737885, + 0.042058058083057404, + 0.03813515231013298, + 0.03598630428314209, + 0.030938349664211273, + 0.0595853365957737, + 0.04068894684314728, + 0.003145829075947404, + 0.023220498114824295, + 0.0037335671950131655, + 0.004503806587308645, + 0.024258920922875404, + 0.00381834851577878, + 0.003974313382059336, + 0.0029157805256545544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.011382521130144596 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective moving forward along a park path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style clean and minimalist with soft gradients and no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, pale-skinned hand entering from the lower right and extending to the center frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand offering a single branch with dark smooth wood and flowers having five petals in soft white and pale pink gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus bokeh of cherry trees creating a canopy of vibrant pink and white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026356877759099007 + ], + "opposite_logprob_tokens": [ + -3.6513569355010986 + ], + "logprob_reward": -0.026356877759099007, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026356877759099007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06214580684900284, + 0.04548249766230583, + 0.050090085715055466, + 0.2631801962852478, + 0.044221073389053345, + 0.10859689861536026, + 0.16527128219604492, + 0.021174326539039612, + 0.18490150570869446, + 0.05493632331490517 + ], + "credit_map": [ + 0.008334405720233917, + 0.009707269258797169, + 0.006426963023841381, + 0.007014696951955557, + 0.00906003825366497, + 0.009139995090663433, + 0.007487050723284483, + 0.004975386429578066, + 0.004793752916157246, + 0.005840135272592306, + 0.006345196161419153, + 0.0067520481534302235, + 0.005728283431380987, + 0.006539180874824524, + 0.005565413739532232, + 0.0039184861816465855, + 0.005715314298868179, + 0.006690211594104767, + 0.006728849373757839, + 0.008475973270833492, + 0.005511768162250519, + 0.006952246651053429, + 0.0060847364366054535, + 0.0039309849962592125, + 0.041799839586019516, + 0.035254985094070435, + 0.03090534172952175, + 0.029281476512551308, + 0.029781457036733627, + 0.049328699707984924, + 0.024097183719277382, + 0.022731227800250053, + 0.004535720683634281, + 0.0059899827465415, + 0.004495220258831978, + 0.005416537635028362, + 0.007604868151247501, + 0.008408837951719761, + 0.004243702162057161, + 0.003526204265654087, + 0.010337870568037033, + 0.014311445876955986, + 0.014595277607440948, + 0.016069212928414345, + 0.013644362799823284, + 0.019502514973282814, + 0.010923745110630989, + 0.009212467819452286, + 0.02083311788737774, + 0.019356733188033104, + 0.016826733946800232, + 0.019066983833909035, + 0.019886191934347153, + 0.025188181549310684, + 0.02269795909523964, + 0.02141537144780159, + 0.0025931280106306076, + 0.0029203903395682573, + 0.0020538149401545525, + 0.0022983269300311804, + 0.0025296839885413647, + 0.003901304677128792, + 0.002204503631219268, + 0.002673174487426877, + 0.016004687175154686, + 0.022965574637055397, + 0.020227713510394096, + 0.021192442625761032, + 0.025047527626156807, + 0.019600167870521545, + 0.03211355209350586, + 0.027749842032790184, + 0.003292868146672845, + 0.016320085152983665, + 0.002026896458119154, + 0.003751029958948493, + 0.01629648171365261, + 0.005063007585704327, + 0.004690384492278099, + 0.003495569108054042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 6 + ], + "selected_reward": -0.026356877759099007 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective moving forward along a park path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style clean and minimalist with soft gradients and no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, pale-skinned hand entering from the lower right and extending to the center frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand offering a single branch with dark smooth wood and flowers having five petals in soft white and pale pink gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus bokeh of cherry trees creating a canopy of vibrant pink and white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006739504635334015 + ], + "opposite_logprob_tokens": [ + -5.006739616394043 + ], + "logprob_reward": -0.006739504635334015, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006739504635334015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09160879999399185, + 0.02238861471414566, + 0.02520967833697796, + 0.023219937458634377, + 0.017403554171323776, + 0.041441623121500015, + 0.03432023897767067, + 0.023122400045394897, + 0.6475637555122375, + 0.07372144609689713 + ], + "credit_map": [ + 0.008505672216415405, + 0.023063378408551216, + 0.009292523376643658, + 0.008191048167645931, + 0.01932586543262005, + 0.007932247593998909, + 0.009050016291439533, + 0.006248037796467543, + 0.0024226221721619368, + 0.0030472707003355026, + 0.0036736156325787306, + 0.0025926693342626095, + 0.003244647989049554, + 0.002297410275787115, + 0.002858326304703951, + 0.0022520488128066063, + 0.002604754874482751, + 0.0036909733898937702, + 0.003709559328854084, + 0.0031814994290471077, + 0.0031033128034323454, + 0.004258584696799517, + 0.002340608276426792, + 0.002320383209735155, + 0.002803641138598323, + 0.0032216105610132217, + 0.002064458094537258, + 0.002249471377581358, + 0.003093486884608865, + 0.005452267825603485, + 0.0023360729683190584, + 0.0019989279098808765, + 0.0018405101727694273, + 0.0022778466809540987, + 0.002547090407460928, + 0.0023090806789696217, + 0.0021750819869339466, + 0.002974479692056775, + 0.001624025055207312, + 0.001655436703003943, + 0.005138936918228865, + 0.005055406130850315, + 0.006342287641018629, + 0.0062348488718271255, + 0.004172987770289183, + 0.006943379063159227, + 0.003710211720317602, + 0.0038435610476881266, + 0.004466681275516748, + 0.004023247864097357, + 0.00317367329262197, + 0.005306673236191273, + 0.0042183673940598965, + 0.004405182786285877, + 0.004903820808976889, + 0.0038225867319852114, + 0.0019807966891676188, + 0.002914737444370985, + 0.0016505724051967263, + 0.0033979052677750587, + 0.0027560852468013763, + 0.0030798676889389753, + 0.0049680666998028755, + 0.002374363364651799, + 0.056249454617500305, + 0.08112547546625137, + 0.07394092530012131, + 0.06935413181781769, + 0.09361667931079865, + 0.07016603648662567, + 0.10603920370340347, + 0.0970718041062355, + 0.007851919159293175, + 0.012258163653314114, + 0.008058239705860615, + 0.010127144865691662, + 0.009854708798229694, + 0.009164857678115368, + 0.010730537585914135, + 0.005675871856510639 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.006739504635334015 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective moving forward along a park path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style clean and minimalist with soft gradients and no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, pale-skinned hand entering from the lower right and extending to the center frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand offering a single branch with dark smooth wood and flowers having five petals in soft white and pale pink gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus bokeh of cherry trees creating a canopy of vibrant pink and white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072972789406776 + ], + "opposite_logprob_tokens": [ + -4.511073112487793 + ], + "logprob_reward": -0.011072972789406776, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072972789406776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09385975450277328, + 0.02549660950899124, + 0.05546471104025841, + 0.05052654817700386, + 0.03472580015659332, + 0.09933813661336899, + 0.0514955073595047, + 0.022951174527406693, + 0.40660133957862854, + 0.1595403552055359 + ], + "credit_map": [ + 0.012041613459587097, + 0.014775441028177738, + 0.010736439377069473, + 0.012847497127950191, + 0.012395917437970638, + 0.014540827833116055, + 0.00952752586454153, + 0.0069945016875863075, + 0.0023517259396612644, + 0.004124640952795744, + 0.0043699657544493675, + 0.0028031538240611553, + 0.0040664286352694035, + 0.0028712621424347162, + 0.0023459235671907663, + 0.002563509624451399, + 0.005724238231778145, + 0.007685848977416754, + 0.009581571444869041, + 0.00843343697488308, + 0.005759701132774353, + 0.00775194400921464, + 0.00567487720400095, + 0.004853095859289169, + 0.008079949766397476, + 0.00602602306753397, + 0.008340368047356606, + 0.005676508415490389, + 0.004754676017910242, + 0.007566192653030157, + 0.0049557979218661785, + 0.0051270355470478535, + 0.004275367595255375, + 0.004182657692581415, + 0.003660145215690136, + 0.0031717480160295963, + 0.0036944039165973663, + 0.008136136457324028, + 0.0038206614553928375, + 0.0037846846971660852, + 0.010768016800284386, + 0.012327420525252819, + 0.0161394365131855, + 0.01455400139093399, + 0.01040653046220541, + 0.01727941446006298, + 0.008232095278799534, + 0.009631235152482986, + 0.006017678417265415, + 0.006129206158220768, + 0.004595956765115261, + 0.005673880688846111, + 0.005826212931424379, + 0.007673928048461676, + 0.007863274775445461, + 0.007715370040386915, + 0.0027148111257702112, + 0.0032721483148634434, + 0.0018839766271412373, + 0.0037165037356317043, + 0.002609384013339877, + 0.0024818708188831806, + 0.0041254484094679356, + 0.0021470333449542522, + 0.031568847596645355, + 0.05578820779919624, + 0.07852799445390701, + 0.051332373172044754, + 0.03566042706370354, + 0.045412737876176834, + 0.06766452640295029, + 0.04064622521400452, + 0.011063319630920887, + 0.04310441017150879, + 0.008840542286634445, + 0.009272177703678608, + 0.0502801239490509, + 0.015615778043866158, + 0.012275838293135166, + 0.009088183753192425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.011072972789406776 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective moving forward along a park path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style clean and minimalist with soft gradients and no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, pale-skinned hand entering from the lower right and extending to the center frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand offering a single branch with dark smooth wood and flowers having five petals in soft white and pale pink gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus bokeh of cherry trees creating a canopy of vibrant pink and white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010412277188152075 + ], + "opposite_logprob_tokens": [ + -6.876041412353516 + ], + "logprob_reward": -0.0010412277188152075, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010412277188152075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09032362699508667, + 0.025717590004205704, + 0.1049690842628479, + 0.11475202441215515, + 0.04858545586466789, + 0.22083641588687897, + 0.05673375725746155, + 0.04310530424118042, + 0.2589893937110901, + 0.03598736226558685 + ], + "credit_map": [ + 0.011780383065342903, + 0.014107754454016685, + 0.010696571320295334, + 0.01272298488765955, + 0.010997021570801735, + 0.012332416139543056, + 0.010439894162118435, + 0.007246595341712236, + 0.003501488361507654, + 0.0036255421582609415, + 0.003281721379607916, + 0.0031852424144744873, + 0.0034977998584508896, + 0.0037284756544977427, + 0.0024788170121610165, + 0.0024185008369386196, + 0.010325531475245953, + 0.014905025251209736, + 0.01765415444970131, + 0.01903288997709751, + 0.010164955630898476, + 0.012019923888146877, + 0.010312597267329693, + 0.010553991422057152, + 0.01661129854619503, + 0.015184128656983376, + 0.014417033642530441, + 0.01480229664593935, + 0.012119163759052753, + 0.02061203122138977, + 0.010335063561797142, + 0.010670995339751244, + 0.003926469944417477, + 0.007152094505727291, + 0.006040493492037058, + 0.005883107427507639, + 0.0077927871607244015, + 0.008167002350091934, + 0.005058464594185352, + 0.004565033130347729, + 0.027731796726584435, + 0.027820032089948654, + 0.038248851895332336, + 0.030146678909659386, + 0.0230392012745142, + 0.03202762454748154, + 0.01731438748538494, + 0.024507828056812286, + 0.005896118935197592, + 0.0070055085234344006, + 0.005338810849934816, + 0.008487626910209656, + 0.007922139018774033, + 0.00887668039649725, + 0.006804970093071461, + 0.006401902996003628, + 0.004324870649725199, + 0.006034702528268099, + 0.0043882159516215324, + 0.009321405552327633, + 0.005100409034639597, + 0.006181806791573763, + 0.003754435805603862, + 0.003999452572315931, + 0.022617878392338753, + 0.028199952095746994, + 0.02776520699262619, + 0.032537221908569336, + 0.03257734328508377, + 0.018910212442278862, + 0.05324127897620201, + 0.043140292167663574, + 0.003737917635589838, + 0.005009220447391272, + 0.0032711587846279144, + 0.00344967283308506, + 0.005853394512087107, + 0.004628568422049284, + 0.005882638040930033, + 0.004154788795858622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 3 + ], + "selected_reward": -0.0010412277188152075 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a first-person perspective, an uninterrupted gliding shot moves smoothly forward along a park path, all rendered in a clean, minimalist art style defined by elegant, soft gradients. The air is thick with a dense, gentle mass of cherry blossom petals, which slowly fall like pink snow, carpeting the stone-paved path with a delicate, textured carpet. In the background, the soft-focus bokeh of countless other cherry trees creates a bright, all-encompassing canopy of vibrant pink and white. Diffuse, ethereal light filters through the blossoms, casting no harsh shadows and bathing the entire scene in a warm, tranquil glow. As the camera advances, a slender, pale-skinned hand enters the frame from the lower right, its form defined by smooth tonal transitions rather than sharp lines. Moving with deliberate grace, the hand extends forward until it reaches its center frame, offering the viewer a single, perfectly formed branch of cherry blossoms. The branch itself is a dark, smooth wood from which sprout several beautifully shaped flowers, with five delicate petals perfectly blending soft white and pale pink gradients, creating a peaceful, intimate and dreamy atmosphere.", + "seed": 8773, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_4/index_8086_seed_8773_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.00600166644435376, + "vlm_logprob": -0.00600166644435376, + "vlm_yes_no_margin": 5.45, + "vlm_token_credit": -0.00600166644435376, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective moving forward along a park path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style clean and minimalist with soft gradients and no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, pale-skinned hand entering from the lower right and extending to the center frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand offering a single branch with dark smooth wood and flowers having five petals in soft white and pale pink gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus bokeh of cherry trees creating a canopy of vibrant pink and white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011841363273561 + ], + "opposite_logprob_tokens": [ + -4.761841297149658 + ], + "logprob_reward": -0.011841363273561, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.011841363273561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19828768074512482, + 0.08425184339284897, + 0.04096218943595886, + 0.10921785980463028, + 0.04341008514165878, + 0.035346344113349915, + 0.0793527290225029, + 0.3146097958087921, + 0.020982204005122185, + 0.07357936352491379 + ], + "credit_map": [ + 0.014692542143166065, + 0.030092285946011543, + 0.03631899505853653, + 0.026347843930125237, + 0.02509443648159504, + 0.021608954295516014, + 0.0264543779194355, + 0.017678234726190567, + 0.008530566468834877, + 0.010369769297540188, + 0.008787698112428188, + 0.01352701522409916, + 0.011997460387647152, + 0.012870602309703827, + 0.00891885720193386, + 0.009249866008758545, + 0.004129861947149038, + 0.003990747965872288, + 0.006641451269388199, + 0.004697452299296856, + 0.004150263499468565, + 0.004877307917922735, + 0.006034384481608868, + 0.006440714932978153, + 0.014196215197443962, + 0.014277871698141098, + 0.011374379508197308, + 0.013491779565811157, + 0.010476839728653431, + 0.010728996247053146, + 0.013324284926056862, + 0.02134748548269272, + 0.004553291015326977, + 0.004336608573794365, + 0.005036856047809124, + 0.007288672029972076, + 0.004920500330626965, + 0.0060677314177155495, + 0.006041013635694981, + 0.005165410693734884, + 0.004617427010089159, + 0.0034117670729756355, + 0.005022082943469286, + 0.0058452291414141655, + 0.0038120741955935955, + 0.004709732253104448, + 0.0030923911835998297, + 0.00483563682064414, + 0.0092538483440876, + 0.009240408428013325, + 0.014271102845668793, + 0.008971449919044971, + 0.012557782232761383, + 0.011031130328774452, + 0.008639751933515072, + 0.005387241952121258, + 0.030302496626973152, + 0.036636240780353546, + 0.025852488353848457, + 0.028539348393678665, + 0.023422623053193092, + 0.03300667554140091, + 0.04929279536008835, + 0.0875571072101593, + 0.0026545056607574224, + 0.0034970752894878387, + 0.0024483101442456245, + 0.003072715364396572, + 0.002281494438648224, + 0.002577185397967696, + 0.0021605028305202723, + 0.0022904134821146727, + 0.010092935524880886, + 0.014470083639025688, + 0.011297337710857391, + 0.0098361661657691, + 0.005903348792344332, + 0.010517997667193413, + 0.0075089819729328156, + 0.003952506929636002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.011841363273561 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective moving forward along a park path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style clean and minimalist with soft gradients and no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, pale-skinned hand entering from the lower right and extending to the center frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand offering a single branch with dark smooth wood and flowers having five petals in soft white and pale pink gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus bokeh of cherry trees creating a canopy of vibrant pink and white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005261858459562063 + ], + "opposite_logprob_tokens": [ + -5.2552618980407715 + ], + "logprob_reward": -0.005261858459562063, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005261858459562063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14671003818511963, + 0.06449726223945618, + 0.03582112118601799, + 0.06336139142513275, + 0.06629716604948044, + 0.24829727411270142, + 0.06112246960401535, + 0.08463874459266663, + 0.040299784392118454, + 0.18895475566387177 + ], + "credit_map": [ + 0.02196900174021721, + 0.017911653965711594, + 0.021980661898851395, + 0.01456650160253048, + 0.020895356312394142, + 0.01820574328303337, + 0.016649438068270683, + 0.014531678520143032, + 0.007154455408453941, + 0.007800114341080189, + 0.006913844496011734, + 0.008587323129177094, + 0.011561468243598938, + 0.009674697183072567, + 0.007237065117806196, + 0.005568290129303932, + 0.0035515681374818087, + 0.003444232977926731, + 0.004710945766419172, + 0.0065746610052883625, + 0.003934424836188555, + 0.0037632170133292675, + 0.005912294611334801, + 0.003929775208234787, + 0.009947645477950573, + 0.007728083059191704, + 0.005730347242206335, + 0.006080883089452982, + 0.005597254261374474, + 0.00719395512714982, + 0.005876195151358843, + 0.015207028016448021, + 0.0072165890596807, + 0.006199810653924942, + 0.00937832985073328, + 0.010654385201632977, + 0.007984932512044907, + 0.008057298138737679, + 0.00867047905921936, + 0.008135337382555008, + 0.029354432597756386, + 0.03345136344432831, + 0.03631157428026199, + 0.0300820954144001, + 0.030059143900871277, + 0.032614439725875854, + 0.025548186153173447, + 0.030876019969582558, + 0.008164508268237114, + 0.008292727172374725, + 0.01214288454502821, + 0.007886850275099277, + 0.00787339173257351, + 0.0066089872270822525, + 0.00541200116276741, + 0.004741122014820576, + 0.009850002825260162, + 0.009101535193622112, + 0.009722930379211903, + 0.009255238808691502, + 0.008716633543372154, + 0.010123088955879211, + 0.011075744405388832, + 0.016793571412563324, + 0.003960107918828726, + 0.0044976393692195415, + 0.0052406154572963715, + 0.006557919550687075, + 0.004371190909296274, + 0.005893723573535681, + 0.005106769502162933, + 0.0046718185767531395, + 0.028687555342912674, + 0.05093485116958618, + 0.02561323530972004, + 0.016242267563939095, + 0.013257184065878391, + 0.025886010378599167, + 0.013076622039079666, + 0.015257024206221104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.005261858459562063 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective moving forward along a park path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style clean and minimalist with soft gradients and no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, pale-skinned hand entering from the lower right and extending to the center frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand offering a single branch with dark smooth wood and flowers having five petals in soft white and pale pink gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus bokeh of cherry trees creating a canopy of vibrant pink and white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052522532641887665 + ], + "opposite_logprob_tokens": [ + -5.255252361297607 + ], + "logprob_reward": -0.0052522532641887665, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052522532641887665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06049728021025658, + 0.02656782977283001, + 0.04768848791718483, + 0.04742695018649101, + 0.052154939621686935, + 0.04530841112136841, + 0.12954890727996826, + 0.08005905896425247, + 0.026589244604110718, + 0.4841589033603668 + ], + "credit_map": [ + 0.006382794119417667, + 0.00803868006914854, + 0.007324810605496168, + 0.0063739363104105, + 0.008726789616048336, + 0.0067965323105454445, + 0.007223796099424362, + 0.009629939682781696, + 0.002834914019331336, + 0.003161047585308552, + 0.002671267371624708, + 0.0036653736606240273, + 0.004623440559953451, + 0.004041924607008696, + 0.0025638770312070847, + 0.00300598400644958, + 0.0038573527708649635, + 0.005153742153197527, + 0.007946758531033993, + 0.005307789426296949, + 0.004396412521600723, + 0.005602939520031214, + 0.0074268365278840065, + 0.007996659725904465, + 0.004650282673537731, + 0.00340161076746881, + 0.0062446147203445435, + 0.007045642472803593, + 0.0035489010624587536, + 0.004902590531855822, + 0.006408818531781435, + 0.011224488727748394, + 0.005907944403588772, + 0.005948189180344343, + 0.005560670979321003, + 0.008160645142197609, + 0.007324767764657736, + 0.007804613560438156, + 0.006053938530385494, + 0.0053941719233989716, + 0.004578339867293835, + 0.007436931598931551, + 0.0054798671044409275, + 0.005593740846961737, + 0.0066132196225225925, + 0.004683997482061386, + 0.004678701516240835, + 0.006243610754609108, + 0.01224515400826931, + 0.01308225467801094, + 0.021705014631152153, + 0.023700876161456108, + 0.015497866086661816, + 0.015253731049597263, + 0.016709325835108757, + 0.011354690417647362, + 0.009541036561131477, + 0.011639358475804329, + 0.012859937734901905, + 0.0081095015630126, + 0.006465097423642874, + 0.008884710259735584, + 0.010525180026888847, + 0.012034231796860695, + 0.00368809187784791, + 0.003882903838530183, + 0.002790337661281228, + 0.004029384348541498, + 0.00304594193585217, + 0.003958613146096468, + 0.003140383632853627, + 0.0020535876974463463, + 0.03992796689271927, + 0.06671254336833954, + 0.05480830371379852, + 0.056926336139440536, + 0.04015367850661278, + 0.04607250168919563, + 0.06574499607086182, + 0.11381258070468903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.0052522532641887665 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective moving forward along a park path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style clean and minimalist with soft gradients and no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, pale-skinned hand entering from the lower right and extending to the center frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand offering a single branch with dark smooth wood and flowers having five petals in soft white and pale pink gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus bokeh of cherry trees creating a canopy of vibrant pink and white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734294816851616 + ], + "opposite_logprob_tokens": [ + -5.006734371185303 + ], + "logprob_reward": -0.006734294816851616, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734294816851616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09933546185493469, + 0.06965792924165726, + 0.04130362346768379, + 0.06494428962469101, + 0.027140116319060326, + 0.06494775414466858, + 0.12285196036100388, + 0.10428840667009354, + 0.024112097918987274, + 0.38141828775405884 + ], + "credit_map": [ + 0.010358600877225399, + 0.016661042347550392, + 0.009543290361762047, + 0.01113953534513712, + 0.016956239938735962, + 0.014937274158000946, + 0.010984282940626144, + 0.008755190297961235, + 0.007338458206504583, + 0.007892871275544167, + 0.008292315527796745, + 0.009169499389827251, + 0.011786941438913345, + 0.010757071897387505, + 0.00735694682225585, + 0.00706382654607296, + 0.004222748335450888, + 0.003767959540709853, + 0.005182321649044752, + 0.00731634721159935, + 0.004415109287947416, + 0.0052191405557096004, + 0.004052724689245224, + 0.007127270568162203, + 0.007261309772729874, + 0.005232276860624552, + 0.00835226196795702, + 0.010623371228575706, + 0.005683817435055971, + 0.006583179347217083, + 0.007282570004463196, + 0.013925503008067608, + 0.003073800588026643, + 0.0032025063410401344, + 0.002835117978975177, + 0.0038005642127245665, + 0.004317224491387606, + 0.004325205460190773, + 0.0029858245979994535, + 0.0025998721830546856, + 0.008308080025017262, + 0.009655396454036236, + 0.008494972251355648, + 0.0103031350299716, + 0.009473963640630245, + 0.006537372712045908, + 0.005389528349041939, + 0.006785308010876179, + 0.014491073787212372, + 0.0169588103890419, + 0.019465051591396332, + 0.020296869799494743, + 0.01504876185208559, + 0.011273786425590515, + 0.01503538154065609, + 0.01028223056346178, + 0.01321104820817709, + 0.01885915733873844, + 0.011600730940699577, + 0.008958542719483376, + 0.008491814136505127, + 0.012471367605030537, + 0.014015138149261475, + 0.016680605709552765, + 0.0026030552107840776, + 0.0028585391119122505, + 0.004559409338980913, + 0.003217418910935521, + 0.002539967419579625, + 0.0037938691675662994, + 0.0021542636677622795, + 0.002385576255619526, + 0.055040519684553146, + 0.1056625247001648, + 0.054044585675001144, + 0.03562207520008087, + 0.025046274065971375, + 0.04605446383357048, + 0.029464123770594597, + 0.03048372082412243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.006734294816851616 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective moving forward along a park path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style clean and minimalist with soft gradients and no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, pale-skinned hand entering from the lower right and extending to the center frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand offering a single branch with dark smooth wood and flowers having five petals in soft white and pale pink gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus bokeh of cherry trees creating a canopy of vibrant pink and white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00091856240760535 + ], + "opposite_logprob_tokens": [ + -7.000918388366699 + ], + "logprob_reward": -0.00091856240760535, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.00091856240760535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18414080142974854, + 0.0960477963089943, + 0.06069561094045639, + 0.07733629643917084, + 0.04354582726955414, + 0.09136636555194855, + 0.0852200835943222, + 0.1189129650592804, + 0.02142592892050743, + 0.22130823135375977 + ], + "credit_map": [ + 0.020925527438521385, + 0.026976874098181725, + 0.01936257816851139, + 0.020500222221016884, + 0.03050864115357399, + 0.022657694295048714, + 0.02231799066066742, + 0.020891286432743073, + 0.009782757610082626, + 0.010691984556615353, + 0.011566892266273499, + 0.011428337544202805, + 0.01397785171866417, + 0.015391879715025425, + 0.009787741117179394, + 0.013420355506241322, + 0.006188431289047003, + 0.005431672558188438, + 0.007238419726490974, + 0.009460016153752804, + 0.005198907572776079, + 0.007433569990098476, + 0.0063749984838068485, + 0.013369598425924778, + 0.009198087267577648, + 0.009768028743565083, + 0.008968153037130833, + 0.014315997250378132, + 0.008729900233447552, + 0.007867393083870411, + 0.011352994479238987, + 0.007135747000575066, + 0.004407131113111973, + 0.004355958197265863, + 0.0050146072171628475, + 0.011202982626855373, + 0.004849375691264868, + 0.005209704395383596, + 0.004742361605167389, + 0.0037637080531567335, + 0.009887357242405415, + 0.017954058945178986, + 0.00904966052621603, + 0.008973892778158188, + 0.018245987594127655, + 0.010059475898742676, + 0.006364939734339714, + 0.010830996558070183, + 0.008838847279548645, + 0.010242022573947906, + 0.011121697723865509, + 0.014853360131382942, + 0.010754082351922989, + 0.01140819676220417, + 0.011202111840248108, + 0.006799774710088968, + 0.010567350313067436, + 0.011952068656682968, + 0.01821768656373024, + 0.012751429341733456, + 0.009950206615030766, + 0.010898038744926453, + 0.018667861819267273, + 0.0259083304554224, + 0.0027043053414672613, + 0.0022533803712576628, + 0.003014255315065384, + 0.0027957672718912363, + 0.0023350322153419256, + 0.0028577533084899187, + 0.0029062777757644653, + 0.002559158718213439, + 0.03563164174556732, + 0.05032772198319435, + 0.026751048862934113, + 0.027223115786910057, + 0.014650373719632626, + 0.02338802069425583, + 0.02457069419324398, + 0.018765628337860107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.00091856240760535 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a first-person perspective, an uninterrupted gliding shot moves smoothly forward along a park path, all rendered in a clean, minimalist art style defined by elegant, soft gradients. The air is thick with a dense, gentle mass of cherry blossom petals, which slowly fall like pink snow, carpeting the stone-paved path with a delicate, textured carpet. In the background, the soft-focus bokeh of countless other cherry trees creates a bright, all-encompassing canopy of vibrant pink and white. Diffuse, ethereal light filters through the blossoms, casting no harsh shadows and bathing the entire scene in a warm, tranquil glow. As the camera advances, a slender, pale-skinned hand enters the frame from the lower right, its form defined by smooth tonal transitions rather than sharp lines. Moving with deliberate grace, the hand extends forward until it reaches its center frame, offering the viewer a single, perfectly formed branch of cherry blossoms. The branch itself is a dark, smooth wood from which sprout several beautifully shaped flowers, with five delicate petals perfectly blending soft white and pale pink gradients, creating a peaceful, intimate and dreamy atmosphere.", + "seed": 8774, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_4/index_8086_seed_8774_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.009630965581163765, + "vlm_logprob": -0.009630965581163765, + "vlm_yes_no_margin": 5.025, + "vlm_token_credit": -0.009630965581163765, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective moving forward along a park path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style clean and minimalist with soft gradients and no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, pale-skinned hand entering from the lower right and extending to the center frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand offering a single branch with dark smooth wood and flowers having five petals in soft white and pale pink gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus bokeh of cherry trees creating a canopy of vibrant pink and white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011387943290174007 + ], + "opposite_logprob_tokens": [ + -5.386387825012207 + ], + "logprob_reward": -0.011387943290174007, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.011387943290174007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06835465878248215, + 0.10198836773633957, + 0.01957572251558304, + 0.08709754049777985, + 0.045593567192554474, + 0.022772420197725296, + 0.08411803841590881, + 0.06232599541544914, + 0.42858636379241943, + 0.07958739995956421 + ], + "credit_map": [ + 0.011824593879282475, + 0.01192636787891388, + 0.006310157012194395, + 0.006219618488103151, + 0.012435062788426876, + 0.008874842897057533, + 0.005689806770533323, + 0.005074207670986652, + 0.01617211103439331, + 0.015059982426464558, + 0.008793658576905727, + 0.01225789450109005, + 0.014040258713066578, + 0.017840709537267685, + 0.009976200759410858, + 0.007847539149224758, + 0.0027163969352841377, + 0.0034295576624572277, + 0.0011280120816081762, + 0.0013942169025540352, + 0.003615766065195203, + 0.004513536114245653, + 0.0013354902621358633, + 0.0014427455607801676, + 0.006226756144315004, + 0.025731805711984634, + 0.0037338894326239824, + 0.004933428950607777, + 0.03001931682229042, + 0.008270123973488808, + 0.004638553597033024, + 0.003543652594089508, + 0.006346928887069225, + 0.007433019578456879, + 0.0037627548445016146, + 0.006118665914982557, + 0.006211928091943264, + 0.006840186659246683, + 0.00466614356264472, + 0.0042139324359595776, + 0.002558706095442176, + 0.0037893326953053474, + 0.0017511692130938172, + 0.002209506696090102, + 0.004725979175418615, + 0.0033923282753676176, + 0.002300502499565482, + 0.00204489310272038, + 0.008041661232709885, + 0.011771110817790031, + 0.007821574807167053, + 0.008237148635089397, + 0.011763516813516617, + 0.009212030097842216, + 0.0206729955971241, + 0.006597989238798618, + 0.00641161622479558, + 0.008768904954195023, + 0.005517272278666496, + 0.006189730484038591, + 0.007858374156057835, + 0.009584209881722927, + 0.008407619781792164, + 0.009588263928890228, + 0.01656230352818966, + 0.16366948187351227, + 0.008253617212176323, + 0.016062842682003975, + 0.17360998690128326, + 0.019408641383051872, + 0.023167148232460022, + 0.007852302864193916, + 0.0067143989726901054, + 0.013567882589995861, + 0.009870834648609161, + 0.011568493209779263, + 0.00911510270088911, + 0.009681404568254948, + 0.010768495500087738, + 0.008300775662064552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 3 + ], + "selected_reward": -0.011387943290174007 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective moving forward along a park path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style clean and minimalist with soft gradients and no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, pale-skinned hand entering from the lower right and extending to the center frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand offering a single branch with dark smooth wood and flowers having five petals in soft white and pale pink gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus bokeh of cherry trees creating a canopy of vibrant pink and white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016085494309663773 + ], + "opposite_logprob_tokens": [ + -4.141085624694824 + ], + "logprob_reward": -0.016085494309663773, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016085494309663773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06637930124998093, + 0.07378675043582916, + 0.045775994658470154, + 0.29868465662002563, + 0.10275498777627945, + 0.02145366184413433, + 0.046469978988170624, + 0.06196599081158638, + 0.24305903911590576, + 0.039669666439294815 + ], + "credit_map": [ + 0.00923981238156557, + 0.007450847886502743, + 0.0058631193824112415, + 0.006899976171553135, + 0.010018638335168362, + 0.014464344829320908, + 0.0070718079805374146, + 0.0053707510232925415, + 0.010172407142817974, + 0.011308366432785988, + 0.00522583769634366, + 0.0079791359603405, + 0.014207134023308754, + 0.012712827883660793, + 0.0068206642754375935, + 0.005360373295843601, + 0.005948059260845184, + 0.007355809677392244, + 0.0031234589405357838, + 0.004325440619140863, + 0.008327649906277657, + 0.009316856041550636, + 0.0042170751839876175, + 0.003161643398925662, + 0.019473930820822716, + 0.0903756320476532, + 0.015384625643491745, + 0.01672506332397461, + 0.10213603079319, + 0.028248168528079987, + 0.01399318128824234, + 0.012347990646958351, + 0.013284450396895409, + 0.01447258424013853, + 0.00885235145688057, + 0.013358402997255325, + 0.014614909887313843, + 0.016517581418156624, + 0.011587953194975853, + 0.010066751390695572, + 0.0021333605982363224, + 0.003110507968813181, + 0.0017277635633945465, + 0.0025568308774381876, + 0.0034897520672529936, + 0.0034501433838158846, + 0.0026752715930342674, + 0.002310031559318304, + 0.0054008448496460915, + 0.006991638336330652, + 0.0039880117401480675, + 0.005758127197623253, + 0.007098859176039696, + 0.006245053838938475, + 0.008287579752504826, + 0.002699862001463771, + 0.005901991855353117, + 0.00584095623344183, + 0.006888484116643667, + 0.007021598517894745, + 0.00762456888332963, + 0.008844776079058647, + 0.010004078969359398, + 0.009839530102908611, + 0.008934061042964458, + 0.09727208316326141, + 0.0044488683342933655, + 0.006978067569434643, + 0.10235913097858429, + 0.012362979352474213, + 0.006977259647101164, + 0.00372658995911479, + 0.0035245493054389954, + 0.0069631957449018955, + 0.004433955531567335, + 0.003916546236723661, + 0.0043758824467659, + 0.003783284919336438, + 0.006133269052952528, + 0.006538979709148407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 4 + ], + "selected_reward": -0.016085494309663773 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective moving forward along a park path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style clean and minimalist with soft gradients and no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, pale-skinned hand entering from the lower right and extending to the center frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand offering a single branch with dark smooth wood and flowers having five petals in soft white and pale pink gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus bokeh of cherry trees creating a canopy of vibrant pink and white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948578007519245 + ], + "opposite_logprob_tokens": [ + -5.130948543548584 + ], + "logprob_reward": -0.005948578007519245, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948578007519245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02817131206393242, + 0.03732423856854439, + 0.02311938814818859, + 0.04351675882935524, + 0.010312918573617935, + 0.03808527812361717, + 0.03177464380860329, + 0.11095210909843445, + 0.6243100166320801, + 0.052433352917432785 + ], + "credit_map": [ + 0.004134383983910084, + 0.0036368886940181255, + 0.0028316047973930836, + 0.002944657579064369, + 0.0049155461601912975, + 0.0036240832414478064, + 0.00364410481415689, + 0.0024400425609201193, + 0.005468098912388086, + 0.005369275342673063, + 0.0027239841874688864, + 0.004437911324203014, + 0.005557905416935682, + 0.006600714288651943, + 0.0038925642147660255, + 0.0032737820874899626, + 0.0031921048648655415, + 0.003995946608483791, + 0.0012339763343334198, + 0.0011459406232461333, + 0.0047003296203911304, + 0.005882401950657368, + 0.0012729460140690207, + 0.001695743529126048, + 0.001679160981439054, + 0.014983760192990303, + 0.0013883194187656045, + 0.002180458279326558, + 0.016860010102391243, + 0.002656659809872508, + 0.0022210648749023676, + 0.0015473278472200036, + 0.0017960070399567485, + 0.001122807152569294, + 0.0009246801491826773, + 0.0010537689086049795, + 0.0019146964186802506, + 0.0016238173702731729, + 0.0011747295502573252, + 0.00070241215871647, + 0.005977229680866003, + 0.004973472096025944, + 0.005372049752622843, + 0.004744088277220726, + 0.00432467320933938, + 0.0059272898361086845, + 0.0037462040781974792, + 0.0030202711932361126, + 0.003737963503226638, + 0.0038733119145035744, + 0.003447445807978511, + 0.0030649930704385042, + 0.0038092697504907846, + 0.004362291190773249, + 0.0069839833304286, + 0.0024953836109489202, + 0.007784736808389425, + 0.01023365929722786, + 0.012379723601043224, + 0.011616870760917664, + 0.01125440001487732, + 0.01829117350280285, + 0.01568569801747799, + 0.023705843836069107, + 0.024993348866701126, + 0.23682166635990143, + 0.01102903950959444, + 0.026166943833231926, + 0.25986629724502563, + 0.03131263703107834, + 0.022666525095701218, + 0.011453550308942795, + 0.005423375871032476, + 0.008045523427426815, + 0.006161695811897516, + 0.005448322743177414, + 0.007870498113334179, + 0.008661879226565361, + 0.006170715671032667, + 0.004651341587305069 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.005948578007519245 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective moving forward along a park path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style clean and minimalist with soft gradients and no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, pale-skinned hand entering from the lower right and extending to the center frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand offering a single branch with dark smooth wood and flowers having five petals in soft white and pale pink gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus bokeh of cherry trees creating a canopy of vibrant pink and white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536648660898209 + ], + "opposite_logprob_tokens": [ + -4.387536525726318 + ], + "logprob_reward": -0.012536648660898209, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536648660898209, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05704014375805855, + 0.016410449519753456, + 0.018620220944285393, + 0.09171491116285324, + 0.01858915016055107, + 0.01186234038323164, + 0.023548614233732224, + 0.021740619093179703, + 0.6277198791503906, + 0.1127537414431572 + ], + "credit_map": [ + 0.009976757690310478, + 0.0073226578533649445, + 0.004967020358890295, + 0.006947297137230635, + 0.007861900143325329, + 0.009528040885925293, + 0.006356069352477789, + 0.004080400802195072, + 0.00268555642105639, + 0.00226701726205647, + 0.0014887963188812137, + 0.0018727162387222052, + 0.002323173452168703, + 0.002865178044885397, + 0.001582033233717084, + 0.0013259798288345337, + 0.002155829453840852, + 0.0036412603221833706, + 0.0011166402837261558, + 0.0017137039685621858, + 0.004303707275539637, + 0.0026020328514277935, + 0.002056529512628913, + 0.0010305170435458422, + 0.0042222184129059315, + 0.03594718873500824, + 0.0021430093329399824, + 0.0021522094029933214, + 0.03818061202764511, + 0.004728881176561117, + 0.002479949500411749, + 0.0018608475802466273, + 0.003385110991075635, + 0.0018458121921867132, + 0.0019772937521338463, + 0.0018002779688686132, + 0.003367332275956869, + 0.0029131791088730097, + 0.0018746500136330724, + 0.0014254952548071742, + 0.0014542518183588982, + 0.001623075339011848, + 0.0014252191176638007, + 0.0015588372480124235, + 0.0019770353101193905, + 0.0016402307664975524, + 0.001238802564330399, + 0.0009448884520679712, + 0.003066318342462182, + 0.0032432852312922478, + 0.002797007793560624, + 0.002226494252681732, + 0.004696765914559364, + 0.0030451114289462566, + 0.0020059330854564905, + 0.0024676986504346132, + 0.003551542293280363, + 0.0023671092931181192, + 0.002097728196531534, + 0.0029775535222142935, + 0.003733251243829727, + 0.002788953948765993, + 0.0026147260796278715, + 0.0016097560292109847, + 0.024833057075738907, + 0.24157054722309113, + 0.011677710339426994, + 0.021595872938632965, + 0.26183485984802246, + 0.03133704885840416, + 0.021307222545146942, + 0.013563542626798153, + 0.01085306704044342, + 0.019422011449933052, + 0.008654077537357807, + 0.009952130727469921, + 0.012506134808063507, + 0.00981493853032589, + 0.020994162186980247, + 0.02055721916258335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.012536648660898209 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective moving forward along a park path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style clean and minimalist with soft gradients and no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, pale-skinned hand entering from the lower right and extending to the center frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand offering a single branch with dark smooth wood and flowers having five petals in soft white and pale pink gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus bokeh of cherry trees creating a canopy of vibrant pink and white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021961636375635862 + ], + "opposite_logprob_tokens": [ + -6.127196311950684 + ], + "logprob_reward": -0.0021961636375635862, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021961636375635862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03379860892891884, + 0.013053961098194122, + 0.047635842114686966, + 0.6919940114021301, + 0.03356800973415375, + 0.06419594585895538, + 0.019464809447526932, + 0.0161434393376112, + 0.06819485127925873, + 0.011950583197176456 + ], + "credit_map": [ + 0.0036964083556085825, + 0.00661322521045804, + 0.002555555198341608, + 0.0028015263378620148, + 0.007419582922011614, + 0.004921354819089174, + 0.0033237445168197155, + 0.0024672066792845726, + 0.0023719933815300465, + 0.002042626729235053, + 0.0010799923911690712, + 0.0010407465742900968, + 0.00241580861620605, + 0.001811670488677919, + 0.001336812973022461, + 0.0009543077321723104, + 0.004706105217337608, + 0.01080762967467308, + 0.0022879624739289284, + 0.002928454428911209, + 0.013047287240624428, + 0.008083222433924675, + 0.003035645466297865, + 0.002739527728408575, + 0.025101646780967712, + 0.26535582542419434, + 0.01352097187191248, + 0.012500971555709839, + 0.29663795232772827, + 0.04815738648176193, + 0.016659213230013847, + 0.014059926383197308, + 0.0044030966237187386, + 0.004766723606735468, + 0.003017784794792533, + 0.004935643635690212, + 0.004123851656913757, + 0.005544996354728937, + 0.003785344772040844, + 0.0029905647970736027, + 0.006060597486793995, + 0.01369755994528532, + 0.004653769079595804, + 0.00672184070572257, + 0.012295524589717388, + 0.009359450079500675, + 0.006762731354683638, + 0.004644464701414108, + 0.0019858339801430702, + 0.002925540553405881, + 0.0019758245907723904, + 0.0021843609865754843, + 0.002817577449604869, + 0.002265963237732649, + 0.003829432185739279, + 0.0014802751829847693, + 0.0014622054295614362, + 0.0017979848198592663, + 0.0020979722030460835, + 0.002030672039836645, + 0.002055468736216426, + 0.00249696453101933, + 0.0023394476156681776, + 0.0018627223325893283, + 0.0029959012754261494, + 0.025384599342942238, + 0.001792581402696669, + 0.003944075666368008, + 0.026042629033327103, + 0.003376347478479147, + 0.002668763045221567, + 0.0019899490289390087, + 0.0011207498610019684, + 0.0019378314027562737, + 0.0010787370847538114, + 0.0013019217876717448, + 0.0014963801950216293, + 0.0012555195717141032, + 0.0023159021511673927, + 0.0014435393968597054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 5 + ], + "selected_reward": -0.0021961636375635862 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_5/index_tensor([2550, 2550])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_5/index_tensor([2550, 2550])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6b63c83bfcfe506f82e9338a10e5ef0317f5a70d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_5/index_tensor([2550, 2550])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, low-angle, wide-angle shot captures a vast field of tall, sun-bleached cornstalks, flapping violently in a chaotic frenzy. A massive dust storm, a billowing wall of thick ochre dust, relentlessly engulfs the landscape, swallowing the horizon and creating an oppressive, claustrophobic atmosphere. A soft, directional light, like that from a dying sun, penetrates the dense cloud of particles, creating dramatic god rays that illuminate millions of individually rendered dust and sand particles dancing in the air. This intense light casts long, stark shadows that stretch and writhe across the ground, highlighting the corn's violent movement. Each stalk is rendered in a hyper-realistic Unreal Engine style, displaying realistic, high-fidelity textures; their dry, brittle leaves are pale yellow and tattered, whipping back and forth with such force that they shred and tear. The entire scene is a symphony of controlled chaos, with cinematic lighting and physically-based rendering emphasizing the oppressive power of nature and the fragility of the harvest.", + "seed": 24564, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_5/index_2550_seed_24564_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1196698123589158, + "vlm_logprob": -0.1196698123589158, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.1196698123589158, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a wide-angle shot of a vast field of tall cornstalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dust storm depicted as a billowing wall of thick ochre dust engulfing the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are long, stark shadows cast by a soft directional light creating god rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cornstalks rendered in a hyper-realistic Unreal Engine style with pale yellow, tattered leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2507864236831665 + ], + "opposite_logprob_tokens": [ + -1.7507864236831665 + ], + "logprob_reward": -0.2507864236831665, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2507864236831665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3870178759098053, + 0.03343651443719864, + 0.025605546310544014, + 0.17380069196224213, + 0.1118917167186737, + 0.013533815741539001, + 0.03611970320343971, + 0.023742562159895897, + 0.0510440468788147, + 0.1438075751066208 + ], + "credit_map": [ + 0.030352484434843063, + 0.05510726571083069, + 0.05064509063959122, + 0.07148368656635284, + 0.05371968075633049, + 0.0440843291580677, + 0.049136336892843246, + 0.03248898684978485, + 0.003600166877731681, + 0.005208812188357115, + 0.004281712230294943, + 0.003828906686976552, + 0.00479245325550437, + 0.004451514221727848, + 0.004202987998723984, + 0.0030699593480676413, + 0.002639995189383626, + 0.0036614667624235153, + 0.003296054434031248, + 0.0035747752990573645, + 0.0032650181092321873, + 0.0039733960293233395, + 0.002341720275580883, + 0.0028531188145279884, + 0.007109691854566336, + 0.007737990003079176, + 0.01960645243525505, + 0.01639147475361824, + 0.040690407156944275, + 0.00614585354924202, + 0.06453795731067657, + 0.011580867692828178, + 0.006878050044178963, + 0.006221294403076172, + 0.009115943685173988, + 0.009563819505274296, + 0.02755279839038849, + 0.005545792169868946, + 0.0390046201646328, + 0.008009406737983227, + 0.0013842051848769188, + 0.0024249469861388206, + 0.0016445302171632648, + 0.0024810347240418196, + 0.001200446393340826, + 0.0010731251677498221, + 0.002113752067089081, + 0.0012117758160457015, + 0.0031627025455236435, + 0.004995335359126329, + 0.0033310074359178543, + 0.004442763514816761, + 0.004104818217456341, + 0.004616730380803347, + 0.008065704256296158, + 0.00340064219199121, + 0.00186559883877635, + 0.005678466986864805, + 0.003091692691668868, + 0.0027590394020080566, + 0.0024679084308445454, + 0.0027452223002910614, + 0.003091673832386732, + 0.0020429606083780527, + 0.0031941444613039494, + 0.008605941198766232, + 0.008869246579706669, + 0.008185552433133125, + 0.004806916695088148, + 0.006310042925179005, + 0.006290800869464874, + 0.004781400319188833, + 0.009005770087242126, + 0.019871331751346588, + 0.02152242138981819, + 0.022653764113783836, + 0.01781505160033703, + 0.01504821702837944, + 0.023324348032474518, + 0.014566680416464806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.2507864236831665 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a wide-angle shot of a vast field of tall cornstalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dust storm depicted as a billowing wall of thick ochre dust engulfing the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are long, stark shadows cast by a soft directional light creating god rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cornstalks rendered in a hyper-realistic Unreal Engine style with pale yellow, tattered leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048692476004362106 + ], + "opposite_logprob_tokens": [ + -3.048692464828491 + ], + "logprob_reward": -0.048692476004362106, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048692476004362106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1682339459657669, + 0.02387434057891369, + 0.018016742542386055, + 0.45928382873535156, + 0.08015971630811691, + 0.015163226053118706, + 0.03757726773619652, + 0.011048618704080582, + 0.05359239503741264, + 0.13304993510246277 + ], + "credit_map": [ + 0.011734414845705032, + 0.03053528442978859, + 0.020648496225476265, + 0.024915756657719612, + 0.023919088765978813, + 0.019658606499433517, + 0.02049342542886734, + 0.016328878700733185, + 0.0023520656395703554, + 0.0027775128837674856, + 0.002995397662743926, + 0.003672150894999504, + 0.0035656699910759926, + 0.0031063882634043694, + 0.0034789317287504673, + 0.0019262252608314157, + 0.00203179894015193, + 0.0020900038070976734, + 0.0022487749811261892, + 0.0024679661728441715, + 0.002219487214460969, + 0.00248745852150023, + 0.0023022640962153673, + 0.002168988576158881, + 0.018659522756934166, + 0.021291585639119148, + 0.048280779272317886, + 0.04186594486236572, + 0.11395808309316635, + 0.01687508448958397, + 0.16819430887699127, + 0.030158527195453644, + 0.004135737195611, + 0.003980762790888548, + 0.00752645218744874, + 0.008896290324628353, + 0.01380649209022522, + 0.00259202066808939, + 0.031313344836235046, + 0.007908613421022892, + 0.0018283044919371605, + 0.0027772679459303617, + 0.0020502281840890646, + 0.0018919029971584678, + 0.0014246396021917462, + 0.0015800442779436707, + 0.002159861149266362, + 0.001450977986678481, + 0.0031089854892343283, + 0.005671248305588961, + 0.0039497059769928455, + 0.00481144106015563, + 0.004540831316262484, + 0.004301995504647493, + 0.007088895421475172, + 0.004104166757315397, + 0.00082253513392061, + 0.002719148527830839, + 0.001397541374899447, + 0.0016478878678753972, + 0.0010566761484369636, + 0.0012481299927458167, + 0.0011474615894258022, + 0.0010092384181916714, + 0.0033975420519709587, + 0.010188689455389977, + 0.007408092729747295, + 0.0081725362688303, + 0.006030467804521322, + 0.007701350376009941, + 0.0054296390153467655, + 0.0052640726789832115, + 0.008395805023610592, + 0.021434951573610306, + 0.019385408610105515, + 0.020208515226840973, + 0.015287032350897789, + 0.014196613803505898, + 0.02038227766752243, + 0.013759332709014416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.048692476004362106 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a wide-angle shot of a vast field of tall cornstalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dust storm depicted as a billowing wall of thick ochre dust engulfing the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are long, stark shadows cast by a soft directional light creating god rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cornstalks rendered in a hyper-realistic Unreal Engine style with pale yellow, tattered leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005959124304354191 + ], + "opposite_logprob_tokens": [ + -5.1309590339660645 + ], + "logprob_reward": -0.005959124304354191, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005959124304354191, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26476114988327026, + 0.054898519068956375, + 0.029968345537781715, + 0.43214771151542664, + 0.056424111127853394, + 0.022019173949956894, + 0.021186785772442818, + 0.011013942770659924, + 0.03406617417931557, + 0.07351412624120712 + ], + "credit_map": [ + 0.02083822712302208, + 0.043347131460905075, + 0.03381387144327164, + 0.04193245619535446, + 0.03951627388596535, + 0.02567450702190399, + 0.033976394683122635, + 0.025662273168563843, + 0.005254258867353201, + 0.00926334410905838, + 0.0069629899226129055, + 0.007238492369651794, + 0.00808656308799982, + 0.006077018566429615, + 0.006855602376163006, + 0.005160247441381216, + 0.0022854269482195377, + 0.005097608547657728, + 0.0039880163967609406, + 0.004717261530458927, + 0.003943101968616247, + 0.004379944875836372, + 0.003137461841106415, + 0.0024195245932787657, + 0.01724395900964737, + 0.02015572041273117, + 0.04528754577040672, + 0.039812829345464706, + 0.10697836428880692, + 0.015642572194337845, + 0.15852749347686768, + 0.028499189764261246, + 0.003678564913570881, + 0.003583083162084222, + 0.0034849164076149464, + 0.004940848797559738, + 0.014130042865872383, + 0.002870480762794614, + 0.01859918050467968, + 0.005136994644999504, + 0.0024676553439348936, + 0.002967192092910409, + 0.003359810449182987, + 0.002403952181339264, + 0.0017634113319218159, + 0.0021729187574237585, + 0.0045881434343755245, + 0.0022960901260375977, + 0.0012786738807335496, + 0.0017171504441648722, + 0.004561838693916798, + 0.001960441004484892, + 0.0017097425879910588, + 0.002957182005047798, + 0.005248243920505047, + 0.0017535114893689752, + 0.0006412930670194328, + 0.0017639934085309505, + 0.0024264641106128693, + 0.0015200928319245577, + 0.0010431729024276137, + 0.0012717470526695251, + 0.001366169541142881, + 0.0009810106130316854, + 0.0020020105876028538, + 0.004676674958318472, + 0.009129790589213371, + 0.0036194685380905867, + 0.002472985303029418, + 0.002588118426501751, + 0.007288087159395218, + 0.0022890379186719656, + 0.004772300366312265, + 0.011769239790737629, + 0.011061688885092735, + 0.0107030738145113, + 0.008583870716392994, + 0.007110279984772205, + 0.011178781278431416, + 0.008334890007972717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.005959124304354191 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a wide-angle shot of a vast field of tall cornstalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dust storm depicted as a billowing wall of thick ochre dust engulfing the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are long, stark shadows cast by a soft directional light creating god rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cornstalks rendered in a hyper-realistic Unreal Engine style with pale yellow, tattered leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296482384204865 + ], + "opposite_logprob_tokens": [ + -2.2379648685455322 + ], + "logprob_reward": -0.11296482384204865, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296482384204865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2156456708908081, + 0.06551545113325119, + 0.14797163009643555, + 0.2591678500175476, + 0.0897110253572464, + 0.04011433199048042, + 0.05468638241291046, + 0.022111989557743073, + 0.06288553774356842, + 0.04219017177820206 + ], + "credit_map": [ + 0.016518503427505493, + 0.0336570218205452, + 0.023590482771396637, + 0.04032762348651886, + 0.028664354234933853, + 0.024817518889904022, + 0.026604866608977318, + 0.021465294063091278, + 0.004881966859102249, + 0.011820724233984947, + 0.00822449941188097, + 0.008547228761017323, + 0.010986752808094025, + 0.009757456369698048, + 0.005917030852288008, + 0.005379785317927599, + 0.015513229183852673, + 0.022962957620620728, + 0.0207584947347641, + 0.01681148260831833, + 0.02440258301794529, + 0.018759513273835182, + 0.013968921266496181, + 0.014794430695474148, + 0.01002972386777401, + 0.011255624704062939, + 0.0298200324177742, + 0.02485148049890995, + 0.058091968297958374, + 0.009195358492434025, + 0.09739214181900024, + 0.018531514331698418, + 0.0057594855315983295, + 0.008922787383198738, + 0.009525010362267494, + 0.007085131015628576, + 0.023489611223340034, + 0.006550043821334839, + 0.018941665068268776, + 0.009437277913093567, + 0.004806485027074814, + 0.007524184882640839, + 0.005426459014415741, + 0.0048566642217338085, + 0.003304602112621069, + 0.003942402545362711, + 0.006716062780469656, + 0.003537464886903763, + 0.004444795195013285, + 0.007996276021003723, + 0.0073037766851484776, + 0.005677934270352125, + 0.006232193671166897, + 0.007107211742550135, + 0.00916439201682806, + 0.006759800016880035, + 0.0017933578928932548, + 0.00534388842061162, + 0.0026590635534375906, + 0.0023895150516182184, + 0.0021812475752085447, + 0.0024445706512778997, + 0.002700828481465578, + 0.0025995178148150444, + 0.0040622903034091, + 0.009808839298784733, + 0.00942317210137844, + 0.010378135368227959, + 0.005899231415241957, + 0.006394815165549517, + 0.01133103296160698, + 0.005588015075773001, + 0.0030391435138881207, + 0.00656869076192379, + 0.005528404377400875, + 0.007178527768701315, + 0.004651803523302078, + 0.004692404065281153, + 0.006741948425769806, + 0.0037892453838139772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.11296482384204865 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a wide-angle shot of a vast field of tall cornstalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dust storm depicted as a billowing wall of thick ochre dust engulfing the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are long, stark shadows cast by a soft directional light creating god rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cornstalks rendered in a hyper-realistic Unreal Engine style with pale yellow, tattered leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17994621396064758 + ], + "opposite_logprob_tokens": [ + -1.8049461841583252 + ], + "logprob_reward": -0.17994621396064758, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17994621396064758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13932079076766968, + 0.019841063767671585, + 0.050582412630319595, + 0.3792664706707001, + 0.27457278966903687, + 0.056178655475378036, + 0.010617496445775032, + 0.019237525761127472, + 0.03300749510526657, + 0.01737525686621666 + ], + "credit_map": [ + 0.010536384768784046, + 0.024270806461572647, + 0.018055981025099754, + 0.022588100284337997, + 0.019011318683624268, + 0.014975517056882381, + 0.017723066732287407, + 0.012159626930952072, + 0.0016001170733943582, + 0.00256944028660655, + 0.0027309292927384377, + 0.0035433287266641855, + 0.0028334634844213724, + 0.0017260574968531728, + 0.0030655767768621445, + 0.0017721491167321801, + 0.005971100181341171, + 0.005932866130024195, + 0.009069406427443027, + 0.0058924066834151745, + 0.004603334236890078, + 0.003687082091346383, + 0.008678934536874294, + 0.006747287232428789, + 0.015308612026274204, + 0.017484748736023903, + 0.039851583540439606, + 0.034979529678821564, + 0.09384448826313019, + 0.01387888565659523, + 0.1390567123889923, + 0.02486192248761654, + 0.01099423412233591, + 0.011742966249585152, + 0.019351737573742867, + 0.028323298320174217, + 0.06724783033132553, + 0.009863343089818954, + 0.10433590412139893, + 0.022713463753461838, + 0.005127832759171724, + 0.007309191860258579, + 0.007857955992221832, + 0.0067010801285505295, + 0.005192478187382221, + 0.005601593758910894, + 0.012904238887131214, + 0.005484283436089754, + 0.0010141602251678705, + 0.0016710059717297554, + 0.0013044329825788736, + 0.0011458182707428932, + 0.0011222752509638667, + 0.0013306388864293694, + 0.0016022579511627555, + 0.001426906674169004, + 0.001523316721431911, + 0.0028222945984452963, + 0.004773355089128017, + 0.0028282315470278263, + 0.0012063265312463045, + 0.001367818797007203, + 0.003175540594384074, + 0.0015406417660415173, + 0.0015338377561420202, + 0.0023023791145533323, + 0.00994475930929184, + 0.002849737647920847, + 0.001935853622853756, + 0.0015853379154577851, + 0.010647870600223541, + 0.0022077197209000587, + 0.0012096050195395947, + 0.0032281146850436926, + 0.0021849432960152626, + 0.0024241474457085133, + 0.0021185765508562326, + 0.0022201661486178637, + 0.0024039747659116983, + 0.0015857293037697673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.17994621396064758 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot captures a vast field of tall, sun-bleached cornstalks, flapping violently in a chaotic frenzy. A massive dust storm, a billowing wall of thick ochre dust, relentlessly engulfs the landscape, swallowing the horizon and creating an oppressive, claustrophobic atmosphere. A soft, directional light, like that from a dying sun, penetrates the dense cloud of particles, creating dramatic god rays that illuminate millions of individually rendered dust and sand particles dancing in the air. This intense light casts long, stark shadows that stretch and writhe across the ground, highlighting the corn's violent movement. Each stalk is rendered in a hyper-realistic Unreal Engine style, displaying realistic, high-fidelity textures; their dry, brittle leaves are pale yellow and tattered, whipping back and forth with such force that they shred and tear. The entire scene is a symphony of controlled chaos, with cinematic lighting and physically-based rendering emphasizing the oppressive power of nature and the fragility of the harvest.", + "seed": 24565, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_5/index_2550_seed_24565_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.16045581102371215, + "vlm_logprob": -0.16045581102371215, + "vlm_yes_no_margin": 2.274999976158142, + "vlm_token_credit": -0.16045581102371215, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a wide-angle shot of a vast field of tall cornstalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dust storm depicted as a billowing wall of thick ochre dust engulfing the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are long, stark shadows cast by a soft directional light creating god rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cornstalks rendered in a hyper-realistic Unreal Engine style with pale yellow, tattered leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.19398844242095947 + ], + "opposite_logprob_tokens": [ + -2.06898832321167 + ], + "logprob_reward": -0.19398844242095947, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.19398844242095947, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24208766222000122, + 0.046894390136003494, + 0.06158864125609398, + 0.07284867763519287, + 0.02592630311846733, + 0.10746405273675919, + 0.13333925604820251, + 0.2465343326330185, + 0.04389297589659691, + 0.01942369155585766 + ], + "credit_map": [ + 0.01744907721877098, + 0.028980396687984467, + 0.042576372623443604, + 0.036295950412750244, + 0.026724928990006447, + 0.038603875786066055, + 0.03004056215286255, + 0.021416492760181427, + 0.0031645530834794044, + 0.003555421018972993, + 0.005902338307350874, + 0.006900675129145384, + 0.004383524414151907, + 0.011110182851552963, + 0.006128411274403334, + 0.005749284755438566, + 0.004336815793067217, + 0.004789875820279121, + 0.01089499332010746, + 0.008128113113343716, + 0.008039223961532116, + 0.006752374116331339, + 0.006978225894272327, + 0.011669020168483257, + 0.0031510177068412304, + 0.0047868723049759865, + 0.006322572007775307, + 0.006939695216715336, + 0.009420063346624374, + 0.016030140221118927, + 0.008846466429531574, + 0.017351849004626274, + 0.002407393418252468, + 0.003170317504554987, + 0.00248975632712245, + 0.0032839544583112, + 0.005595307797193527, + 0.0026923632249236107, + 0.0023931546602398157, + 0.0038940580561757088, + 0.005423045717179775, + 0.02041422203183174, + 0.007842770777642727, + 0.014294186607003212, + 0.02755373902618885, + 0.01189652644097805, + 0.009144232608377934, + 0.010895327664911747, + 0.009544172324240208, + 0.02642764151096344, + 0.02122531644999981, + 0.021446440368890762, + 0.018841702491044998, + 0.009612091816961765, + 0.013862787745893002, + 0.012379099614918232, + 0.006287658121436834, + 0.09569230675697327, + 0.00457820063456893, + 0.009395184926688671, + 0.10536409169435501, + 0.01444084569811821, + 0.006804430857300758, + 0.003971615806221962, + 0.002788687590509653, + 0.008331282995641232, + 0.005560359917581081, + 0.007539748214185238, + 0.0068178558722138405, + 0.004266245756298304, + 0.00409880792722106, + 0.004489983897656202, + 0.0009444482275284827, + 0.002165940124541521, + 0.0024949030485004187, + 0.00282183475792408, + 0.004066140856593847, + 0.0015482158632948995, + 0.001882562879472971, + 0.0034996457397937775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.19398844242095947 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a wide-angle shot of a vast field of tall cornstalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dust storm depicted as a billowing wall of thick ochre dust engulfing the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are long, stark shadows cast by a soft directional light creating god rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cornstalks rendered in a hyper-realistic Unreal Engine style with pale yellow, tattered leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054989613592624664 + ], + "opposite_logprob_tokens": [ + -2.9299895763397217 + ], + "logprob_reward": -0.054989613592624664, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054989613592624664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33058103919029236, + 0.09544575959444046, + 0.06308965384960175, + 0.15588606894016266, + 0.03771166503429413, + 0.07052341848611832, + 0.09183913469314575, + 0.09339050948619843, + 0.03602178767323494, + 0.025510963052511215 + ], + "credit_map": [ + 0.025996074080467224, + 0.05741732195019722, + 0.03897804021835327, + 0.03555276617407799, + 0.0465017631649971, + 0.043186724185943604, + 0.04678252339363098, + 0.03616582974791527, + 0.007943459786474705, + 0.010973273776471615, + 0.011564700864255428, + 0.0150759257376194, + 0.009453666396439075, + 0.013867836445569992, + 0.01338259968906641, + 0.013184288516640663, + 0.004210858140140772, + 0.005888620391488075, + 0.008519253693521023, + 0.009634487330913544, + 0.01002227608114481, + 0.006064954679459333, + 0.009576494805514812, + 0.00917270127683878, + 0.0028702456038445234, + 0.013286799192428589, + 0.005080912262201309, + 0.004833451006561518, + 0.06028424948453903, + 0.05555527284741402, + 0.0068932995200157166, + 0.007081829011440277, + 0.002371501177549362, + 0.0036063697189092636, + 0.0051251244731247425, + 0.009020044468343258, + 0.00439039571210742, + 0.003819369478151202, + 0.005632875487208366, + 0.003745982889086008, + 0.004086620640009642, + 0.008984418585896492, + 0.00479491101577878, + 0.010981129482388496, + 0.01751577854156494, + 0.0077251712791621685, + 0.007830909453332424, + 0.008604471571743488, + 0.004409688524901867, + 0.014843546785414219, + 0.01396110188215971, + 0.017751293256878853, + 0.010814860463142395, + 0.006742116529494524, + 0.014371023513376713, + 0.008945493958890438, + 0.0031324555166065693, + 0.03071429952979088, + 0.004622101783752441, + 0.00980496034026146, + 0.029639404267072678, + 0.0046561891213059425, + 0.004671177361160517, + 0.006149911787360907, + 0.0025413483381271362, + 0.008940580300986767, + 0.005119482055306435, + 0.004671470262110233, + 0.005953793879598379, + 0.002899136859923601, + 0.002581879263743758, + 0.0033140925224870443, + 0.0017289271345362067, + 0.003928383346647024, + 0.004121711011976004, + 0.004642724525183439, + 0.0026206274051219225, + 0.0020653533283621073, + 0.0020890841260552406, + 0.004314151592552662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.054989613592624664 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a wide-angle shot of a vast field of tall cornstalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dust storm depicted as a billowing wall of thick ochre dust engulfing the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are long, stark shadows cast by a soft directional light creating god rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cornstalks rendered in a hyper-realistic Unreal Engine style with pale yellow, tattered leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014199413359165192 + ], + "opposite_logprob_tokens": [ + -4.264199256896973 + ], + "logprob_reward": -0.014199413359165192, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014199413359165192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19826865196228027, + 0.15211842954158783, + 0.05220876261591911, + 0.15174375474452972, + 0.02920995093882084, + 0.07536743581295013, + 0.044381096959114075, + 0.1790449172258377, + 0.10624945908784866, + 0.011407519690692425 + ], + "credit_map": [ + 0.016586339101195335, + 0.03940173238515854, + 0.020785775035619736, + 0.019122112542390823, + 0.03202106058597565, + 0.025326872244477272, + 0.022118927910923958, + 0.02290583960711956, + 0.011183305643498898, + 0.015796206891536713, + 0.015211185440421104, + 0.019070899114012718, + 0.019995860755443573, + 0.03690351918339729, + 0.01822194829583168, + 0.015735501423478127, + 0.006648921873420477, + 0.004515817854553461, + 0.007857326418161392, + 0.007924996316432953, + 0.005299078766256571, + 0.00601368909701705, + 0.006905360147356987, + 0.007043574936687946, + 0.003953434061259031, + 0.012410266324877739, + 0.005574648268520832, + 0.0035685452166944742, + 0.05820363759994507, + 0.054097242653369904, + 0.006350320763885975, + 0.007585656363517046, + 0.001671963487751782, + 0.0028782119043171406, + 0.003209550864994526, + 0.007402829825878143, + 0.0034517922904342413, + 0.002675058087334037, + 0.0045254104770720005, + 0.0033951359800994396, + 0.004541145171970129, + 0.009739243425428867, + 0.005828084424138069, + 0.010970224626362324, + 0.018667468801140785, + 0.009442704729735851, + 0.006749155465513468, + 0.009429407306015491, + 0.004282883368432522, + 0.004973758012056351, + 0.009528480470180511, + 0.008379754610359669, + 0.003676649881526828, + 0.0028639400843530893, + 0.004702108446508646, + 0.0059735216200351715, + 0.005040836986154318, + 0.07236262410879135, + 0.004781300202012062, + 0.005787939298897982, + 0.07418587058782578, + 0.008496569469571114, + 0.005311641842126846, + 0.003078148001804948, + 0.006389813032001257, + 0.023988449946045876, + 0.01533804927021265, + 0.01061128918081522, + 0.020297614857554436, + 0.011360213160514832, + 0.009728281758725643, + 0.008535753935575485, + 0.0006735724164173007, + 0.0010384550550952554, + 0.0017007917631417513, + 0.0022021005861461163, + 0.0014286339282989502, + 0.0011330002453178167, + 0.0017825892427936196, + 0.0014483770355582237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.014199413359165192 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a wide-angle shot of a vast field of tall cornstalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dust storm depicted as a billowing wall of thick ochre dust engulfing the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are long, stark shadows cast by a soft directional light creating god rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cornstalks rendered in a hyper-realistic Unreal Engine style with pale yellow, tattered leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31349220871925354 + ], + "opposite_logprob_tokens": [ + -1.3134921789169312 + ], + "logprob_reward": -0.31349220871925354, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31349220871925354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.144450843334198, + 0.08000382781028748, + 0.11546482890844345, + 0.11015579849481583, + 0.04363131523132324, + 0.1371231973171234, + 0.12090539932250977, + 0.16905216872692108, + 0.03065398521721363, + 0.04855860024690628 + ], + "credit_map": [ + 0.013710379600524902, + 0.022795535624027252, + 0.016686342656612396, + 0.015815041959285736, + 0.020475083962082863, + 0.022256294265389442, + 0.020463967695832253, + 0.012248185463249683, + 0.00583504606038332, + 0.007864370942115784, + 0.011707672849297523, + 0.014685910195112228, + 0.007738928310573101, + 0.01234246976673603, + 0.012395636178553104, + 0.007433794904500246, + 0.010036149993538857, + 0.012153196148574352, + 0.015669193118810654, + 0.016382260248064995, + 0.01187326293438673, + 0.018328094854950905, + 0.01790502853691578, + 0.013117645867168903, + 0.005127317272126675, + 0.01162379328161478, + 0.009217149578034878, + 0.008743573911488056, + 0.02353818342089653, + 0.01784699223935604, + 0.020388567820191383, + 0.013670212589204311, + 0.0035247423220425844, + 0.00576142268255353, + 0.004972822032868862, + 0.005848153028637171, + 0.008432679809629917, + 0.004144595004618168, + 0.005765033420175314, + 0.005181864835321903, + 0.006152983754873276, + 0.020552633330225945, + 0.0138225257396698, + 0.021303026005625725, + 0.03077416680753231, + 0.015712061896920204, + 0.013377911411225796, + 0.015427893027663231, + 0.008251173421740532, + 0.010860010981559753, + 0.01752319000661373, + 0.03156726807355881, + 0.008632756769657135, + 0.011091606691479683, + 0.02205711044371128, + 0.010922287590801716, + 0.007507270202040672, + 0.02400653064250946, + 0.017994672060012817, + 0.03811013698577881, + 0.03716016560792923, + 0.014540174044668674, + 0.011968537233769894, + 0.017764687538146973, + 0.0021331575699150562, + 0.004600947257131338, + 0.003674605395644903, + 0.006488665007054806, + 0.003651021746918559, + 0.0021543961483985186, + 0.0028852655086666346, + 0.0050659277476370335, + 0.002439397620037198, + 0.010440781712532043, + 0.004233075305819511, + 0.0051123639568686485, + 0.014414140954613686, + 0.003079991554841399, + 0.002852632664144039, + 0.005986217875033617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.31349220871925354 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a wide-angle shot of a vast field of tall cornstalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dust storm depicted as a billowing wall of thick ochre dust engulfing the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are long, stark shadows cast by a soft directional light creating god rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cornstalks rendered in a hyper-realistic Unreal Engine style with pale yellow, tattered leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22560937702655792 + ], + "opposite_logprob_tokens": [ + -1.6006094217300415 + ], + "logprob_reward": -0.22560937702655792, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22560937702655792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17583070695400238, + 0.09611427038908005, + 0.08893575519323349, + 0.1514875590801239, + 0.02738535962998867, + 0.11513622105121613, + 0.09560813009738922, + 0.16090430319309235, + 0.03168826922774315, + 0.05690939351916313 + ], + "credit_map": [ + 0.012601526454091072, + 0.028301915153861046, + 0.028305090963840485, + 0.023584147915244102, + 0.02013162150979042, + 0.01651168242096901, + 0.02626596949994564, + 0.020128771662712097, + 0.005638871807605028, + 0.00832243263721466, + 0.012444091029465199, + 0.017582055181264877, + 0.0099201500415802, + 0.014988063834607601, + 0.012661474756896496, + 0.014557141810655594, + 0.00506998086348176, + 0.006407370790839195, + 0.012205841019749641, + 0.01688886620104313, + 0.009565197862684727, + 0.012181860394775867, + 0.01482664979994297, + 0.011789996176958084, + 0.006521675270050764, + 0.007636508904397488, + 0.0106937475502491, + 0.011012384667992592, + 0.037157606333494186, + 0.03912471607327461, + 0.022264903411269188, + 0.017076028510928154, + 0.0014580084243789315, + 0.0024905765894800425, + 0.0031615246552973986, + 0.005819970276206732, + 0.0032512866891920567, + 0.003397484542801976, + 0.00403023324906826, + 0.0037762788124382496, + 0.0059789028018713, + 0.014703291468322277, + 0.007595112547278404, + 0.021141204982995987, + 0.026266803964972496, + 0.012204928323626518, + 0.011403163895010948, + 0.015842830762267113, + 0.006929157767444849, + 0.011164908297359943, + 0.019095780327916145, + 0.024232493713498116, + 0.007491193246096373, + 0.006013886071741581, + 0.008408128283917904, + 0.012272585183382034, + 0.004376232158392668, + 0.05164628103375435, + 0.006392702925950289, + 0.01177906896919012, + 0.0567077100276947, + 0.008969582617282867, + 0.009124292060732841, + 0.011908444575965405, + 0.0014995761448517442, + 0.00793871097266674, + 0.0036352677270770073, + 0.005493419244885445, + 0.005648604594171047, + 0.0020834747701883316, + 0.0018058124696835876, + 0.0035834063310176134, + 0.0027302915696054697, + 0.013776063919067383, + 0.005650268867611885, + 0.0038081977982074022, + 0.020088816061615944, + 0.0045668939128518105, + 0.003332334104925394, + 0.0029565333388745785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.22560937702655792 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot captures a vast field of tall, sun-bleached cornstalks, flapping violently in a chaotic frenzy. A massive dust storm, a billowing wall of thick ochre dust, relentlessly engulfs the landscape, swallowing the horizon and creating an oppressive, claustrophobic atmosphere. A soft, directional light, like that from a dying sun, penetrates the dense cloud of particles, creating dramatic god rays that illuminate millions of individually rendered dust and sand particles dancing in the air. This intense light casts long, stark shadows that stretch and writhe across the ground, highlighting the corn's violent movement. Each stalk is rendered in a hyper-realistic Unreal Engine style, displaying realistic, high-fidelity textures; their dry, brittle leaves are pale yellow and tattered, whipping back and forth with such force that they shred and tear. The entire scene is a symphony of controlled chaos, with cinematic lighting and physically-based rendering emphasizing the oppressive power of nature and the fragility of the harvest.", + "seed": 24566, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_5/index_2550_seed_24566_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.167684331163764, + "vlm_logprob": -0.167684331163764, + "vlm_yes_no_margin": 2.075, + "vlm_token_credit": -0.167684331163764, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a wide-angle shot of a vast field of tall cornstalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dust storm depicted as a billowing wall of thick ochre dust engulfing the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are long, stark shadows cast by a soft directional light creating god rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cornstalks rendered in a hyper-realistic Unreal Engine style with pale yellow, tattered leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.19442704319953918 + ], + "opposite_logprob_tokens": [ + -2.069427013397217 + ], + "logprob_reward": -0.19442704319953918, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.19442704319953918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.618159294128418, + 0.05591927841305733, + 0.0808195173740387, + 0.042460180819034576, + 0.020096048712730408, + 0.015662111341953278, + 0.013629913330078125, + 0.011985545046627522, + 0.07212020456790924, + 0.06914790719747543 + ], + "credit_map": [ + 0.053556594997644424, + 0.08705560117959976, + 0.07608821243047714, + 0.1329873651266098, + 0.06845782697200775, + 0.05546826869249344, + 0.08215272426605225, + 0.062392693012952805, + 0.0047526308335363865, + 0.010281126014888287, + 0.00569870974868536, + 0.006404144689440727, + 0.008434890769422054, + 0.006917287595570087, + 0.008695696480572224, + 0.004734794143587351, + 0.007137986831367016, + 0.01702207326889038, + 0.007038943935185671, + 0.0064535001292824745, + 0.008970599621534348, + 0.014130576513707638, + 0.00893291737884283, + 0.0111329210922122, + 0.002927168505266309, + 0.007042675744742155, + 0.003470857860520482, + 0.005784019827842712, + 0.003954462707042694, + 0.006450952496379614, + 0.007709319703280926, + 0.0051207225769758224, + 0.001677896361798048, + 0.002522937720641494, + 0.0045461771078407764, + 0.0020108772441744804, + 0.002500802744179964, + 0.003049143124371767, + 0.0016885304357856512, + 0.0020996849052608013, + 0.0009735459461808205, + 0.0015512845711782575, + 0.002222670940682292, + 0.0020029025617986917, + 0.002098557772114873, + 0.0024834054056555033, + 0.0020328278187662363, + 0.0022969183046370745, + 0.0010653110221028328, + 0.0018255023751407862, + 0.0015522525645792484, + 0.0014755260199308395, + 0.0017916643992066383, + 0.001976106548681855, + 0.0014817704213783145, + 0.002461780561134219, + 0.0011582804145291448, + 0.0011941218981519341, + 0.0013002986088395119, + 0.0015163084026426077, + 0.0014487827429547906, + 0.0018475535325706005, + 0.00142256252001971, + 0.0020976366940885782, + 0.0048462022095918655, + 0.007687668316066265, + 0.006474845577031374, + 0.01106257550418377, + 0.01188125554472208, + 0.008341786451637745, + 0.010887346230447292, + 0.010938526131212711, + 0.0056721824221313, + 0.006284405943006277, + 0.008916007354855537, + 0.009785394184291363, + 0.010407461784780025, + 0.007895425893366337, + 0.009351683780550957, + 0.010835342109203339 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.19442704319953918 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a wide-angle shot of a vast field of tall cornstalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dust storm depicted as a billowing wall of thick ochre dust engulfing the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are long, stark shadows cast by a soft directional light creating god rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cornstalks rendered in a hyper-realistic Unreal Engine style with pale yellow, tattered leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17989353835582733 + ], + "opposite_logprob_tokens": [ + -1.8048934936523438 + ], + "logprob_reward": -0.17989353835582733, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17989353835582733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22924542427062988, + 0.1091013103723526, + 0.16270150244235992, + 0.07365871965885162, + 0.011966357938945293, + 0.042706601321697235, + 0.07628197968006134, + 0.04019712656736374, + 0.19338855147361755, + 0.06075245514512062 + ], + "credit_map": [ + 0.01629938930273056, + 0.027100684121251106, + 0.024273790419101715, + 0.06194048747420311, + 0.028001826256513596, + 0.021584656089544296, + 0.02874005027115345, + 0.021304523572325706, + 0.009714370593428612, + 0.013249089941382408, + 0.009621464647352695, + 0.012383501045405865, + 0.017428532242774963, + 0.016593631356954575, + 0.01883849874138832, + 0.011272220872342587, + 0.014613807201385498, + 0.03453291952610016, + 0.014869668520987034, + 0.01176745817065239, + 0.019143030047416687, + 0.027776101604104042, + 0.01891997456550598, + 0.021078545600175858, + 0.005489025730639696, + 0.013323807157576084, + 0.005319975782185793, + 0.0070237889885902405, + 0.005869051441550255, + 0.01252977829426527, + 0.016172267496585846, + 0.007931024767458439, + 0.0010396711295470595, + 0.0012598743196576834, + 0.0019614235498011112, + 0.0015048041241243482, + 0.0014533897629007697, + 0.001984360860660672, + 0.001120254979468882, + 0.0016425787471234798, + 0.004789249040186405, + 0.010903375223279, + 0.003997043240815401, + 0.005205220077186823, + 0.00570390559732914, + 0.004700548946857452, + 0.003800218226388097, + 0.0036070425994694233, + 0.005166435614228249, + 0.008786496706306934, + 0.009829310700297356, + 0.011774403974413872, + 0.012022437527775764, + 0.011936680413782597, + 0.007545622531324625, + 0.00922059640288353, + 0.0022995900362730026, + 0.0032809260301291943, + 0.0039278543554246426, + 0.004205125849694014, + 0.003358003916218877, + 0.005137245636433363, + 0.003652045736089349, + 0.014336338266730309, + 0.013339675962924957, + 0.018192525953054428, + 0.02079210802912712, + 0.023422444239258766, + 0.027644870802760124, + 0.029667144641280174, + 0.028649721294641495, + 0.03168005496263504, + 0.0044062878005206585, + 0.005788502749055624, + 0.007905937731266022, + 0.0063805896788835526, + 0.011335334740579128, + 0.008452428504824638, + 0.008124474436044693, + 0.00835889670997858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.17989353835582733 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a wide-angle shot of a vast field of tall cornstalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dust storm depicted as a billowing wall of thick ochre dust engulfing the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are long, stark shadows cast by a soft directional light creating god rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cornstalks rendered in a hyper-realistic Unreal Engine style with pale yellow, tattered leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635142020881176 + ], + "opposite_logprob_tokens": [ + -3.6513514518737793 + ], + "logprob_reward": -0.02635142020881176, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635142020881176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13701976835727692, + 0.3552103638648987, + 0.06003536656498909, + 0.03380342945456505, + 0.025905927643179893, + 0.12178680300712585, + 0.05528925731778145, + 0.04347512871026993, + 0.11262688040733337, + 0.05484698340296745 + ], + "credit_map": [ + 0.011342798359692097, + 0.014913087710738182, + 0.018710166215896606, + 0.033194322139024734, + 0.012068920768797398, + 0.013122672215104103, + 0.01888568140566349, + 0.014782154001295567, + 0.02818707749247551, + 0.0639512836933136, + 0.0893491804599762, + 0.06470116227865219, + 0.030540866777300835, + 0.021470854058861732, + 0.03582305088639259, + 0.0211869515478611, + 0.003995833452790976, + 0.00878323707729578, + 0.006158006377518177, + 0.004962226841598749, + 0.007635475602000952, + 0.011256910860538483, + 0.009878108277916908, + 0.0073655820451676846, + 0.002973480150103569, + 0.00353250396437943, + 0.0037095113657414913, + 0.004358264151960611, + 0.002785331103950739, + 0.007692764513194561, + 0.004204249009490013, + 0.004547329619526863, + 0.0024995163548737764, + 0.0034564374946057796, + 0.0043016900308430195, + 0.0034339004196226597, + 0.0025398891884833574, + 0.0032935435883700848, + 0.0031710180919617414, + 0.0032099357340484858, + 0.009610984474420547, + 0.01940177194774151, + 0.020999353379011154, + 0.01356562227010727, + 0.01411010604351759, + 0.01709873601794243, + 0.011856828816235065, + 0.015143418684601784, + 0.00434001674875617, + 0.007880247198045254, + 0.006411082576960325, + 0.006284734699875116, + 0.011562391184270382, + 0.0055036386474967, + 0.005088185891509056, + 0.008218969218432903, + 0.003968451172113419, + 0.004635998513549566, + 0.00417802669107914, + 0.006511495914310217, + 0.0036798231303691864, + 0.005029485560953617, + 0.004720588214695454, + 0.010751266963779926, + 0.007928879000246525, + 0.01177690178155899, + 0.013024340383708477, + 0.014096497558057308, + 0.013765815645456314, + 0.01795097254216671, + 0.018561668694019318, + 0.015521829016506672, + 0.0056181540712714195, + 0.007829323410987854, + 0.00810849480330944, + 0.007765836548060179, + 0.008108275942504406, + 0.00543171726167202, + 0.005086609628051519, + 0.00689858291298151 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.02635142020881176 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a wide-angle shot of a vast field of tall cornstalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dust storm depicted as a billowing wall of thick ochre dust engulfing the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are long, stark shadows cast by a soft directional light creating god rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cornstalks rendered in a hyper-realistic Unreal Engine style with pale yellow, tattered leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34867289662361145 + ], + "opposite_logprob_tokens": [ + -1.223672866821289 + ], + "logprob_reward": -0.34867289662361145, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34867289662361145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32879653573036194, + 0.2213059812784195, + 0.053166553378105164, + 0.05766909196972847, + 0.019369812682271004, + 0.08047211915254593, + 0.023206310346722603, + 0.03908120468258858, + 0.028507571667432785, + 0.14842481911182404 + ], + "credit_map": [ + 0.025708835572004318, + 0.03917988762259483, + 0.04444278031587601, + 0.08134133368730545, + 0.04366447776556015, + 0.027493052184581757, + 0.03776707500219345, + 0.02919909730553627, + 0.016563795506954193, + 0.04062178358435631, + 0.050212085247039795, + 0.03686270862817764, + 0.020905205979943275, + 0.016239896416664124, + 0.02616928145289421, + 0.013731240294873714, + 0.003430865006521344, + 0.006620946805924177, + 0.007632941473275423, + 0.004884440917521715, + 0.008137338794767857, + 0.009516806341707706, + 0.006778279785066843, + 0.006164934486150742, + 0.003965951967984438, + 0.011554128490388393, + 0.004525075666606426, + 0.005592600908130407, + 0.005567735992372036, + 0.008737842552363873, + 0.011187379248440266, + 0.006538372952491045, + 0.001562668476253748, + 0.0032059368677437305, + 0.00409972108900547, + 0.0029962975531816483, + 0.0020947528537362814, + 0.0013911506393924356, + 0.0022220814134925604, + 0.0017972036730498075, + 0.0063753509894013405, + 0.011508052237331867, + 0.01058795116841793, + 0.0072555504739284515, + 0.009470321238040924, + 0.013488613069057465, + 0.00913363415747881, + 0.01265264768153429, + 0.0015244745882228017, + 0.0035821106284856796, + 0.002425904618576169, + 0.002754726679995656, + 0.005980427842587233, + 0.0024514745455235243, + 0.002351291012018919, + 0.002135901479050517, + 0.002747387858107686, + 0.004242877010256052, + 0.004088373389095068, + 0.0041268630884587765, + 0.003970947116613388, + 0.006562739610671997, + 0.004856751766055822, + 0.008485264144837856, + 0.0021527945064008236, + 0.0027441787533462048, + 0.004622871056199074, + 0.0034467761870473623, + 0.004802213981747627, + 0.002969368826597929, + 0.0027689493726938963, + 0.0050004213117063046, + 0.014876734465360641, + 0.016418324783444405, + 0.02081364579498768, + 0.02184135466814041, + 0.02666810341179371, + 0.01139417290687561, + 0.016839316114783287, + 0.019573166966438293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.34867289662361145 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a wide-angle shot of a vast field of tall cornstalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dust storm depicted as a billowing wall of thick ochre dust engulfing the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are long, stark shadows cast by a soft directional light creating god rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cornstalks rendered in a hyper-realistic Unreal Engine style with pale yellow, tattered leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08907675743103027 + ], + "opposite_logprob_tokens": [ + -2.4640767574310303 + ], + "logprob_reward": -0.08907675743103027, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08907675743103027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26800450682640076, + 0.10378329455852509, + 0.12582075595855713, + 0.0631980150938034, + 0.0805753692984581, + 0.046576790511608124, + 0.05376933887600899, + 0.043241310864686966, + 0.15464359521865845, + 0.06038704141974449 + ], + "credit_map": [ + 0.025335567072033882, + 0.0362512432038784, + 0.031195761635899544, + 0.034745581448078156, + 0.0284363254904747, + 0.04111148789525032, + 0.0295955091714859, + 0.04133300110697746, + 0.010133231058716774, + 0.012781264260411263, + 0.019145485013723373, + 0.01605299673974514, + 0.013084754347801208, + 0.012699658051133156, + 0.011644936166703701, + 0.008240973576903343, + 0.010178394615650177, + 0.021675148978829384, + 0.011920183897018433, + 0.01056244969367981, + 0.012993797659873962, + 0.027856310829520226, + 0.010934270918369293, + 0.01970020681619644, + 0.005232096184045076, + 0.00632736599072814, + 0.011234639212489128, + 0.0076152365654706955, + 0.005059308838099241, + 0.012393255718052387, + 0.007470808923244476, + 0.0078653022646904, + 0.0053764586336910725, + 0.008907861076295376, + 0.020814768970012665, + 0.009484944865107536, + 0.010634437203407288, + 0.010465839877724648, + 0.00807237345725298, + 0.006818684283643961, + 0.005115031264722347, + 0.005738977342844009, + 0.0070338440127670765, + 0.006131726782768965, + 0.005019438918679953, + 0.005344398785382509, + 0.006100236903876066, + 0.006093135569244623, + 0.003976165316998959, + 0.005581878591328859, + 0.008274907246232033, + 0.006323019973933697, + 0.005573520436882973, + 0.010867731645703316, + 0.004757992457598448, + 0.008414125069975853, + 0.00377278751693666, + 0.00450319005176425, + 0.003938760608434677, + 0.005163864232599735, + 0.004566535819321871, + 0.006392660550773144, + 0.005533710587769747, + 0.009369799867272377, + 0.010634087026119232, + 0.015157879330217838, + 0.017781365662813187, + 0.020073702558875084, + 0.018109634518623352, + 0.02307252213358879, + 0.02522185444831848, + 0.024592548608779907, + 0.0077788312919437885, + 0.006278302054852247, + 0.009956412017345428, + 0.0055114515125751495, + 0.005397238302975893, + 0.00740363635122776, + 0.008606983348727226, + 0.009454185143113136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.08907675743103027 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot captures a vast field of tall, sun-bleached cornstalks, flapping violently in a chaotic frenzy. A massive dust storm, a billowing wall of thick ochre dust, relentlessly engulfs the landscape, swallowing the horizon and creating an oppressive, claustrophobic atmosphere. A soft, directional light, like that from a dying sun, penetrates the dense cloud of particles, creating dramatic god rays that illuminate millions of individually rendered dust and sand particles dancing in the air. This intense light casts long, stark shadows that stretch and writhe across the ground, highlighting the corn's violent movement. Each stalk is rendered in a hyper-realistic Unreal Engine style, displaying realistic, high-fidelity textures; their dry, brittle leaves are pale yellow and tattered, whipping back and forth with such force that they shred and tear. The entire scene is a symphony of controlled chaos, with cinematic lighting and physically-based rendering emphasizing the oppressive power of nature and the fragility of the harvest.", + "seed": 24567, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_5/index_2550_seed_24567_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10394996479153633, + "vlm_logprob": -0.10394996479153633, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.10394996479153633, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a wide-angle shot of a vast field of tall cornstalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dust storm depicted as a billowing wall of thick ochre dust engulfing the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are long, stark shadows cast by a soft directional light creating god rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cornstalks rendered in a hyper-realistic Unreal Engine style with pale yellow, tattered leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.21102957427501678 + ], + "opposite_logprob_tokens": [ + -1.8360295295715332 + ], + "logprob_reward": -0.21102957427501678, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.21102957427501678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1530986726284027, + 0.2836551368236542, + 0.018135175108909607, + 0.07830186933279037, + 0.019715894013643265, + 0.04191315174102783, + 0.06941014528274536, + 0.1222824901342392, + 0.060492876917123795, + 0.1529945731163025 + ], + "credit_map": [ + 0.01752626895904541, + 0.01447437983006239, + 0.026441244408488274, + 0.02423231303691864, + 0.012880486436188221, + 0.01975650154054165, + 0.020621689036488533, + 0.017165787518024445, + 0.011977901682257652, + 0.022207407280802727, + 0.02113289386034012, + 0.11889311671257019, + 0.012004788964986801, + 0.011663484387099743, + 0.020597906783223152, + 0.06517763435840607, + 0.0013432526029646397, + 0.003094632178544998, + 0.003585540922358632, + 0.0037217021454125643, + 0.0015601052436977625, + 0.0021344299893826246, + 0.0011900237295776606, + 0.0015054884133860469, + 0.008245508186519146, + 0.007673874031752348, + 0.010732768103480339, + 0.0074665741994977, + 0.008224988356232643, + 0.021565239876508713, + 0.006859177723526955, + 0.007533737923949957, + 0.0016425088979303837, + 0.002919234335422516, + 0.0017767937388271093, + 0.002260899869725108, + 0.0023795603774487972, + 0.00480515556409955, + 0.0018267417326569557, + 0.002104999264702201, + 0.0068895467557013035, + 0.003920038230717182, + 0.007012743502855301, + 0.006055453326553106, + 0.0035632734652608633, + 0.007233931217342615, + 0.0031453650444746017, + 0.004092798568308353, + 0.018692247569561005, + 0.005926256999373436, + 0.011998371221125126, + 0.007211041171103716, + 0.006200650706887245, + 0.004642421845346689, + 0.007905050180852413, + 0.006834112107753754, + 0.020040839910507202, + 0.011223974637687206, + 0.014815903268754482, + 0.01509969774633646, + 0.01538569200783968, + 0.014616306871175766, + 0.019883986562490463, + 0.011216087266802788, + 0.006437868345528841, + 0.006507265381515026, + 0.005083180032670498, + 0.010050703771412373, + 0.011280430480837822, + 0.009547006338834763, + 0.006422477308660746, + 0.005163948517292738, + 0.013008886948227882, + 0.013462220318615437, + 0.01435152068734169, + 0.019375013187527657, + 0.029298560693860054, + 0.024264607578516006, + 0.018275588750839233, + 0.020958170294761658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.21102957427501678 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a wide-angle shot of a vast field of tall cornstalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dust storm depicted as a billowing wall of thick ochre dust engulfing the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are long, stark shadows cast by a soft directional light creating god rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cornstalks rendered in a hyper-realistic Unreal Engine style with pale yellow, tattered leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033719539642333984 + ], + "opposite_logprob_tokens": [ + -3.408719539642334 + ], + "logprob_reward": -0.033719539642333984, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033719539642333984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11528880894184113, + 0.48107561469078064, + 0.022975226864218712, + 0.02275359071791172, + 0.04724440723657608, + 0.05994251370429993, + 0.04598503187298775, + 0.06396809220314026, + 0.0830562487244606, + 0.05771049112081528 + ], + "credit_map": [ + 0.007326753810048103, + 0.01778213307261467, + 0.015071187168359756, + 0.018863433972001076, + 0.010347248055040836, + 0.011959348805248737, + 0.017825229093432426, + 0.01611347496509552, + 0.019302168861031532, + 0.03789721056818962, + 0.03626783564686775, + 0.20245526731014252, + 0.019257087260484695, + 0.020151866599917412, + 0.036440808326005936, + 0.10930333286523819, + 0.0019832453690469265, + 0.0033721730578690767, + 0.004417486488819122, + 0.003512855852022767, + 0.002010451862588525, + 0.003559581469744444, + 0.001951668644323945, + 0.0021677652839571238, + 0.0020426514092832804, + 0.002730431267991662, + 0.003912447486072779, + 0.002307221991941333, + 0.0021804561838507652, + 0.004959625191986561, + 0.002122600097209215, + 0.0024981589522212744, + 0.0038531008176505566, + 0.00619353074580431, + 0.0034830619115382433, + 0.005683097988367081, + 0.008480928838253021, + 0.010520555078983307, + 0.004179937299340963, + 0.004850191995501518, + 0.008569450117647648, + 0.00628664530813694, + 0.007818966172635555, + 0.013354484923183918, + 0.005863497965037823, + 0.007180873304605484, + 0.005454738158732653, + 0.00541386054828763, + 0.010269952937960625, + 0.006807410623878241, + 0.00797003973275423, + 0.005175373516976833, + 0.004417724907398224, + 0.0028388972859829664, + 0.004434117116034031, + 0.004071514122188091, + 0.006381831597536802, + 0.009141347371041775, + 0.006026741117238998, + 0.006998199969530106, + 0.014609714969992638, + 0.007238965481519699, + 0.009323718957602978, + 0.004247569479048252, + 0.010663524270057678, + 0.007768082432448864, + 0.008587497286498547, + 0.008684992790222168, + 0.012228213250637054, + 0.015312018804252148, + 0.01066953968256712, + 0.009142380207777023, + 0.005104050040245056, + 0.004887609742581844, + 0.008322197012603283, + 0.008958042599260807, + 0.009824777022004128, + 0.005779845640063286, + 0.008870825171470642, + 0.005963147152215242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.033719539642333984 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a wide-angle shot of a vast field of tall cornstalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dust storm depicted as a billowing wall of thick ochre dust engulfing the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are long, stark shadows cast by a soft directional light creating god rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cornstalks rendered in a hyper-realistic Unreal Engine style with pale yellow, tattered leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016073528677225113 + ], + "opposite_logprob_tokens": [ + -4.141073703765869 + ], + "logprob_reward": -0.016073528677225113, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016073528677225113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08591094613075256, + 0.33045095205307007, + 0.020148981362581253, + 0.016421401873230934, + 0.19137166440486908, + 0.07631196081638336, + 0.04502291977405548, + 0.055764760822057724, + 0.12189523875713348, + 0.05670119449496269 + ], + "credit_map": [ + 0.0075521343387663364, + 0.0077460710890591145, + 0.010324115864932537, + 0.021426085382699966, + 0.007329818792641163, + 0.008546710945665836, + 0.01302003487944603, + 0.009965978562831879, + 0.01987556368112564, + 0.02590145170688629, + 0.040982525795698166, + 0.13113538920879364, + 0.018388662487268448, + 0.019870148971676826, + 0.030563611537218094, + 0.04373357445001602, + 0.0023421519435942173, + 0.0034115882590413094, + 0.004157244693487883, + 0.002292313612997532, + 0.0018800281686708331, + 0.0021844550501555204, + 0.002147601218894124, + 0.0017335987649857998, + 0.0011208131909370422, + 0.0023508137091994286, + 0.0023954883217811584, + 0.0014929494354873896, + 0.001760337152518332, + 0.004132451955229044, + 0.0014588134363293648, + 0.0017097365343943238, + 0.01607607677578926, + 0.02080785110592842, + 0.01548351626843214, + 0.023605521768331528, + 0.034362711012363434, + 0.042472753673791885, + 0.016874447464942932, + 0.0216887928545475, + 0.009100859053432941, + 0.009542170912027359, + 0.009043589234352112, + 0.01795867085456848, + 0.009770134463906288, + 0.007362390402704477, + 0.007204554509371519, + 0.0063295867294073105, + 0.008036252111196518, + 0.006470991764217615, + 0.0057643987238407135, + 0.0043342444114387035, + 0.006225347984582186, + 0.004405840300023556, + 0.005841575097292662, + 0.003944268450140953, + 0.007204726804047823, + 0.0060144029557704926, + 0.006932517979294062, + 0.007869395427405834, + 0.009080610238015652, + 0.005364078562706709, + 0.008431402035057545, + 0.0048676240257918835, + 0.01260551530867815, + 0.013917661271989346, + 0.014008847996592522, + 0.02818741649389267, + 0.010841243900358677, + 0.00977866817265749, + 0.01832607015967369, + 0.014229819178581238, + 0.005035263020545244, + 0.003988178912550211, + 0.006812695413827896, + 0.009038658812642097, + 0.00627519516274333, + 0.005796157754957676, + 0.011585009284317493, + 0.008170043118298054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 8 + ], + "selected_reward": -0.016073528677225113 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a wide-angle shot of a vast field of tall cornstalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dust storm depicted as a billowing wall of thick ochre dust engulfing the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are long, stark shadows cast by a soft directional light creating god rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cornstalks rendered in a hyper-realistic Unreal Engine style with pale yellow, tattered leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789995789527893 + ], + "opposite_logprob_tokens": [ + -2.5789995193481445 + ], + "logprob_reward": -0.0789995789527893, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789995789527893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060536667704582214, + 0.20780467987060547, + 0.016580546274781227, + 0.04407665133476257, + 0.04778173193335533, + 0.27405816316604614, + 0.03638841211795807, + 0.09149834513664246, + 0.1595907211303711, + 0.061684008687734604 + ], + "credit_map": [ + 0.004594873636960983, + 0.003728632815182209, + 0.008991800248622894, + 0.013199172914028168, + 0.0057886443100869656, + 0.003670681267976761, + 0.01405108068138361, + 0.006511789280921221, + 0.010269137099385262, + 0.01671796850860119, + 0.015617047436535358, + 0.08728394657373428, + 0.008665190078318119, + 0.010149293579161167, + 0.015030749142169952, + 0.04407137632369995, + 0.002049235859885812, + 0.002038307022303343, + 0.003805257612839341, + 0.001856646267697215, + 0.0011466341093182564, + 0.0017458449583500624, + 0.0019472575513646007, + 0.0019913651049137115, + 0.004202401731163263, + 0.004271408077329397, + 0.006718701682984829, + 0.004005621187388897, + 0.004710861947387457, + 0.011622938327491283, + 0.004128850996494293, + 0.004415872972458601, + 0.003861329285427928, + 0.004385205451399088, + 0.005756269674748182, + 0.008735042065382004, + 0.005619822070002556, + 0.007770298048853874, + 0.006029108539223671, + 0.005624661687761545, + 0.03211905062198639, + 0.05406457185745239, + 0.048891179263591766, + 0.03329591453075409, + 0.03400058299303055, + 0.024481171742081642, + 0.029099518433213234, + 0.01810619793832302, + 0.004645066801458597, + 0.003601928474381566, + 0.005816021468490362, + 0.006689152680337429, + 0.005057475995272398, + 0.002911534858867526, + 0.004457768052816391, + 0.0032094642519950867, + 0.012004584074020386, + 0.00936431996524334, + 0.009717116132378578, + 0.014478029683232307, + 0.01712208241224289, + 0.009904855862259865, + 0.011695082299411297, + 0.007212282158434391, + 0.021536413580179214, + 0.014311470091342926, + 0.019371118396520615, + 0.03337971866130829, + 0.01970895193517208, + 0.011581684462726116, + 0.01956530474126339, + 0.020136069506406784, + 0.006427090149372816, + 0.00608807522803545, + 0.007359686773270369, + 0.009344391524791718, + 0.0060192495584487915, + 0.008904955349862576, + 0.010157758370041847, + 0.007382803596556187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 8 + ], + "selected_reward": -0.0789995789527893 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a wide-angle shot of a vast field of tall cornstalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dust storm depicted as a billowing wall of thick ochre dust engulfing the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are long, stark shadows cast by a soft directional light creating god rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cornstalks rendered in a hyper-realistic Unreal Engine style with pale yellow, tattered leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17992760241031647 + ], + "opposite_logprob_tokens": [ + -1.8049275875091553 + ], + "logprob_reward": -0.17992760241031647, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17992760241031647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15705372393131256, + 0.31353452801704407, + 0.029897166416049004, + 0.04783722013235092, + 0.05332577973604202, + 0.19233499467372894, + 0.022009508684277534, + 0.07556422799825668, + 0.04216640070080757, + 0.0662764310836792 + ], + "credit_map": [ + 0.015886222943663597, + 0.018322432413697243, + 0.02233358658850193, + 0.032390136271715164, + 0.015347747132182121, + 0.010974623262882233, + 0.025210121646523476, + 0.01658884435892105, + 0.013291592709720135, + 0.02439914643764496, + 0.03436420485377312, + 0.13322629034519196, + 0.014947451651096344, + 0.014623926021158695, + 0.023298941552639008, + 0.05538301542401314, + 0.003083159914240241, + 0.0034369793720543385, + 0.007741850800812244, + 0.007577951531857252, + 0.0019424906931817532, + 0.002164238365367055, + 0.0021040975116193295, + 0.0018464010208845139, + 0.0035947714932262897, + 0.008674275130033493, + 0.006072451360523701, + 0.004220199305564165, + 0.0048316302709281445, + 0.012849709019064903, + 0.0036264273803681135, + 0.003967758733779192, + 0.004314745310693979, + 0.008292504586279392, + 0.007218466140329838, + 0.007505955174565315, + 0.007160133216530085, + 0.008672349154949188, + 0.005632326938211918, + 0.004529302474111319, + 0.025317823514342308, + 0.0360872782766819, + 0.031699180603027344, + 0.03246479108929634, + 0.02184411697089672, + 0.012828324921429157, + 0.018696140497922897, + 0.01339735183864832, + 0.004286499228328466, + 0.0024672886356711388, + 0.003261832520365715, + 0.002700281096622348, + 0.0029847074765712023, + 0.0016583058750256896, + 0.0022275454830378294, + 0.0024230489507317543, + 0.01197129301726818, + 0.007526110392063856, + 0.007728991564363241, + 0.01042191218584776, + 0.014965028502047062, + 0.007037903182208538, + 0.010752866975963116, + 0.005160131026059389, + 0.005364482756704092, + 0.003799991449341178, + 0.004847515840083361, + 0.006312099285423756, + 0.004758983384817839, + 0.005540317390114069, + 0.007605365011841059, + 0.003937646746635437, + 0.00564727745950222, + 0.004963689483702183, + 0.009073413908481598, + 0.008458556607365608, + 0.0071720038540661335, + 0.012069519609212875, + 0.009385853074491024, + 0.009506125934422016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.17992760241031647 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_6/index_tensor([2282, 2282])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_6/index_tensor([2282, 2282])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..63a42d722dc4128b4fb506614d3b5dd3fb972056 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_6/index_tensor([2282, 2282])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunningly close-up, static shot, captured from a level angle, focuses squarely on a single work of art, immersed in a soft, picture-book-like macaron-hued scene. A spherical bomb, composed of countless tightly packed, overlapping layers of pink, pale yellow, and mint green rose petals, hovers in the center of the frame. A warm, diffuse light gently streams down from above, giving the petals a velvety texture and casting almost invisible, soft shadows. Several emerald-green, flexible vines, like delicate threads, entwine and secure the bomb's structure. One reaches upward, spiraling into an elegant fuse, tipped with a small, pre-opening white bud. Fine cracks appear on the sphere's surface, and the outermost petals delicately curl and peel away with almost imperceptible grace, as if a miracle of life is slowly emerging from its cocoon. The background is treated as a dreamy scene composed of soft light spots and blurred blocks of color. The gradient of lavender and cream creates an extreme shallow depth of field effect, thus directing all visual focus to the moment of birth of this unique work of art, creating a dreamy atmosphere that is tranquil, wonderful and full of the beauty of the budding life.", + "seed": 5648, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_6/index_2282_seed_5648_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.027668017614632844, + "vlm_logprob": -0.027668017614632844, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.027668017614632844, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical bomb made of overlapping pink, pale yellow, and mint green rose petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are emerald-green flexible vines visible entwining the bomb and spiraling upward into a fuse with a white bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks appear on the sphere's surface while the outermost petals curl and peel away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dreamy scene with soft light spots, blurred color blocks, and a lavender-to-cream gradient?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027116285637021065 + ], + "opposite_logprob_tokens": [ + -4.152116298675537 + ], + "logprob_reward": -0.027116285637021065, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.027116285637021065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14738678932189941, + 0.1173379123210907, + 0.05403032898902893, + 0.10068944096565247, + 0.05095153674483299, + 0.19700777530670166, + 0.013883396983146667, + 0.02830638363957405, + 0.04176272079348564, + 0.24864371120929718 + ], + "credit_map": [ + 0.02226608246564865, + 0.012380407191812992, + 0.03417975455522537, + 0.020444152876734734, + 0.013084065169095993, + 0.018166789785027504, + 0.01144570205360651, + 0.015419834293425083, + 0.01634748838841915, + 0.009771620854735374, + 0.025282390415668488, + 0.013041201047599316, + 0.009608396328985691, + 0.015534364618360996, + 0.012899958528578281, + 0.014852496795356274, + 0.006267849821597338, + 0.005032129120081663, + 0.00864718109369278, + 0.007219390943646431, + 0.005767313297837973, + 0.006273728795349598, + 0.007785303983837366, + 0.007037431467324495, + 0.011644801124930382, + 0.012457105331122875, + 0.012450425885617733, + 0.013270474970340729, + 0.013751217164099216, + 0.016182437539100647, + 0.01046703103929758, + 0.010465945117175579, + 0.004902561195194721, + 0.005069982726126909, + 0.012393967248499393, + 0.006452995352447033, + 0.005464829504489899, + 0.007762469816952944, + 0.0043017202988266945, + 0.004603009205311537, + 0.017239365726709366, + 0.022925706580281258, + 0.024910079315304756, + 0.022390643134713173, + 0.024864966049790382, + 0.027037853375077248, + 0.0359300896525383, + 0.021709078922867775, + 0.002189185004681349, + 0.0020950559992343187, + 0.0013783115427941084, + 0.0018589921528473496, + 0.0017003521788865328, + 0.0017185858450829983, + 0.0014002665411680937, + 0.0015426467871293426, + 0.003144863061606884, + 0.0044841039925813675, + 0.003061247756704688, + 0.003961579874157906, + 0.004063834901899099, + 0.0039411671459674835, + 0.002614170778542757, + 0.003035417292267084, + 0.005286527331918478, + 0.006941388826817274, + 0.006676011253148317, + 0.007638162467628717, + 0.003137414576485753, + 0.00460810586810112, + 0.003878359915688634, + 0.003596749622374773, + 0.028553467243909836, + 0.06583081930875778, + 0.039036307483911514, + 0.021227367222309113, + 0.01595311611890793, + 0.05055833235383034, + 0.00918157771229744, + 0.018302718177437782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.027116285637021065 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical bomb made of overlapping pink, pale yellow, and mint green rose petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are emerald-green flexible vines visible entwining the bomb and spiraling upward into a fuse with a white bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks appear on the sphere's surface while the outermost petals curl and peel away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dreamy scene with soft light spots, blurred color blocks, and a lavender-to-cream gradient?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003200054634362459 + ], + "opposite_logprob_tokens": [ + -5.753200054168701 + ], + "logprob_reward": -0.003200054634362459, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003200054634362459, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12423151731491089, + 0.06146777793765068, + 0.10176557302474976, + 0.06803515553474426, + 0.03260212764143944, + 0.3073943257331848, + 0.06196261942386627, + 0.036475613713264465, + 0.08549748361110687, + 0.12056775391101837 + ], + "credit_map": [ + 0.017944524064660072, + 0.012330744415521622, + 0.021793361753225327, + 0.015328796580433846, + 0.010377643629908562, + 0.016207583248615265, + 0.012417756952345371, + 0.017831111326813698, + 0.006983479950577021, + 0.007543881889432669, + 0.009714160114526749, + 0.008323907852172852, + 0.008082905784249306, + 0.008522652089595795, + 0.006343227345496416, + 0.005953565239906311, + 0.010723087005317211, + 0.009063425473868847, + 0.012745370157063007, + 0.014487681910395622, + 0.011108392849564552, + 0.011755174957215786, + 0.015983665362000465, + 0.015898777171969414, + 0.008449962362647057, + 0.005485308822244406, + 0.008449179120361805, + 0.01006549783051014, + 0.007075316272675991, + 0.011333785951137543, + 0.007188261952251196, + 0.009987846948206425, + 0.0027316331397742033, + 0.0029922528192400932, + 0.004850267898291349, + 0.0055392226204276085, + 0.0035515923518687487, + 0.0049662161618471146, + 0.0039293281733989716, + 0.004041615873575211, + 0.027071183547377586, + 0.035693347454071045, + 0.040865082293748856, + 0.03729404881596565, + 0.03670313581824303, + 0.04084073007106781, + 0.05477011203765869, + 0.03415669873356819, + 0.007065006531774998, + 0.007427415810525417, + 0.0061774118803441525, + 0.007711527403444052, + 0.0075006671249866486, + 0.007153891958296299, + 0.011059262789785862, + 0.007867436856031418, + 0.004054239019751549, + 0.005279724486172199, + 0.004200816620141268, + 0.005468256771564484, + 0.005226869136095047, + 0.004978579003363848, + 0.0039309607818722725, + 0.003336166962981224, + 0.006456295028328896, + 0.011314768344163895, + 0.01045288611203432, + 0.016339195892214775, + 0.006939199287444353, + 0.009635169059038162, + 0.015096748247742653, + 0.009263209998607635, + 0.013492520898580551, + 0.03353813663125038, + 0.019229810684919357, + 0.010380628518760204, + 0.007903757505118847, + 0.02155807986855507, + 0.006479319650679827, + 0.007985497824847698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.003200054634362459 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical bomb made of overlapping pink, pale yellow, and mint green rose petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are emerald-green flexible vines visible entwining the bomb and spiraling upward into a fuse with a white bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks appear on the sphere's surface while the outermost petals curl and peel away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dreamy scene with soft light spots, blurred color blocks, and a lavender-to-cream gradient?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252490285784006 + ], + "opposite_logprob_tokens": [ + -5.255252361297607 + ], + "logprob_reward": -0.005252490285784006, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252490285784006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0759311094880104, + 0.08411002904176712, + 0.05972675979137421, + 0.08752471208572388, + 0.05872345715761185, + 0.31991493701934814, + 0.06847824156284332, + 0.035515155643224716, + 0.03965628892183304, + 0.1704193353652954 + ], + "credit_map": [ + 0.011506623588502407, + 0.006749482359737158, + 0.013031835667788982, + 0.008508501574397087, + 0.006758664734661579, + 0.010663115419447422, + 0.00783941987901926, + 0.010873465798795223, + 0.012383149936795235, + 0.006794087588787079, + 0.019256403669714928, + 0.007691673468798399, + 0.009263863787055016, + 0.012482781894505024, + 0.008853667415678501, + 0.007384397555142641, + 0.007487057242542505, + 0.006075794342905283, + 0.010364504531025887, + 0.006538124289363623, + 0.00634591793641448, + 0.00976928137242794, + 0.00675823213532567, + 0.006387847941368818, + 0.010136892087757587, + 0.00952435564249754, + 0.012154429219663143, + 0.008095047436654568, + 0.020002612844109535, + 0.009930658154189587, + 0.006677966099232435, + 0.011002748273313046, + 0.007285183295607567, + 0.006190501153469086, + 0.007369360886514187, + 0.006742935627698898, + 0.011910461820662022, + 0.006840383633971214, + 0.005870833061635494, + 0.006513797678053379, + 0.028188148513436317, + 0.03834373503923416, + 0.04233267530798912, + 0.0371578149497509, + 0.04126907140016556, + 0.041806962341070175, + 0.055845923721790314, + 0.03497057780623436, + 0.00986461155116558, + 0.008963257074356079, + 0.007033808156847954, + 0.008019618690013885, + 0.007151375059038401, + 0.008002012968063354, + 0.011598266661167145, + 0.007845287211239338, + 0.0036759451031684875, + 0.005802197847515345, + 0.0037173854652792215, + 0.004213833715766668, + 0.0053687854669988155, + 0.0059541440568864346, + 0.003015217836946249, + 0.003767642891034484, + 0.004596235696226358, + 0.005598174873739481, + 0.00554231321439147, + 0.006371728610247374, + 0.003929418046027422, + 0.00508923688903451, + 0.004123752471059561, + 0.004405423998832703, + 0.020496021956205368, + 0.04195374622941017, + 0.02629956416785717, + 0.015186922624707222, + 0.011336296796798706, + 0.03550334647297859, + 0.006562245078384876, + 0.013081174343824387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 3 + ], + "selected_reward": -0.005252490285784006 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical bomb made of overlapping pink, pale yellow, and mint green rose petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are emerald-green flexible vines visible entwining the bomb and spiraling upward into a fuse with a white bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks appear on the sphere's surface while the outermost petals curl and peel away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dreamy scene with soft light spots, blurred color blocks, and a lavender-to-cream gradient?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10028454661369324 + ], + "opposite_logprob_tokens": [ + -2.3502845764160156 + ], + "logprob_reward": -0.10028454661369324, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10028454661369324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11720999330282211, + 0.1455772966146469, + 0.06530290842056274, + 0.11389034241437912, + 0.03232979774475098, + 0.08396299183368683, + 0.05061257258057594, + 0.04050661250948906, + 0.15417879819869995, + 0.19642867147922516 + ], + "credit_map": [ + 0.019840417429804802, + 0.009250933304429054, + 0.025197576731443405, + 0.011102533899247646, + 0.0072762989439070225, + 0.019624704495072365, + 0.01007142849266529, + 0.01484611351042986, + 0.023786110803484917, + 0.01060490496456623, + 0.031930726021528244, + 0.01429112907499075, + 0.013117496855556965, + 0.02572278492152691, + 0.012505324557423592, + 0.01361881010234356, + 0.00841959286481142, + 0.004905284382402897, + 0.013830002397298813, + 0.007222082931548357, + 0.006503721699118614, + 0.012396099977195263, + 0.005216233432292938, + 0.006809895392507315, + 0.01439300924539566, + 0.013358226977288723, + 0.013213313184678555, + 0.01336375530809164, + 0.01654178276658058, + 0.01744728907942772, + 0.011926408857107162, + 0.013646557927131653, + 0.0025422596372663975, + 0.0025932893622666597, + 0.00813370756804943, + 0.0029985143337398767, + 0.002777647227048874, + 0.0065650236792862415, + 0.0026936917565762997, + 0.004025664646178484, + 0.009352290071547031, + 0.009025965817272663, + 0.009251190349459648, + 0.010188027285039425, + 0.012167018838226795, + 0.010961990803480148, + 0.014379044063389301, + 0.00863745715469122, + 0.008627653121948242, + 0.008817854337394238, + 0.004395091440528631, + 0.0054246410727500916, + 0.004812671337276697, + 0.011596132069826126, + 0.002919271122664213, + 0.00401925528421998, + 0.004456759430468082, + 0.004640149883925915, + 0.0044198292307555676, + 0.007343139033764601, + 0.007687875535339117, + 0.00455848965793848, + 0.0038959914818406105, + 0.003504377556964755, + 0.011362587101757526, + 0.019730063155293465, + 0.019185427576303482, + 0.026537466794252396, + 0.020304996520280838, + 0.017287543043494225, + 0.023112060502171516, + 0.016658658161759377, + 0.02337038144469261, + 0.05035893991589546, + 0.03084009885787964, + 0.01645461656153202, + 0.01273968443274498, + 0.04016784206032753, + 0.007474198471754789, + 0.015022925101220608 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.10028454661369324 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical bomb made of overlapping pink, pale yellow, and mint green rose petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are emerald-green flexible vines visible entwining the bomb and spiraling upward into a fuse with a white bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks appear on the sphere's surface while the outermost petals curl and peel away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dreamy scene with soft light spots, blurred color blocks, and a lavender-to-cream gradient?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024867109023034573 + ], + "opposite_logprob_tokens": [ + -6.002486705780029 + ], + "logprob_reward": -0.0024867109023034573, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024867109023034573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09138825535774231, + 0.03065720945596695, + 0.07471112161874771, + 0.01535970438271761, + 0.012800760567188263, + 0.08769918233156204, + 0.024801118299365044, + 0.005400576628744602, + 0.018719883635640144, + 0.6384621262550354 + ], + "credit_map": [ + 0.01229820679873228, + 0.007781362161040306, + 0.017911134287714958, + 0.015097232535481453, + 0.00660290801897645, + 0.012524396181106567, + 0.010382615961134434, + 0.008790411055088043, + 0.0035588503815233707, + 0.0025389688089489937, + 0.004894403740763664, + 0.003524050349369645, + 0.002920545404776931, + 0.0033455053344368935, + 0.004676945507526398, + 0.005197943188250065, + 0.009321190416812897, + 0.010489247739315033, + 0.011923175305128098, + 0.007107081823050976, + 0.009621869772672653, + 0.011780139058828354, + 0.007083308417350054, + 0.0073851183988153934, + 0.001874067122116685, + 0.0014419879298657179, + 0.0022690973710268736, + 0.0017461816314607859, + 0.0025715231895446777, + 0.0024655272718518972, + 0.0012563714990392327, + 0.001734950696118176, + 0.0010628991294652224, + 0.0012895175023004413, + 0.0019972799345850945, + 0.001979710767045617, + 0.0021594094578176737, + 0.0016768808709457517, + 0.001084920484572649, + 0.001550143351778388, + 0.007857943885028362, + 0.010003778152167797, + 0.011166688986122608, + 0.010237165726721287, + 0.010830298066139221, + 0.012102922424674034, + 0.0162864550948143, + 0.009213944897055626, + 0.005003877449780703, + 0.004352516029030085, + 0.0024073824752122164, + 0.0027466232422739267, + 0.0023250728845596313, + 0.00355980871245265, + 0.0025625391863286495, + 0.0018433022778481245, + 0.0006859906134195626, + 0.000669535540509969, + 0.0007889063563197851, + 0.0008524596923962235, + 0.000866088317707181, + 0.0005496407393366098, + 0.00047524095862172544, + 0.0005127150216139853, + 0.0019198836525902152, + 0.002833497477695346, + 0.0022551186848431826, + 0.0037822804879397154, + 0.0015726406127214432, + 0.002271316945552826, + 0.002273504389449954, + 0.0018116437131538987, + 0.07516784965991974, + 0.16052168607711792, + 0.10245416313409805, + 0.05398957431316376, + 0.04100077971816063, + 0.13359887897968292, + 0.022637277841567993, + 0.049092017114162445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.0024867109023034573 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunningly close-up, static shot, captured from a level angle, focuses squarely on a single work of art, immersed in a soft, picture-book-like macaron-hued scene. A spherical bomb, composed of countless tightly packed, overlapping layers of pink, pale yellow, and mint green rose petals, hovers in the center of the frame. A warm, diffuse light gently streams down from above, giving the petals a velvety texture and casting almost invisible, soft shadows. Several emerald-green, flexible vines, like delicate threads, entwine and secure the bomb's structure. One reaches upward, spiraling into an elegant fuse, tipped with a small, pre-opening white bud. Fine cracks appear on the sphere's surface, and the outermost petals delicately curl and peel away with almost imperceptible grace, as if a miracle of life is slowly emerging from its cocoon. The background is treated as a dreamy scene composed of soft light spots and blurred blocks of color. The gradient of lavender and cream creates an extreme shallow depth of field effect, thus directing all visual focus to the moment of birth of this unique work of art, creating a dreamy atmosphere that is tranquil, wonderful and full of the beauty of the budding life.", + "seed": 5649, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_6/index_2282_seed_5649_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0160491423914209, + "vlm_logprob": -0.0160491423914209, + "vlm_yes_no_margin": 5.15, + "vlm_token_credit": -0.0160491423914209, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical bomb made of overlapping pink, pale yellow, and mint green rose petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are emerald-green flexible vines visible entwining the bomb and spiraling upward into a fuse with a white bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks appear on the sphere's surface while the outermost petals curl and peel away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dreamy scene with soft light spots, blurred color blocks, and a lavender-to-cream gradient?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017407961189746857 + ], + "opposite_logprob_tokens": [ + -4.8924078941345215 + ], + "logprob_reward": -0.017407961189746857, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.017407961189746857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09173725545406342, + 0.03904847800731659, + 0.03517000004649162, + 0.04110565409064293, + 0.022690746933221817, + 0.012666909955441952, + 0.30836743116378784, + 0.3340924084186554, + 0.10361005365848541, + 0.011511074379086494 + ], + "credit_map": [ + 0.006925282068550587, + 0.009412921965122223, + 0.01587316021323204, + 0.01015675999224186, + 0.01175367459654808, + 0.006657849997282028, + 0.012347209267318249, + 0.018610401079058647, + 0.003754663746803999, + 0.004898001439869404, + 0.003967747557908297, + 0.006285219918936491, + 0.004777418915182352, + 0.003142809495329857, + 0.007017775438725948, + 0.005204837769269943, + 0.0036192547995597124, + 0.002739532385021448, + 0.003878877265378833, + 0.0052547212690114975, + 0.003754599019885063, + 0.00297540589235723, + 0.007045371923595667, + 0.005902238190174103, + 0.004161386284977198, + 0.0038474579341709614, + 0.004052407573908567, + 0.006539476104080677, + 0.005094787105917931, + 0.0035197408869862556, + 0.008298799395561218, + 0.005591596942394972, + 0.0024140928871929646, + 0.002495063468813896, + 0.0021175434812903404, + 0.0036802785471081734, + 0.003023612778633833, + 0.0029166587628424168, + 0.0031840975861996412, + 0.002859399188309908, + 0.0016802154714241624, + 0.001231788657605648, + 0.0019247184973210096, + 0.0018093885155394673, + 0.0011176824336871505, + 0.002060977276414633, + 0.001156982616521418, + 0.0016851574182510376, + 0.01872205175459385, + 0.007560521364212036, + 0.05445567145943642, + 0.008269176818430424, + 0.006545579992234707, + 0.13097640872001648, + 0.006992773152887821, + 0.07484523952007294, + 0.05808800831437111, + 0.01488080807030201, + 0.10171215236186981, + 0.011378016322851181, + 0.00777259049937129, + 0.09077398478984833, + 0.008966365829110146, + 0.04052048549056053, + 0.023886635899543762, + 0.013909132219851017, + 0.021294157952070236, + 0.009264853782951832, + 0.0076229386031627655, + 0.008692368865013123, + 0.011650580912828445, + 0.007289390079677105, + 0.001279001822695136, + 0.001738575752824545, + 0.001362221548333764, + 0.0012287310091778636, + 0.001189531059935689, + 0.001585703925229609, + 0.0016409901436418295, + 0.0014863192336633801 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.017407961189746857 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical bomb made of overlapping pink, pale yellow, and mint green rose petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are emerald-green flexible vines visible entwining the bomb and spiraling upward into a fuse with a white bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks appear on the sphere's surface while the outermost petals curl and peel away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dreamy scene with soft light spots, blurred color blocks, and a lavender-to-cream gradient?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031981533393263817 + ], + "opposite_logprob_tokens": [ + -5.753198146820068 + ], + "logprob_reward": -0.0031981533393263817, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031981533393263817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10551882535219193, + 0.05565238371491432, + 0.059341929852962494, + 0.09488736838102341, + 0.032136183232069016, + 0.0164541807025671, + 0.344160795211792, + 0.20947971940040588, + 0.06998681277036667, + 0.01238179486244917 + ], + "credit_map": [ + 0.008791404776275158, + 0.009751920588314533, + 0.014239290729165077, + 0.012329481542110443, + 0.012779521755874157, + 0.013389578089118004, + 0.018145250156521797, + 0.016092374920845032, + 0.004926484078168869, + 0.005873073358088732, + 0.008227787911891937, + 0.007137257605791092, + 0.006219274364411831, + 0.006204010918736458, + 0.0075986492447555065, + 0.009465848095715046, + 0.005639411509037018, + 0.006267638877034187, + 0.005446117836982012, + 0.01318993791937828, + 0.006249917205423117, + 0.004980963654816151, + 0.01003309153020382, + 0.007534845266491175, + 0.008915256708860397, + 0.010598551481962204, + 0.00847773626446724, + 0.014327132143080235, + 0.015096970833837986, + 0.006774544715881348, + 0.018111905083060265, + 0.012585273012518883, + 0.0035047775600105524, + 0.0033558455761522055, + 0.00353551865555346, + 0.005250935442745686, + 0.003603273769840598, + 0.0033199575264006853, + 0.004463620949536562, + 0.005102255381643772, + 0.0015148903476074338, + 0.001891659339889884, + 0.0021959650330245495, + 0.00251022563315928, + 0.0016310581704601645, + 0.001654021325521171, + 0.0016886757221072912, + 0.00336768408305943, + 0.02642037346959114, + 0.009810849092900753, + 0.05673632770776749, + 0.01021097507327795, + 0.008194728754460812, + 0.14427076280117035, + 0.010286382399499416, + 0.07823041081428528, + 0.0343109630048275, + 0.009231793694198132, + 0.0634632334113121, + 0.007181845605373383, + 0.004984254017472267, + 0.058988917618989944, + 0.006178592797368765, + 0.02514011412858963, + 0.016410445794463158, + 0.006157702300697565, + 0.01945328339934349, + 0.005022003315389156, + 0.0027832903433591127, + 0.009057704359292984, + 0.004710004199296236, + 0.006392382085323334, + 0.0014263017801567912, + 0.001980919623747468, + 0.0014249433297663927, + 0.0014293816639110446, + 0.0012394338846206665, + 0.001877480186522007, + 0.0018040937138721347, + 0.0011992409126833081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.0031981533393263817 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical bomb made of overlapping pink, pale yellow, and mint green rose petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are emerald-green flexible vines visible entwining the bomb and spiraling upward into a fuse with a white bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks appear on the sphere's surface while the outermost petals curl and peel away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dreamy scene with soft light spots, blurred color blocks, and a lavender-to-cream gradient?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031914988067001104 + ], + "opposite_logprob_tokens": [ + -5.7531914710998535 + ], + "logprob_reward": -0.0031914988067001104, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031914988067001104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13997718691825867, + 0.07454885542392731, + 0.06616981327533722, + 0.20212024450302124, + 0.054011497646570206, + 0.032920952886343, + 0.1316404789686203, + 0.07778195291757584, + 0.18910714983940125, + 0.03172187879681587 + ], + "credit_map": [ + 0.011459899134933949, + 0.013630256056785583, + 0.017937077209353447, + 0.014893711544573307, + 0.0221854280680418, + 0.01517814863473177, + 0.02427614852786064, + 0.0204165019094944, + 0.006548305042088032, + 0.011021187528967857, + 0.006901971530169249, + 0.009304031729698181, + 0.011834204196929932, + 0.010345634073019028, + 0.008194895461201668, + 0.01039862260222435, + 0.00600104033946991, + 0.00792471319437027, + 0.005694873631000519, + 0.012195365503430367, + 0.009195178747177124, + 0.005208036862313747, + 0.010447636246681213, + 0.009502962231636047, + 0.017835427075624466, + 0.020624671131372452, + 0.01705218106508255, + 0.03266805782914162, + 0.033046286553144455, + 0.01346550788730383, + 0.03886653482913971, + 0.028561577200889587, + 0.007780508138239384, + 0.006491061300039291, + 0.007189271505922079, + 0.007908355444669724, + 0.0062376148998737335, + 0.005506798624992371, + 0.005415207240730524, + 0.007482671644538641, + 0.0031024687923491, + 0.0030643842183053493, + 0.006148025393486023, + 0.0034151666332036257, + 0.0023397773038595915, + 0.004072597250342369, + 0.0035900985822081566, + 0.0071884342469275, + 0.013954712077975273, + 0.012696709483861923, + 0.0156633909791708, + 0.020203731954097748, + 0.008307707495987415, + 0.014205674640834332, + 0.017065316438674927, + 0.02954322285950184, + 0.014328338205814362, + 0.0038990636821836233, + 0.023503102362155914, + 0.004083672072738409, + 0.0024065272882580757, + 0.017371851950883865, + 0.005023197270929813, + 0.0071661886759102345, + 0.04414895176887512, + 0.019119752570986748, + 0.04909021407365799, + 0.015061364509165287, + 0.009811119176447392, + 0.021048638969659805, + 0.015097296796739101, + 0.01572980359196663, + 0.005635000299662352, + 0.00374703505076468, + 0.004270902834832668, + 0.002891742857173085, + 0.0026258020661771297, + 0.005924862809479237, + 0.003389391815289855, + 0.0032371405977755785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.0031914988067001104 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical bomb made of overlapping pink, pale yellow, and mint green rose petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are emerald-green flexible vines visible entwining the bomb and spiraling upward into a fuse with a white bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks appear on the sphere's surface while the outermost petals curl and peel away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dreamy scene with soft light spots, blurred color blocks, and a lavender-to-cream gradient?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493635684251785 + ], + "opposite_logprob_tokens": [ + -2.929936408996582 + ], + "logprob_reward": -0.05493635684251785, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493635684251785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12906238436698914, + 0.05149506777524948, + 0.09359630197286606, + 0.2285003662109375, + 0.04112200438976288, + 0.05821586772799492, + 0.11181367933750153, + 0.1098419725894928, + 0.13725149631500244, + 0.03910088911652565 + ], + "credit_map": [ + 0.01180224772542715, + 0.013777455314993858, + 0.015596347860991955, + 0.019670603796839714, + 0.019692273810505867, + 0.00907230470329523, + 0.019945092499256134, + 0.019506065174937248, + 0.005461442284286022, + 0.009485199116170406, + 0.0050005060620605946, + 0.007079227361828089, + 0.008619933389127254, + 0.004540922585874796, + 0.0060276188887655735, + 0.005280217621475458, + 0.008115222677588463, + 0.009670134633779526, + 0.008502773009240627, + 0.016610069200396538, + 0.011645634658634663, + 0.006129089277237654, + 0.018634358420968056, + 0.014289028942584991, + 0.01814059168100357, + 0.020589541643857956, + 0.020363837480545044, + 0.0358705036342144, + 0.024363184347748756, + 0.019617177546024323, + 0.04473152756690979, + 0.04482399299740791, + 0.00398773280903697, + 0.0036668372340500355, + 0.004462163429707289, + 0.0061095296405255795, + 0.004386811517179012, + 0.004385086707770824, + 0.005314277019351721, + 0.008809564635157585, + 0.0047917235642671585, + 0.0034450856037437916, + 0.009941458702087402, + 0.00499565526843071, + 0.0042060669511556625, + 0.007316488306969404, + 0.005131874233484268, + 0.01838751509785652, + 0.014425233006477356, + 0.008657821454107761, + 0.030092602595686913, + 0.00923425704240799, + 0.00739680603146553, + 0.02017243020236492, + 0.012316207401454449, + 0.009518320672214031, + 0.030535833910107613, + 0.00717938644811511, + 0.032852571457624435, + 0.006934916600584984, + 0.00406593456864357, + 0.012660290114581585, + 0.0062982672825455666, + 0.00931477826088667, + 0.028894703835248947, + 0.016588930040597916, + 0.02960335649549961, + 0.012878641486167908, + 0.009636026807129383, + 0.012197413481771946, + 0.014423738233745098, + 0.013028676621615887, + 0.008500142954289913, + 0.003659021109342575, + 0.006555045489221811, + 0.0036676679737865925, + 0.0022607375867664814, + 0.0069992924109101295, + 0.0037411856465041637, + 0.0037177966441959143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.05493635684251785 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical bomb made of overlapping pink, pale yellow, and mint green rose petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are emerald-green flexible vines visible entwining the bomb and spiraling upward into a fuse with a white bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks appear on the sphere's surface while the outermost petals curl and peel away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dreamy scene with soft light spots, blurred color blocks, and a lavender-to-cream gradient?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015117417788133025 + ], + "opposite_logprob_tokens": [ + -6.501511573791504 + ], + "logprob_reward": -0.0015117417788133025, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015117417788133025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05134308710694313, + 0.02543490193784237, + 0.012522049248218536, + 0.07054692506790161, + 0.010277695953845978, + 0.025802556425333023, + 0.05786938965320587, + 0.694919228553772, + 0.03853147104382515, + 0.012752752751111984 + ], + "credit_map": [ + 0.0041174329817295074, + 0.0035624890588223934, + 0.009752023033797741, + 0.005749966949224472, + 0.004583350848406553, + 0.006613715086132288, + 0.005101875402033329, + 0.011862228624522686, + 0.003077545203268528, + 0.002254737541079521, + 0.002980064833536744, + 0.002681048819795251, + 0.002699206117540598, + 0.0052810958586633205, + 0.0031701684929430485, + 0.003291034372523427, + 0.0012026710901409388, + 0.0016388337826356292, + 0.001383287482894957, + 0.0015727217542007565, + 0.001688854768872261, + 0.0014224735787138343, + 0.0019690615590661764, + 0.0016441454645246267, + 0.006829012185335159, + 0.006591834127902985, + 0.00698771420866251, + 0.010482562705874443, + 0.009751171804964542, + 0.006147589068859816, + 0.014237403869628906, + 0.009519631043076515, + 0.0011485372669994831, + 0.0013905641390010715, + 0.0013048561522737145, + 0.001377082895487547, + 0.0009605571394786239, + 0.001048850012011826, + 0.0012881638249382377, + 0.0017590845236554742, + 0.004144003614783287, + 0.001844076905399561, + 0.004643733147531748, + 0.001601274125277996, + 0.0015784522984176874, + 0.005024378653615713, + 0.0021387715823948383, + 0.004827864933758974, + 0.003299805335700512, + 0.0014881780371069908, + 0.011386588215827942, + 0.0017042006365954876, + 0.001301926444284618, + 0.024192435666918755, + 0.0016794218681752682, + 0.012816826812922955, + 0.11947551369667053, + 0.03040996938943863, + 0.21488068997859955, + 0.023732805624604225, + 0.015589214861392975, + 0.19136899709701538, + 0.017239123582839966, + 0.08222281187772751, + 0.006286955904215574, + 0.002603161847218871, + 0.011540554463863373, + 0.0021425613667815924, + 0.0014747364912182093, + 0.008583297953009605, + 0.0018991493852809072, + 0.004001051653176546, + 0.00208452669903636, + 0.0014784643426537514, + 0.002697764663025737, + 0.0010749957291409373, + 0.0010376553982496262, + 0.0018496267730370164, + 0.001211608643643558, + 0.0013181093381717801 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.0015117417788133025 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunningly close-up, static shot, captured from a level angle, focuses squarely on a single work of art, immersed in a soft, picture-book-like macaron-hued scene. A spherical bomb, composed of countless tightly packed, overlapping layers of pink, pale yellow, and mint green rose petals, hovers in the center of the frame. A warm, diffuse light gently streams down from above, giving the petals a velvety texture and casting almost invisible, soft shadows. Several emerald-green, flexible vines, like delicate threads, entwine and secure the bomb's structure. One reaches upward, spiraling into an elegant fuse, tipped with a small, pre-opening white bud. Fine cracks appear on the sphere's surface, and the outermost petals delicately curl and peel away with almost imperceptible grace, as if a miracle of life is slowly emerging from its cocoon. The background is treated as a dreamy scene composed of soft light spots and blurred blocks of color. The gradient of lavender and cream creates an extreme shallow depth of field effect, thus directing all visual focus to the moment of birth of this unique work of art, creating a dreamy atmosphere that is tranquil, wonderful and full of the beauty of the budding life.", + "seed": 5650, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_6/index_2282_seed_5650_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.018438932893332094, + "vlm_logprob": -0.018438932893332094, + "vlm_yes_no_margin": 5.5, + "vlm_token_credit": -0.018438932893332094, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical bomb made of overlapping pink, pale yellow, and mint green rose petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are emerald-green flexible vines visible entwining the bomb and spiraling upward into a fuse with a white bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks appear on the sphere's surface while the outermost petals curl and peel away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dreamy scene with soft light spots, blurred color blocks, and a lavender-to-cream gradient?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0184086412191391 + ], + "opposite_logprob_tokens": [ + -4.64340877532959 + ], + "logprob_reward": -0.0184086412191391, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0184086412191391, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14620104432106018, + 0.031036093831062317, + 0.06980933994054794, + 0.050605665892362595, + 0.41010889410972595, + 0.05424933508038521, + 0.04110509529709816, + 0.05769442021846771, + 0.11089708656072617, + 0.028293080627918243 + ], + "credit_map": [ + 0.010431424714624882, + 0.008730265311896801, + 0.05013956502079964, + 0.011221783235669136, + 0.008285894058644772, + 0.008083789609372616, + 0.010458813048899174, + 0.03884949907660484, + 0.003223894163966179, + 0.0028036339208483696, + 0.0041276090778410435, + 0.005198880564421415, + 0.0051046814769506454, + 0.002844276139512658, + 0.005041231866925955, + 0.0026918849907815456, + 0.005395733751356602, + 0.003682205220684409, + 0.023043138906359673, + 0.0028193872421979904, + 0.00533876521512866, + 0.007695771288126707, + 0.002372552640736103, + 0.019461791962385178, + 0.005049323663115501, + 0.004153238609433174, + 0.013057941570878029, + 0.003674031002447009, + 0.0059566497802734375, + 0.011269518174231052, + 0.003604237688705325, + 0.0038407226093113422, + 0.0319087877869606, + 0.01422129850834608, + 0.16366776823997498, + 0.015816563740372658, + 0.01647900976240635, + 0.1495133638381958, + 0.008813822641968727, + 0.00968828983604908, + 0.004442764446139336, + 0.006962591782212257, + 0.007469849660992622, + 0.00654680235311389, + 0.009174734354019165, + 0.006759846583008766, + 0.00864944327622652, + 0.0042433044873178005, + 0.004895164165645838, + 0.004236484412103891, + 0.004524005576968193, + 0.007054480724036694, + 0.003768470836803317, + 0.005413955543190241, + 0.007596577517688274, + 0.0036159558221697807, + 0.006674729287624359, + 0.008193787187337875, + 0.005278296768665314, + 0.010569592006504536, + 0.007233727257698774, + 0.006599567364901304, + 0.008888794109225273, + 0.0042559243738651276, + 0.012186464853584766, + 0.009293420240283012, + 0.009884252212941647, + 0.018954791128635406, + 0.011286511085927486, + 0.014765121974050999, + 0.02447471208870411, + 0.010051809251308441, + 0.004129429813474417, + 0.002255589235574007, + 0.002360112965106964, + 0.004442160949110985, + 0.0024896885734051466, + 0.0027642319910228252, + 0.0068147797137498856, + 0.003037089481949806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.0184086412191391 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical bomb made of overlapping pink, pale yellow, and mint green rose petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are emerald-green flexible vines visible entwining the bomb and spiraling upward into a fuse with a white bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks appear on the sphere's surface while the outermost petals curl and peel away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dreamy scene with soft light spots, blurred color blocks, and a lavender-to-cream gradient?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001944795367307961 + ], + "opposite_logprob_tokens": [ + -6.2519450187683105 + ], + "logprob_reward": -0.001944795367307961, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001944795367307961, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20379504561424255, + 0.028771422803401947, + 0.013865919783711433, + 0.015130124054849148, + 0.4056258499622345, + 0.022712010890245438, + 0.05292516201734543, + 0.07630615681409836, + 0.170432910323143, + 0.010435291565954685 + ], + "credit_map": [ + 0.014437650330364704, + 0.013279731385409832, + 0.08017099648714066, + 0.009947685524821281, + 0.011532437987625599, + 0.010188273154199123, + 0.006788711994886398, + 0.0574495866894722, + 0.00290377507917583, + 0.002057745587080717, + 0.006012993864715099, + 0.0043828110210597515, + 0.0029830799903720617, + 0.0032577982638031244, + 0.00391355762258172, + 0.003259665332734585, + 0.0012746279826387763, + 0.0007492869044654071, + 0.0043909745290875435, + 0.0007907418766990304, + 0.000998908537440002, + 0.0019573383033275604, + 0.0007567292777821422, + 0.002947313943877816, + 0.001732472563162446, + 0.001435976941138506, + 0.0018063305178657174, + 0.0017966043669730425, + 0.0021068903151899576, + 0.002034163800999522, + 0.001803743652999401, + 0.0024139436427503824, + 0.03163954243063927, + 0.014010703191161156, + 0.1621648669242859, + 0.01533830538392067, + 0.01639888621866703, + 0.148504838347435, + 0.008076966740190983, + 0.00949180219322443, + 0.0024694353342056274, + 0.003262390848249197, + 0.002971055917441845, + 0.0024183818604797125, + 0.003641266142949462, + 0.0026661825831979513, + 0.0032484179828315973, + 0.0020348837133497, + 0.006143580190837383, + 0.0050742910243570805, + 0.004840576555579901, + 0.009633849374949932, + 0.0049546146765351295, + 0.006948273163288832, + 0.011380777694284916, + 0.0039492095820605755, + 0.007459575310349464, + 0.009629587642848492, + 0.00656141759827733, + 0.015637002885341644, + 0.009817550890147686, + 0.008264961652457714, + 0.01343511138111353, + 0.005500960163772106, + 0.015271059237420559, + 0.012957037426531315, + 0.014853461645543575, + 0.030909936875104904, + 0.017121894285082817, + 0.021762382239103317, + 0.04226663336157799, + 0.015290518291294575, + 0.0019984578248113394, + 0.0009694354375824332, + 0.0011847614077851176, + 0.0011856600176542997, + 0.001018108450807631, + 0.0014482588740065694, + 0.0015148305101320148, + 0.0011157802073284984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.001944795367307961 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical bomb made of overlapping pink, pale yellow, and mint green rose petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are emerald-green flexible vines visible entwining the bomb and spiraling upward into a fuse with a white bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks appear on the sphere's surface while the outermost petals curl and peel away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dreamy scene with soft light spots, blurred color blocks, and a lavender-to-cream gradient?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000920706195756793 + ], + "opposite_logprob_tokens": [ + -7.00092077255249 + ], + "logprob_reward": -0.000920706195756793, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.000920706195756793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20802226662635803, + 0.03229917585849762, + 0.03418247774243355, + 0.01426511351019144, + 0.42873281240463257, + 0.016474757343530655, + 0.04324060678482056, + 0.06480799615383148, + 0.1479240506887436, + 0.010050687938928604 + ], + "credit_map": [ + 0.014788085594773293, + 0.013596014119684696, + 0.08569051325321198, + 0.008580917492508888, + 0.008699294179677963, + 0.008815357461571693, + 0.004602060653269291, + 0.06325004249811172, + 0.0021985925268381834, + 0.002271683420985937, + 0.008910143747925758, + 0.003607154358178377, + 0.004099720157682896, + 0.002457137918099761, + 0.0032363052014261484, + 0.005518444813787937, + 0.0026072643231600523, + 0.0025438256561756134, + 0.009726905263960361, + 0.001656333333812654, + 0.003780236467719078, + 0.004308137111365795, + 0.0016221289988607168, + 0.007937651127576828, + 0.002325985813513398, + 0.001512213610112667, + 0.0017587117617949843, + 0.0014272533589974046, + 0.0018831229535862803, + 0.002452578628435731, + 0.0011520396219566464, + 0.0017532084602862597, + 0.03378300368785858, + 0.014420905150473118, + 0.17156584560871124, + 0.016290688887238503, + 0.01741691119968891, + 0.1567649245262146, + 0.00860076304525137, + 0.009889806620776653, + 0.0018321634270250797, + 0.0015648030675947666, + 0.0018788257148116827, + 0.002310100244358182, + 0.002614453202113509, + 0.00168272340670228, + 0.0030178276356309652, + 0.0015738614602014422, + 0.005313247907906771, + 0.004092799033969641, + 0.00416279723867774, + 0.007537576835602522, + 0.003927827347069979, + 0.005914154928177595, + 0.00886443443596363, + 0.0034277713857591152, + 0.006809561047703028, + 0.008112712763249874, + 0.005725422874093056, + 0.01365450769662857, + 0.008099446073174477, + 0.007115751039236784, + 0.010691694915294647, + 0.004598907195031643, + 0.012406697496771812, + 0.0128012141212821, + 0.012811347842216492, + 0.02599642239511013, + 0.01708981581032276, + 0.01833476312458515, + 0.03553839772939682, + 0.012945408932864666, + 0.0021757585927844048, + 0.0008455979987047613, + 0.0009334533824585378, + 0.001218373072333634, + 0.0009310251334682107, + 0.0012903391616418958, + 0.0016726270550861955, + 0.0009835140081122518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.000920706195756793 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical bomb made of overlapping pink, pale yellow, and mint green rose petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are emerald-green flexible vines visible entwining the bomb and spiraling upward into a fuse with a white bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks appear on the sphere's surface while the outermost petals curl and peel away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dreamy scene with soft light spots, blurred color blocks, and a lavender-to-cream gradient?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000136375427246 + ], + "opposite_logprob_tokens": [ + -2.6950013637542725 + ], + "logprob_reward": -0.07000136375427246, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000136375427246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15188492834568024, + 0.015069854445755482, + 0.01326737180352211, + 0.017675302922725677, + 0.6801661849021912, + 0.016434289515018463, + 0.019654888659715652, + 0.016651922836899757, + 0.061515022069215775, + 0.007680273614823818 + ], + "credit_map": [ + 0.010522057302296162, + 0.009787468239665031, + 0.06169932708144188, + 0.007426377385854721, + 0.006253859959542751, + 0.0060574705712497234, + 0.004787561949342489, + 0.045350782573223114, + 0.00177925662137568, + 0.0015061238082125783, + 0.0015062512829899788, + 0.001632993808016181, + 0.002687644213438034, + 0.002493714913725853, + 0.0017269401578232646, + 0.0017369284760206938, + 0.0010904992232099175, + 0.0005666413344442844, + 0.003738554660230875, + 0.0009415557142347097, + 0.0006970237009227276, + 0.0013561140513047576, + 0.0007115587941370904, + 0.0041654231026768684, + 0.0016795856645330787, + 0.0015098345465958118, + 0.00418331241235137, + 0.0017028333386406302, + 0.0019176672212779522, + 0.0036476452369242907, + 0.0016609635204076767, + 0.0013734595850110054, + 0.05282490327954292, + 0.023460134863853455, + 0.27248644828796387, + 0.02575927972793579, + 0.027271220460534096, + 0.24878433346748352, + 0.01371887605637312, + 0.015860948711633682, + 0.002420274540781975, + 0.0014284038916230202, + 0.0016393682453781366, + 0.002795332809910178, + 0.001584726618602872, + 0.0019619003869593143, + 0.0029962067492306232, + 0.001608072780072689, + 0.0020718793384730816, + 0.002307809889316559, + 0.0015875811222940683, + 0.0032310551032423973, + 0.002871959237381816, + 0.002411517081782222, + 0.0037220518570393324, + 0.0014510342152789235, + 0.0019453526474535465, + 0.0016263241413980722, + 0.0016170172020792961, + 0.002459391253069043, + 0.0023020030930638313, + 0.002042290521785617, + 0.00326558668166399, + 0.0013939564814791083, + 0.005175268743187189, + 0.0041767386719584465, + 0.0052836318500339985, + 0.011619058437645435, + 0.005566911771893501, + 0.007494651712477207, + 0.016514793038368225, + 0.005683962255716324, + 0.001961965812370181, + 0.0006612055585719645, + 0.0008437174838036299, + 0.0007676216773688793, + 0.0008217861177399755, + 0.0010578078217804432, + 0.0008981234859675169, + 0.000668044900521636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.07000136375427246 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical bomb made of overlapping pink, pale yellow, and mint green rose petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are emerald-green flexible vines visible entwining the bomb and spiraling upward into a fuse with a white bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks appear on the sphere's surface while the outermost petals curl and peel away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dreamy scene with soft light spots, blurred color blocks, and a lavender-to-cream gradient?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009191579301841557 + ], + "opposite_logprob_tokens": [ + -7.000919342041016 + ], + "logprob_reward": -0.0009191579301841557, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009191579301841557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16841457784175873, + 0.10433056950569153, + 0.015158547088503838, + 0.04264148324728012, + 0.28407421708106995, + 0.024659790098667145, + 0.05071748048067093, + 0.07656856626272202, + 0.1915699988603592, + 0.041864778846502304 + ], + "credit_map": [ + 0.013282966800034046, + 0.011115288361907005, + 0.06119764968752861, + 0.010439899750053883, + 0.008792618289589882, + 0.009800213389098644, + 0.007990977726876736, + 0.04579496011137962, + 0.006515242625027895, + 0.008350647985935211, + 0.02658415026962757, + 0.008457688614726067, + 0.01658758334815502, + 0.011509610339999199, + 0.0076505993492901325, + 0.018675047904253006, + 0.001738969935104251, + 0.001501290942542255, + 0.001811063615605235, + 0.001792000257410109, + 0.0023574791848659515, + 0.0022531719878315926, + 0.0014054353814572096, + 0.0022991360165178776, + 0.004529192578047514, + 0.0037749512121081352, + 0.005741223227232695, + 0.0047681634314358234, + 0.005703863222151995, + 0.005762835033237934, + 0.005167869850993156, + 0.007193390280008316, + 0.02239229343831539, + 0.010031178593635559, + 0.11340126395225525, + 0.010580864734947681, + 0.011445372365415096, + 0.10406889766454697, + 0.005553745198994875, + 0.006600586697459221, + 0.002750921528786421, + 0.0028171215672045946, + 0.0039016157388687134, + 0.0023756662849336863, + 0.0036834385246038437, + 0.003631298430263996, + 0.0030007415916770697, + 0.0024989857338368893, + 0.005889581982046366, + 0.005207641050219536, + 0.005088374949991703, + 0.00835970975458622, + 0.005653727799654007, + 0.007107395213097334, + 0.00906157586723566, + 0.004349471069872379, + 0.010463448241353035, + 0.01005459763109684, + 0.007382640149444342, + 0.014818867668509483, + 0.00900449138134718, + 0.009116742759943008, + 0.010924674570560455, + 0.004803106654435396, + 0.01766328327357769, + 0.016313664615154266, + 0.01649930141866207, + 0.03343825787305832, + 0.02177485078573227, + 0.024833420291543007, + 0.04453866928815842, + 0.016508549451828003, + 0.005120666231960058, + 0.0036282215733081102, + 0.004073943942785263, + 0.0063261669129133224, + 0.003612012602388859, + 0.00453820638358593, + 0.009794741868972778, + 0.004770820494741201 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.0009191579301841557 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunningly close-up, static shot, captured from a level angle, focuses squarely on a single work of art, immersed in a soft, picture-book-like macaron-hued scene. A spherical bomb, composed of countless tightly packed, overlapping layers of pink, pale yellow, and mint green rose petals, hovers in the center of the frame. A warm, diffuse light gently streams down from above, giving the petals a velvety texture and casting almost invisible, soft shadows. Several emerald-green, flexible vines, like delicate threads, entwine and secure the bomb's structure. One reaches upward, spiraling into an elegant fuse, tipped with a small, pre-opening white bud. Fine cracks appear on the sphere's surface, and the outermost petals delicately curl and peel away with almost imperceptible grace, as if a miracle of life is slowly emerging from its cocoon. The background is treated as a dreamy scene composed of soft light spots and blurred blocks of color. The gradient of lavender and cream creates an extreme shallow depth of field effect, thus directing all visual focus to the moment of birth of this unique work of art, creating a dreamy atmosphere that is tranquil, wonderful and full of the beauty of the budding life.", + "seed": 5651, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_6/index_2282_seed_5651_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03434534352272749, + "vlm_logprob": -0.03434534352272749, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.03434534352272749, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical bomb made of overlapping pink, pale yellow, and mint green rose petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are emerald-green flexible vines visible entwining the bomb and spiraling upward into a fuse with a white bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks appear on the sphere's surface while the outermost petals curl and peel away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dreamy scene with soft light spots, blurred color blocks, and a lavender-to-cream gradient?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017028551548719406 + ], + "opposite_logprob_tokens": [ + -5.142028331756592 + ], + "logprob_reward": -0.017028551548719406, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.017028551548719406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1441575139760971, + 0.04019707441329956, + 0.049248822033405304, + 0.07566337287425995, + 0.2090120017528534, + 0.16395747661590576, + 0.14452804625034332, + 0.06885312497615814, + 0.05811154469847679, + 0.04627098888158798 + ], + "credit_map": [ + 0.01908269338309765, + 0.018960848450660706, + 0.018954560160636902, + 0.014811985194683075, + 0.02397064119577408, + 0.016606664285063744, + 0.012414360418915749, + 0.019355762749910355, + 0.00663750572130084, + 0.005870051681995392, + 0.003861295524984598, + 0.004838614258915186, + 0.004275746177881956, + 0.005045005585998297, + 0.004444179590791464, + 0.005224676802754402, + 0.005887530278414488, + 0.006249086931347847, + 0.006096990313380957, + 0.006419803947210312, + 0.004539999645203352, + 0.005592812784016132, + 0.0050164698623120785, + 0.009446130134165287, + 0.009341186843812466, + 0.008394020609557629, + 0.007100292015820742, + 0.011623038910329342, + 0.011230611242353916, + 0.007870021276175976, + 0.010582919232547283, + 0.009521282278001308, + 0.02746017649769783, + 0.024594658985733986, + 0.021374735981225967, + 0.029263844713568687, + 0.03055858425796032, + 0.021546103060245514, + 0.03151862323284149, + 0.022695275023579597, + 0.02859685570001602, + 0.024131493642926216, + 0.0164638664573431, + 0.01959242857992649, + 0.021654291078448296, + 0.017886916175484657, + 0.01947134919464588, + 0.016160279512405396, + 0.022184086963534355, + 0.018198251724243164, + 0.013289347290992737, + 0.016463326290249825, + 0.02594020590186119, + 0.013691694475710392, + 0.018838953226804733, + 0.015922177582979202, + 0.009896738454699516, + 0.008777369745075703, + 0.009830942377448082, + 0.008625432848930359, + 0.010092876851558685, + 0.007581316400319338, + 0.006880042143166065, + 0.007168404757976532, + 0.005386324133723974, + 0.007382987532764673, + 0.008285537362098694, + 0.007639931980520487, + 0.00587645685300231, + 0.007036304101347923, + 0.0077741872519254684, + 0.008729816414415836, + 0.0072056627832353115, + 0.006780178286135197, + 0.0055760773830115795, + 0.0036703431978821754, + 0.008337752893567085, + 0.007371885702013969, + 0.0030837077647447586, + 0.004245383199304342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.017028551548719406 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical bomb made of overlapping pink, pale yellow, and mint green rose petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are emerald-green flexible vines visible entwining the bomb and spiraling upward into a fuse with a white bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks appear on the sphere's surface while the outermost petals curl and peel away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dreamy scene with soft light spots, blurred color blocks, and a lavender-to-cream gradient?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046432530507445335 + ], + "opposite_logprob_tokens": [ + -5.379643440246582 + ], + "logprob_reward": -0.0046432530507445335, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046432530507445335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12582947313785553, + 0.03769863024353981, + 0.02821977809071541, + 0.09450170397758484, + 0.24915607273578644, + 0.2297162264585495, + 0.05231475085020065, + 0.04100760444998741, + 0.03579476475715637, + 0.10576105117797852 + ], + "credit_map": [ + 0.01682630367577076, + 0.017647655680775642, + 0.015239442698657513, + 0.01426903996616602, + 0.016816237941384315, + 0.01521079707890749, + 0.011753806844353676, + 0.01806619204580784, + 0.004075460601598024, + 0.005117474589496851, + 0.0036142501048743725, + 0.004790142644196749, + 0.0035608247853815556, + 0.004945986904203892, + 0.0049659935757517815, + 0.00662849610671401, + 0.0033962007146328688, + 0.004371175076812506, + 0.0031290207989513874, + 0.003611123189330101, + 0.004009254276752472, + 0.003438045037910342, + 0.0031879469752311707, + 0.0030770108569413424, + 0.01633215881884098, + 0.012979762628674507, + 0.008748191408813, + 0.012697205878794193, + 0.009133151732385159, + 0.011202397756278515, + 0.012189652770757675, + 0.011219191364943981, + 0.03696588799357414, + 0.03313791751861572, + 0.024655403569340706, + 0.032789044082164764, + 0.03543492406606674, + 0.026687875390052795, + 0.034829091280698776, + 0.02465592697262764, + 0.035526342689991, + 0.027422772720456123, + 0.020321696996688843, + 0.029043953865766525, + 0.04131374508142471, + 0.022808915004134178, + 0.03061794489622116, + 0.022660857066512108, + 0.007454004604369402, + 0.00941733829677105, + 0.005785925313830376, + 0.005664808209985495, + 0.008359194733202457, + 0.005740637890994549, + 0.0037995190359652042, + 0.006093320902436972, + 0.0049443840980529785, + 0.005062352865934372, + 0.005776199046522379, + 0.004355302546173334, + 0.009039647877216339, + 0.004405526909977198, + 0.003068559570237994, + 0.004355631303042173, + 0.004153353627771139, + 0.004950657952576876, + 0.003677905537188053, + 0.003307156264781952, + 0.007290486712008715, + 0.0034831175580620766, + 0.004664917476475239, + 0.004267173353582621, + 0.013793868944048882, + 0.014935515820980072, + 0.01749066263437271, + 0.012323448434472084, + 0.00922585092484951, + 0.01305802259594202, + 0.010864291340112686, + 0.014069389551877975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.0046432530507445335 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical bomb made of overlapping pink, pale yellow, and mint green rose petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are emerald-green flexible vines visible entwining the bomb and spiraling upward into a fuse with a white bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks appear on the sphere's surface while the outermost petals curl and peel away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dreamy scene with soft light spots, blurred color blocks, and a lavender-to-cream gradient?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00319126108661294 + ], + "opposite_logprob_tokens": [ + -5.7531914710998535 + ], + "logprob_reward": -0.00319126108661294, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00319126108661294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14299653470516205, + 0.03806490823626518, + 0.03268394246697426, + 0.05916018411517143, + 0.20123213529586792, + 0.15379640460014343, + 0.18543604016304016, + 0.053915612399578094, + 0.06300599873065948, + 0.0697082132101059 + ], + "credit_map": [ + 0.016732722520828247, + 0.020153097808361053, + 0.016862496733665466, + 0.015030289068818092, + 0.02551424503326416, + 0.016420159488916397, + 0.014761391095817089, + 0.017522137612104416, + 0.005629031918942928, + 0.005915426183491945, + 0.0047174193896353245, + 0.003770427545532584, + 0.0038164490833878517, + 0.005030019208788872, + 0.004661442246288061, + 0.0045246947556734085, + 0.005441985093057156, + 0.005868921987712383, + 0.0035037959460169077, + 0.003758413251489401, + 0.0033272586297243834, + 0.0038766353391110897, + 0.0030694978777319193, + 0.003837436670437455, + 0.006456852424889803, + 0.005668227095156908, + 0.005942102521657944, + 0.010029476135969162, + 0.006861270405352116, + 0.005628703162074089, + 0.009802171029150486, + 0.008771385997533798, + 0.029437517747282982, + 0.025832833722233772, + 0.02021758444607258, + 0.026339231058955193, + 0.028083307668566704, + 0.022284768521785736, + 0.02853243611752987, + 0.02050446905195713, + 0.023948032408952713, + 0.020579436793923378, + 0.013800575397908688, + 0.021014714613556862, + 0.021660275757312775, + 0.015978379175066948, + 0.02060054801404476, + 0.016214456409215927, + 0.02868027612566948, + 0.024687208235263824, + 0.016996590420603752, + 0.024799412116408348, + 0.026168815791606903, + 0.018366001546382904, + 0.02528495155274868, + 0.020452793687582016, + 0.00502526992931962, + 0.007600261364132166, + 0.0042439778335392475, + 0.0038552447222173214, + 0.01621902920305729, + 0.0068571013398468494, + 0.0053719282150268555, + 0.004742803983390331, + 0.005493755917996168, + 0.007039661053568125, + 0.007609065622091293, + 0.0064879474230110645, + 0.012077132239937782, + 0.006830926053225994, + 0.009430713020265102, + 0.008036807179450989, + 0.008658445440232754, + 0.009357541799545288, + 0.006847795099020004, + 0.006736537907272577, + 0.01387905515730381, + 0.007482362445443869, + 0.007677168119698763, + 0.009069306775927544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.00319126108661294 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical bomb made of overlapping pink, pale yellow, and mint green rose petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are emerald-green flexible vines visible entwining the bomb and spiraling upward into a fuse with a white bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks appear on the sphere's surface while the outermost petals curl and peel away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dreamy scene with soft light spots, blurred color blocks, and a lavender-to-cream gradient?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14277337491512299 + ], + "opposite_logprob_tokens": [ + -2.017773389816284 + ], + "logprob_reward": -0.14277337491512299, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14277337491512299, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17574074864387512, + 0.05506308749318123, + 0.05255768820643425, + 0.04879103600978851, + 0.13528680801391602, + 0.13351301848888397, + 0.18623486161231995, + 0.06053014099597931, + 0.07727602124214172, + 0.07500655949115753 + ], + "credit_map": [ + 0.020908014848828316, + 0.023978354409337044, + 0.016305575147271156, + 0.020438531413674355, + 0.04075276851654053, + 0.017264273017644882, + 0.019010860472917557, + 0.01708236336708069, + 0.0065179625526070595, + 0.007930280640721321, + 0.00473497761413455, + 0.006419240031391382, + 0.00649050110951066, + 0.007854192517697811, + 0.005769321229308844, + 0.009346610866487026, + 0.00845377054065466, + 0.008247477002441883, + 0.00596168776974082, + 0.006278855726122856, + 0.0049610198475420475, + 0.006007667630910873, + 0.0062063406221568584, + 0.0064408667385578156, + 0.005240024067461491, + 0.005754166282713413, + 0.004618988372385502, + 0.006382797844707966, + 0.009669704362750053, + 0.005264973267912865, + 0.006186673417687416, + 0.005673708859831095, + 0.020483890548348427, + 0.01669207587838173, + 0.013714957982301712, + 0.015968775376677513, + 0.020186426118016243, + 0.016000866889953613, + 0.018328074365854263, + 0.013911732472479343, + 0.021063150838017464, + 0.01851344294846058, + 0.011963706463575363, + 0.018636591732501984, + 0.017967060208320618, + 0.01406621839851141, + 0.01657150126993656, + 0.014731346629559994, + 0.029562238603830338, + 0.025350991636514664, + 0.016121022403240204, + 0.02651735208928585, + 0.02277461253106594, + 0.019420938566327095, + 0.025770388543605804, + 0.020717332139611244, + 0.005736853461712599, + 0.008764653466641903, + 0.004723798483610153, + 0.004246395081281662, + 0.01744786649942398, + 0.008871831931173801, + 0.0051049599424004555, + 0.005633778404444456, + 0.007363937329500914, + 0.008931350894272327, + 0.008790284395217896, + 0.007420264650136232, + 0.017479579895734787, + 0.008896789513528347, + 0.009418437257409096, + 0.008975381031632423, + 0.008330535143613815, + 0.00985690951347351, + 0.007584136910736561, + 0.006820305250585079, + 0.019162049517035484, + 0.007119297981262207, + 0.00830787979066372, + 0.007825448177754879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.14277337491512299 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical bomb made of overlapping pink, pale yellow, and mint green rose petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are emerald-green flexible vines visible entwining the bomb and spiraling upward into a fuse with a white bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks appear on the sphere's surface while the outermost petals curl and peel away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dreamy scene with soft light spots, blurred color blocks, and a lavender-to-cream gradient?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004090277012437582 + ], + "opposite_logprob_tokens": [ + -5.504090309143066 + ], + "logprob_reward": -0.004090277012437582, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004090277012437582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45439019799232483, + 0.14680959284305573, + 0.05723012611269951, + 0.05152552202343941, + 0.037952713668346405, + 0.06651850044727325, + 0.039452359080314636, + 0.056150294840335846, + 0.032792218029499054, + 0.05717845633625984 + ], + "credit_map": [ + 0.08320049941539764, + 0.07967241108417511, + 0.04221423715353012, + 0.05353005975484848, + 0.04510123282670975, + 0.05720105394721031, + 0.03588499128818512, + 0.057585712522268295, + 0.020116718485951424, + 0.0236256942152977, + 0.014983941800892353, + 0.01949235424399376, + 0.013615834526717663, + 0.020367294549942017, + 0.015055811032652855, + 0.019551949575543404, + 0.00640737172216177, + 0.00697951577603817, + 0.006540903355926275, + 0.0077321515418589115, + 0.009132849983870983, + 0.0064282710663974285, + 0.006984064821153879, + 0.007024992257356644, + 0.006917533464729786, + 0.007141896989196539, + 0.004743359517306089, + 0.005433640442788601, + 0.009478620253503323, + 0.006948776077479124, + 0.004190228879451752, + 0.006671464536339045, + 0.003953221719712019, + 0.005439846310764551, + 0.004219359718263149, + 0.004335411358624697, + 0.00704686064273119, + 0.004044234286993742, + 0.0043020392768085, + 0.004611739423125982, + 0.011339502409100533, + 0.0084608756005764, + 0.006546791642904282, + 0.007325856946408749, + 0.010547099635004997, + 0.006191017106175423, + 0.00812207069247961, + 0.007985283620655537, + 0.004582828376442194, + 0.006266802083700895, + 0.0035042527597397566, + 0.004670273512601852, + 0.005603705998510122, + 0.00663301395252347, + 0.0033480427227914333, + 0.004843441769480705, + 0.006731690838932991, + 0.008428824134171009, + 0.008738932199776173, + 0.008260301314294338, + 0.007953261956572533, + 0.0061493744142353535, + 0.004649351350963116, + 0.0052385590970516205, + 0.00402018241584301, + 0.004359323065727949, + 0.0031384655740112066, + 0.004592051263898611, + 0.0041580805554986, + 0.004093403462320566, + 0.004113369155675173, + 0.0043173423036932945, + 0.010690588504076004, + 0.008062784560024738, + 0.0077825384214520454, + 0.006422385573387146, + 0.0054739248007535934, + 0.006622329354286194, + 0.005027898587286472, + 0.007096003275364637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.004090277012437582 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_7/index_tensor([7599, 7599])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_7/index_tensor([7599, 7599])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..00ffbde4a916403ce0c0b891958e5a43a90773e4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_7/index_tensor([7599, 7599])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "focused close-up begins on the hands of an elderly woman, her skin delicately wrinkled with age, but her grip steady and cautious, her fingers wrapped around the sleek, matte black, ergonomic handlebars of a sleek, pearly white premium mobility scooter. Then, the camera slowly and gently tilts upward, gliding over her arm, draped in a soft, cream-colored cashmere cardigan. The movement continues, smoothly bringing her face into the beautifully composed portrait frame. Her wispy, silver-white hair, styled in a neatly combed bob, gently frames her face, her kind eyes framed by gentle crow's feet, and a contented, soft smile graces her lips. The background is a warm, out-of-focus blur of a sun-drenched room, the soft green shapes of indoor plants creating a delightful bokeh effect. As she gracefully settles into the vehicle's comfortable seat, the soft, diffuse morning light streaming from an unseen window illuminates her serene expression, capturing a quiet, optimistic moment of newfound independence and peaceful anticipation before she prepares for action.", + "seed": 8266, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_7/index_7599_seed_8266_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.011010555946268142, + "vlm_logprob": -0.011010555946268142, + "vlm_yes_no_margin": 5.0, + "vlm_token_credit": -0.011010555946268142, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled skin and silver-white hair styled in a bob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding the handlebars of a pearly white mobility scooter with a matte black ergonomic grip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere cardigan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a close-up on her hands and slowly tilt upward to reveal her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a warm, out-of-focus blur of a sun-drenched room with green indoor plants creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005320555530488491 + ], + "opposite_logprob_tokens": [ + -6.0053205490112305 + ], + "logprob_reward": -0.005320555530488491, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.005320555530488491, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08313249796628952, + 0.04570753127336502, + 0.03838609531521797, + 0.0849793553352356, + 0.04971382021903992, + 0.048097189515829086, + 0.10547763854265213, + 0.08273324370384216, + 0.04625008627772331, + 0.4155225157737732 + ], + "credit_map": [ + 0.017833538353443146, + 0.009854485280811787, + 0.015327127650380135, + 0.008870745077729225, + 0.005924465134739876, + 0.006712225265800953, + 0.010374323464930058, + 0.008235583081841469, + 0.005980547983199358, + 0.00539432093501091, + 0.00908222422003746, + 0.005277650896459818, + 0.003910701256245375, + 0.006616098340600729, + 0.00573680829256773, + 0.003709180746227503, + 0.0036665599327534437, + 0.007060336880385876, + 0.0032845118548721075, + 0.006760863121598959, + 0.0036996256094425917, + 0.005650429520756006, + 0.00577402813360095, + 0.002489738631993532, + 0.00841116439551115, + 0.010677496902644634, + 0.01261352002620697, + 0.006895693950355053, + 0.011901132762432098, + 0.01809725910425186, + 0.011980916373431683, + 0.0044021750800311565, + 0.0039538792334496975, + 0.007781325373798609, + 0.00382933858782053, + 0.004686151631176472, + 0.007863401435315609, + 0.009624273516237736, + 0.004507536068558693, + 0.007467913907021284, + 0.003413017140701413, + 0.01025321613997221, + 0.003476526355370879, + 0.008045610040426254, + 0.0055120475590229034, + 0.005664553958922625, + 0.006524840835481882, + 0.005207380745559931, + 0.011321685276925564, + 0.021187586709856987, + 0.01241445355117321, + 0.011681938543915749, + 0.012219690717756748, + 0.012557636015117168, + 0.014019438996911049, + 0.010075218975543976, + 0.009279216639697552, + 0.008289354853332043, + 0.019977889955043793, + 0.009721079841256142, + 0.009197835810482502, + 0.007782644126564264, + 0.009273589588701725, + 0.009211638942360878, + 0.0034743233118206263, + 0.005581073462963104, + 0.0034111267887055874, + 0.0042925067245960236, + 0.00729396753013134, + 0.005421129520982504, + 0.011323720216751099, + 0.005452238954603672, + 0.027099866420030594, + 0.011994118802249432, + 0.02855636551976204, + 0.02317361906170845, + 0.06597062200307846, + 0.03090294636785984, + 0.15359845757484436, + 0.07422653585672379 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 3 + ], + "selected_reward": -0.005320555530488491 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled skin and silver-white hair styled in a bob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding the handlebars of a pearly white mobility scooter with a matte black ergonomic grip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere cardigan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a close-up on her hands and slowly tilt upward to reveal her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a warm, out-of-focus blur of a sun-drenched room with green indoor plants creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016086315736174583 + ], + "opposite_logprob_tokens": [ + -4.141086101531982 + ], + "logprob_reward": -0.016086315736174583, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016086315736174583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26840895414352417, + 0.07631098479032516, + 0.1751008927822113, + 0.05478568375110626, + 0.0502479188144207, + 0.08276134729385376, + 0.10856500267982483, + 0.06204978749155998, + 0.04330246523022652, + 0.07846692949533463 + ], + "credit_map": [ + 0.03274470940232277, + 0.027340278029441833, + 0.03272449970245361, + 0.04514920711517334, + 0.022653920575976372, + 0.02606351487338543, + 0.036872975528240204, + 0.044859860092401505, + 0.011490656062960625, + 0.008559728972613811, + 0.014774040319025517, + 0.010147962719202042, + 0.006285711657255888, + 0.0075749182142317295, + 0.01178333256393671, + 0.00569463986903429, + 0.012555388733744621, + 0.043588947504758835, + 0.02208913303911686, + 0.02243771217763424, + 0.014923143200576305, + 0.0388464629650116, + 0.013591689057648182, + 0.00706841517239809, + 0.006053195800632238, + 0.008322292938828468, + 0.008252467028796673, + 0.004214408807456493, + 0.006937381811439991, + 0.009307538159191608, + 0.008047282695770264, + 0.0036511148791760206, + 0.0028872580733150244, + 0.011113599874079227, + 0.004593639634549618, + 0.007554714567959309, + 0.005850653629750013, + 0.006319822743535042, + 0.004331388510763645, + 0.007596844341605902, + 0.006200909614562988, + 0.011513195000588894, + 0.006188704166561365, + 0.0247181486338377, + 0.006281587295234203, + 0.007393174804747105, + 0.01144755631685257, + 0.009018069133162498, + 0.011176622472703457, + 0.024054186418652534, + 0.012621951289474964, + 0.011699097231030464, + 0.012096489779651165, + 0.0132930688560009, + 0.01291130855679512, + 0.010712282732129097, + 0.007651294115930796, + 0.004791887942701578, + 0.01511820126324892, + 0.009338065050542355, + 0.005947887431830168, + 0.006672492250800133, + 0.005419483408331871, + 0.007110476493835449, + 0.0051245056092739105, + 0.01043066568672657, + 0.004708974622189999, + 0.00489383190870285, + 0.005725560709834099, + 0.0034948578104376793, + 0.0062882606871426105, + 0.002635808428749442, + 0.005570817738771439, + 0.003947087563574314, + 0.004609738476574421, + 0.00434178626164794, + 0.012482009828090668, + 0.006195643916726112, + 0.027597524225711823, + 0.013722323812544346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.016086315736174583 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled skin and silver-white hair styled in a bob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding the handlebars of a pearly white mobility scooter with a matte black ergonomic grip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere cardigan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a close-up on her hands and slowly tilt upward to reveal her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a warm, out-of-focus blur of a sun-drenched room with green indoor plants creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023292383179068565 + ], + "opposite_logprob_tokens": [ + -3.773292303085327 + ], + "logprob_reward": -0.023292383179068565, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023292383179068565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03811780363321304, + 0.04722056910395622, + 0.2498861402273178, + 0.28663480281829834, + 0.030461685732007027, + 0.17290613055229187, + 0.0975722074508667, + 0.027032149955630302, + 0.02944788709282875, + 0.02072068117558956 + ], + "credit_map": [ + 0.004854774568229914, + 0.006945449858903885, + 0.005008156877011061, + 0.005465185735374689, + 0.004006873816251755, + 0.004074634984135628, + 0.004456114489585161, + 0.0033066144678741693, + 0.006367451976984739, + 0.005092781502753496, + 0.01022884901612997, + 0.00612040376290679, + 0.004138540010899305, + 0.004319541621953249, + 0.008062049746513367, + 0.0028909489046782255, + 0.010712016373872757, + 0.0904705822467804, + 0.016220886260271072, + 0.011361976154148579, + 0.015265432186424732, + 0.08938553184270859, + 0.008907719515264034, + 0.007561977952718735, + 0.03000066801905632, + 0.040113430470228195, + 0.04174785315990448, + 0.027151506394147873, + 0.03697005659341812, + 0.06127839908003807, + 0.03539570793509483, + 0.013977193273603916, + 0.0025934362784028053, + 0.0048166923224925995, + 0.004652400501072407, + 0.004547806456685066, + 0.003803845727816224, + 0.004153183661401272, + 0.003263922641053796, + 0.00263039767742157, + 0.01134345680475235, + 0.02048131451010704, + 0.011110156774520874, + 0.04687374085187912, + 0.013602236285805702, + 0.015945350751280785, + 0.02852972224354744, + 0.025020165368914604, + 0.010442998260259628, + 0.021404651924967766, + 0.011571883223950863, + 0.010514519177377224, + 0.010573933832347393, + 0.01190498098731041, + 0.011872558854520321, + 0.009286686778068542, + 0.0026541119441390038, + 0.0029988170135766268, + 0.0026917699724435806, + 0.006574692204594612, + 0.0026802285574376583, + 0.0033598309382796288, + 0.0030058289412409067, + 0.0030668710824102163, + 0.0036892590578645468, + 0.00500263087451458, + 0.003616300644353032, + 0.003528254572302103, + 0.003476290265098214, + 0.0022474040742963552, + 0.005128960125148296, + 0.002758786780759692, + 0.0018226462416350842, + 0.0016107000410556793, + 0.003469437826424837, + 0.0015901769511401653, + 0.0028310567140579224, + 0.002084236592054367, + 0.004708316642791033, + 0.0026041092351078987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.023292383179068565 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled skin and silver-white hair styled in a bob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding the handlebars of a pearly white mobility scooter with a matte black ergonomic grip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere cardigan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a close-up on her hands and slowly tilt upward to reveal her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a warm, out-of-focus blur of a sun-drenched room with green indoor plants creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017133570509031415 + ], + "opposite_logprob_tokens": [ + -6.376713275909424 + ], + "logprob_reward": -0.0017133570509031415, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017133570509031415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10481159389019012, + 0.06643231958150864, + 0.07349785417318344, + 0.13459040224552155, + 0.13981366157531738, + 0.13314767181873322, + 0.16085900366306305, + 0.07276301085948944, + 0.037823066115379333, + 0.07626139372587204 + ], + "credit_map": [ + 0.015391902066767216, + 0.018177716061472893, + 0.024151360616087914, + 0.011056007817387581, + 0.006088475231081247, + 0.008670126087963581, + 0.011262529529631138, + 0.010013478808104992, + 0.0082691116258502, + 0.010151155292987823, + 0.01295614056289196, + 0.008288213983178139, + 0.004911630880087614, + 0.005700796842575073, + 0.010219993069767952, + 0.005935278721153736, + 0.005241053644567728, + 0.021452557295560837, + 0.00549713522195816, + 0.006864209193736315, + 0.004298906773328781, + 0.02265612967312336, + 0.005457483697682619, + 0.0020303709898144007, + 0.005947261583060026, + 0.019936412572860718, + 0.00874410942196846, + 0.008763526566326618, + 0.03193361312150955, + 0.03865651413798332, + 0.014807197265326977, + 0.00580176617950201, + 0.013608520850539207, + 0.02604096569120884, + 0.015727126970887184, + 0.015665268525481224, + 0.01928768679499626, + 0.02739587053656578, + 0.009981009177863598, + 0.01210721954703331, + 0.010484564118087292, + 0.01979520358145237, + 0.009834005497395992, + 0.040860556066036224, + 0.008274376392364502, + 0.014755847863852978, + 0.01698201708495617, + 0.012161105871200562, + 0.016082625836133957, + 0.03716316446661949, + 0.01757526583969593, + 0.017462199553847313, + 0.015382885932922363, + 0.020396197214722633, + 0.02010178379714489, + 0.016694877296686172, + 0.008729948662221432, + 0.007280835881829262, + 0.012013143859803677, + 0.014469938352704048, + 0.0072507006116211414, + 0.008230643346905708, + 0.007698749657720327, + 0.007089052814990282, + 0.003937304485589266, + 0.005547875538468361, + 0.004629849921911955, + 0.004697451367974281, + 0.004698523785918951, + 0.0037893380504101515, + 0.00716623617336154, + 0.003356486326083541, + 0.005967539735138416, + 0.004985404200851917, + 0.007416103035211563, + 0.005020609125494957, + 0.012471925467252731, + 0.00523700937628746, + 0.024502674117684364, + 0.010660122148692608 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 3 + ], + "selected_reward": -0.0017133570509031415 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled skin and silver-white hair styled in a bob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding the handlebars of a pearly white mobility scooter with a matte black ergonomic grip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere cardigan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a close-up on her hands and slowly tilt upward to reveal her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a warm, out-of-focus blur of a sun-drenched room with green indoor plants creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640168234705925 + ], + "opposite_logprob_tokens": [ + -4.758640289306641 + ], + "logprob_reward": -0.008640168234705925, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640168234705925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04135353863239288, + 0.021409697830677032, + 0.08172482252120972, + 0.06212692707777023, + 0.08512978255748749, + 0.060460276901721954, + 0.3673595190048218, + 0.04849117249250412, + 0.15293188393115997, + 0.0790124237537384 + ], + "credit_map": [ + 0.00728677399456501, + 0.004745865240693092, + 0.009937198832631111, + 0.005076481029391289, + 0.002463981509208679, + 0.003402839880436659, + 0.004307151306420565, + 0.004133246373385191, + 0.003458982100710273, + 0.002720046089962125, + 0.0039970711804926395, + 0.002728617750108242, + 0.0020007456187158823, + 0.0018475806573405862, + 0.003231375478208065, + 0.001425279537215829, + 0.00312389456667006, + 0.030419394373893738, + 0.003948440309613943, + 0.0035485385451465845, + 0.0054178740829229355, + 0.029695585370063782, + 0.0031105855014175177, + 0.002460507210344076, + 0.003964642528444529, + 0.0087563656270504, + 0.0062417928129434586, + 0.007090033032000065, + 0.006484447978436947, + 0.010033106431365013, + 0.014557857066392899, + 0.004998684395104647, + 0.006161638535559177, + 0.02357144095003605, + 0.008585111238062382, + 0.013395813293755054, + 0.007163508329540491, + 0.010382326319813728, + 0.00761820375919342, + 0.008251738734543324, + 0.005825561936944723, + 0.012651694007217884, + 0.008045568130910397, + 0.012116281315684319, + 0.005244751460850239, + 0.005018370691686869, + 0.0064470721408724785, + 0.005110978614538908, + 0.03881436213850975, + 0.08612589538097382, + 0.042347345501184464, + 0.03859845921397209, + 0.037996917963027954, + 0.04380718618631363, + 0.043676190078258514, + 0.03599314019083977, + 0.007549984846264124, + 0.00619665440171957, + 0.004891492426395416, + 0.011250889860093594, + 0.003953087143599987, + 0.003945848438888788, + 0.005809966009110212, + 0.004893250297755003, + 0.014574208296835423, + 0.017564957961440086, + 0.015580423176288605, + 0.016219230368733406, + 0.02322922833263874, + 0.009028404019773006, + 0.04807046055793762, + 0.008664969354867935, + 0.00509383250027895, + 0.005154544021934271, + 0.008823764510452747, + 0.005329302046447992, + 0.01004788838326931, + 0.006441865116357803, + 0.02456924505531788, + 0.013551984913647175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 4 + ], + "selected_reward": -0.008640168234705925 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "focused close-up begins on the hands of an elderly woman, her skin delicately wrinkled with age, but her grip steady and cautious, her fingers wrapped around the sleek, matte black, ergonomic handlebars of a sleek, pearly white premium mobility scooter. Then, the camera slowly and gently tilts upward, gliding over her arm, draped in a soft, cream-colored cashmere cardigan. The movement continues, smoothly bringing her face into the beautifully composed portrait frame. Her wispy, silver-white hair, styled in a neatly combed bob, gently frames her face, her kind eyes framed by gentle crow's feet, and a contented, soft smile graces her lips. The background is a warm, out-of-focus blur of a sun-drenched room, the soft green shapes of indoor plants creating a delightful bokeh effect. As she gracefully settles into the vehicle's comfortable seat, the soft, diffuse morning light streaming from an unseen window illuminates her serene expression, capturing a quiet, optimistic moment of newfound independence and peaceful anticipation before she prepares for action.", + "seed": 8267, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_7/index_7599_seed_8267_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.01360140643082559, + "vlm_logprob": -0.01360140643082559, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.01360140643082559, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled skin and silver-white hair styled in a bob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding the handlebars of a pearly white mobility scooter with a matte black ergonomic grip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere cardigan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a close-up on her hands and slowly tilt upward to reveal her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a warm, out-of-focus blur of a sun-drenched room with green indoor plants creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007800947409123182 + ], + "opposite_logprob_tokens": [ + -5.007801055908203 + ], + "logprob_reward": -0.007800947409123182, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.007800947409123182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09906188398599625, + 0.02852938510477543, + 0.05646437779068947, + 0.05339508503675461, + 0.1445731222629547, + 0.15142665803432465, + 0.2149636596441269, + 0.10140091925859451, + 0.05670345574617386, + 0.09348151832818985 + ], + "credit_map": [ + 0.022865861654281616, + 0.009360459633171558, + 0.014768903143703938, + 0.004515017848461866, + 0.016039816662669182, + 0.016987547278404236, + 0.00521077960729599, + 0.009313504211604595, + 0.004881735425442457, + 0.004535870160907507, + 0.004795608576387167, + 0.0024197855964303017, + 0.003454666119068861, + 0.003926585894078016, + 0.0020125587470829487, + 0.0025025762151926756, + 0.005892621353268623, + 0.00524548627436161, + 0.0120637696236372, + 0.006609734147787094, + 0.005945701152086258, + 0.008404092863202095, + 0.0052461340092122555, + 0.0070568351075053215, + 0.0038857590407133102, + 0.003955218940973282, + 0.011875090189278126, + 0.0034936610609292984, + 0.0072179194539785385, + 0.00892542488873005, + 0.004548490978777409, + 0.009493515826761723, + 0.007946805097162724, + 0.007571078836917877, + 0.013049585744738579, + 0.02713790535926819, + 0.023599334061145782, + 0.01955888792872429, + 0.028474098071455956, + 0.017235424369573593, + 0.025939421728253365, + 0.015449626371264458, + 0.019658301025629044, + 0.019200200214982033, + 0.016694486141204834, + 0.014721202664077282, + 0.019170572981238365, + 0.020592840388417244, + 0.017748907208442688, + 0.035280704498291016, + 0.01243156660348177, + 0.02298481948673725, + 0.05033506453037262, + 0.033648278564214706, + 0.021010559052228928, + 0.021523749455809593, + 0.012556967325508595, + 0.017108801752328873, + 0.008931608870625496, + 0.01103017758578062, + 0.01614910550415516, + 0.008935901336371899, + 0.013894004747271538, + 0.012794352136552334, + 0.00784261990338564, + 0.004598400089889765, + 0.005345358978956938, + 0.005622152239084244, + 0.00850525964051485, + 0.005156469531357288, + 0.012133289128541946, + 0.007499902509152889, + 0.007176402490586042, + 0.016079654917120934, + 0.010598227381706238, + 0.01120690442621708, + 0.01798253506422043, + 0.01024396251887083, + 0.00891693402081728, + 0.011276897974312305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.007800947409123182 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled skin and silver-white hair styled in a bob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding the handlebars of a pearly white mobility scooter with a matte black ergonomic grip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere cardigan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a close-up on her hands and slowly tilt upward to reveal her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a warm, out-of-focus blur of a sun-drenched room with green indoor plants creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033736251294612885 + ], + "opposite_logprob_tokens": [ + -3.408736228942871 + ], + "logprob_reward": -0.033736251294612885, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033736251294612885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12706151604652405, + 0.07692054659128189, + 0.193332701921463, + 0.06556081771850586, + 0.0577298067510128, + 0.14829598367214203, + 0.07235478609800339, + 0.18842007219791412, + 0.04506754130125046, + 0.025256238877773285 + ], + "credit_map": [ + 0.015904689207673073, + 0.013983319513499737, + 0.01643410511314869, + 0.012892528437077999, + 0.02180328220129013, + 0.01782183162868023, + 0.010518196038901806, + 0.017703570425510406, + 0.017103878781199455, + 0.008400761522352695, + 0.011680102907121181, + 0.006571547593921423, + 0.00799867045134306, + 0.011584163643419743, + 0.006092808675020933, + 0.007488614413887262, + 0.01972712203860283, + 0.027730515226721764, + 0.028582166880369186, + 0.019871816039085388, + 0.02433280274271965, + 0.028726503252983093, + 0.02426326833665371, + 0.0200984887778759, + 0.003909537568688393, + 0.00505705876275897, + 0.011376800015568733, + 0.004798491019755602, + 0.01088041253387928, + 0.01220521330833435, + 0.007092297542840242, + 0.010240995325148106, + 0.005379306152462959, + 0.0051838550716638565, + 0.004890800919383764, + 0.006469027139246464, + 0.010216373950242996, + 0.007935397326946259, + 0.009392484091222286, + 0.008262560702860355, + 0.017087619751691818, + 0.008069458417594433, + 0.015464610420167446, + 0.00870241317898035, + 0.030622659251093864, + 0.007458627689629793, + 0.045733578503131866, + 0.015157024376094341, + 0.006530212704092264, + 0.009580815210938454, + 0.006514962762594223, + 0.00890950858592987, + 0.012353459373116493, + 0.014405467547476292, + 0.007829030975699425, + 0.006231328938156366, + 0.027614697813987732, + 0.02295754663646221, + 0.01653462089598179, + 0.026390280574560165, + 0.024479849264025688, + 0.014964296482503414, + 0.03201430290937424, + 0.023464487865567207, + 0.005405270494520664, + 0.003900175681337714, + 0.0051050940528512, + 0.006912860553711653, + 0.00716019794344902, + 0.00353077775798738, + 0.006830703932791948, + 0.006222460884600878, + 0.002598162740468979, + 0.0036851000040769577, + 0.002866882598027587, + 0.003412278601899743, + 0.0031202412210404873, + 0.0027742672245949507, + 0.003805840387940407, + 0.002993466332554817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 5 + ], + "selected_reward": -0.033736251294612885 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled skin and silver-white hair styled in a bob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding the handlebars of a pearly white mobility scooter with a matte black ergonomic grip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere cardigan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a close-up on her hands and slowly tilt upward to reveal her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a warm, out-of-focus blur of a sun-drenched room with green indoor plants creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014206817373633385 + ], + "opposite_logprob_tokens": [ + -4.264206886291504 + ], + "logprob_reward": -0.014206817373633385, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014206817373633385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09739907830953598, + 0.027022438123822212, + 0.08686664700508118, + 0.07395943254232407, + 0.3255460262298584, + 0.1683330088853836, + 0.09498515725135803, + 0.027251074090600014, + 0.06279775500297546, + 0.03583940118551254 + ], + "credit_map": [ + 0.022305035963654518, + 0.006576707120984793, + 0.014701123349368572, + 0.00549266766756773, + 0.01486651599407196, + 0.017007308080792427, + 0.006560439243912697, + 0.009889272972941399, + 0.0030892957001924515, + 0.004184074234217405, + 0.004672103561460972, + 0.0030232579447329044, + 0.0030299804639071226, + 0.0034153698943555355, + 0.0024801024701446295, + 0.0031282519921660423, + 0.008841448463499546, + 0.00922525767236948, + 0.02248336561024189, + 0.008061591535806656, + 0.008882046677172184, + 0.01673874445259571, + 0.005370111204683781, + 0.0072640771977603436, + 0.004307625815272331, + 0.0064498381689190865, + 0.013438574969768524, + 0.0059380969032645226, + 0.012672885321080685, + 0.01309935376048088, + 0.007507352624088526, + 0.010545700788497925, + 0.01771506480872631, + 0.02062019146978855, + 0.02824908308684826, + 0.048780713230371475, + 0.0597708486020565, + 0.04321518540382385, + 0.05915537476539612, + 0.048039522022008896, + 0.02208797261118889, + 0.014209812507033348, + 0.016161520034074783, + 0.02015458606183529, + 0.029812823981046677, + 0.02213919349014759, + 0.022797008976340294, + 0.020970070734620094, + 0.006821397226303816, + 0.010800972580909729, + 0.005327410995960236, + 0.008836551569402218, + 0.02503623627126217, + 0.016388962045311928, + 0.01063890103250742, + 0.011134719476103783, + 0.0035066555719822645, + 0.0045400336384773254, + 0.002822523470968008, + 0.006179070565849543, + 0.0022841612808406353, + 0.002002744236961007, + 0.0027606687508523464, + 0.0031552123837172985, + 0.01017401460558176, + 0.006036829669028521, + 0.007687459699809551, + 0.006364826578646898, + 0.011984916403889656, + 0.0050275810062885284, + 0.010005769319832325, + 0.005516353063285351, + 0.0037183554377406836, + 0.0045655155554413795, + 0.005724621936678886, + 0.003894653869792819, + 0.0049645728431642056, + 0.00411309115588665, + 0.005194187164306641, + 0.003664397168904543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.014206817373633385 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled skin and silver-white hair styled in a bob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding the handlebars of a pearly white mobility scooter with a matte black ergonomic grip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere cardigan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a close-up on her hands and slowly tilt upward to reveal her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a warm, out-of-focus blur of a sun-drenched room with green indoor plants creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036174117121845484 + ], + "opposite_logprob_tokens": [ + -5.628617286682129 + ], + "logprob_reward": -0.0036174117121845484, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036174117121845484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09129612892866135, + 0.027764588594436646, + 0.08565686643123627, + 0.17441973090171814, + 0.3007826507091522, + 0.12564176321029663, + 0.06266120821237564, + 0.08065302670001984, + 0.026290196925401688, + 0.02483384683728218 + ], + "credit_map": [ + 0.0185202956199646, + 0.008127324283123016, + 0.014432220719754696, + 0.0052042086608707905, + 0.011271405033767223, + 0.014958220534026623, + 0.0048246257938444614, + 0.013957817107439041, + 0.0032217109110206366, + 0.004645116627216339, + 0.004568294156342745, + 0.002656160620972514, + 0.0033226883970201015, + 0.0038417724426835775, + 0.0024905013851821423, + 0.003018346382305026, + 0.00945896003395319, + 0.007304870989173651, + 0.023335065692663193, + 0.008181009441614151, + 0.00609892001375556, + 0.014326177537441254, + 0.008278947323560715, + 0.008672911673784256, + 0.01194816268980503, + 0.010060076601803303, + 0.04288341477513313, + 0.010585247538983822, + 0.01668364368379116, + 0.03141460940241814, + 0.016575083136558533, + 0.034269485622644424, + 0.014799597673118114, + 0.013835419900715351, + 0.026391178369522095, + 0.05550004169344902, + 0.052113961428403854, + 0.041791655123233795, + 0.05963103100657463, + 0.03671977296471596, + 0.019026221707463264, + 0.007692576851695776, + 0.018417637795209885, + 0.009039676748216152, + 0.02404416725039482, + 0.009014609269797802, + 0.027283774688839912, + 0.011123101226985455, + 0.006087315734475851, + 0.013145625591278076, + 0.00486796023324132, + 0.007921288721263409, + 0.009130842983722687, + 0.009381620213389397, + 0.005990701261907816, + 0.006135846488177776, + 0.009720293805003166, + 0.011675356887280941, + 0.006085210479795933, + 0.008285071700811386, + 0.014505097642540932, + 0.008199093863368034, + 0.012657605111598969, + 0.009525292553007603, + 0.0037024624180048704, + 0.00227748672477901, + 0.0026084035634994507, + 0.0032032986637204885, + 0.0034394694957882166, + 0.002721028868108988, + 0.004821981303393841, + 0.0035160670522600412, + 0.0027421582490205765, + 0.002541821449995041, + 0.0036177481524646282, + 0.0025578425265848637, + 0.003824299667030573, + 0.003050603438168764, + 0.003395823761820793, + 0.003103549126535654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.0036174117121845484 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled skin and silver-white hair styled in a bob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding the handlebars of a pearly white mobility scooter with a matte black ergonomic grip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere cardigan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a close-up on her hands and slowly tilt upward to reveal her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a warm, out-of-focus blur of a sun-drenched room with green indoor plants creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008645604364573956 + ], + "opposite_logprob_tokens": [ + -4.758645534515381 + ], + "logprob_reward": -0.008645604364573956, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008645604364573956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049414001405239105, + 0.02011275477707386, + 0.03723474591970444, + 0.036267682909965515, + 0.27992013096809387, + 0.34446656703948975, + 0.03186009079217911, + 0.05007527023553848, + 0.06084500253200531, + 0.0898037850856781 + ], + "credit_map": [ + 0.011671490035951138, + 0.004202244803309441, + 0.006503426469862461, + 0.002686412539333105, + 0.00812870915979147, + 0.008200050331652164, + 0.003188547445461154, + 0.004833118990063667, + 0.0021611766424030066, + 0.0022022449411451817, + 0.002873830497264862, + 0.002767738653346896, + 0.002275726990774274, + 0.0028962462674826384, + 0.002273677848279476, + 0.002662111772224307, + 0.0034044680651277304, + 0.005371919833123684, + 0.007651531137526035, + 0.0030109784565865993, + 0.004293190315365791, + 0.006963543128222227, + 0.0028733822982758284, + 0.0036657301243394613, + 0.0020742490887641907, + 0.0029949937015771866, + 0.006663907319307327, + 0.004217629786580801, + 0.004702316131442785, + 0.005792089272290468, + 0.0039433944039046764, + 0.005879102740436792, + 0.018600190058350563, + 0.01818303018808365, + 0.022261913865804672, + 0.04995791241526604, + 0.04771997407078743, + 0.0363105870783329, + 0.053691837936639786, + 0.033194657415151596, + 0.07496363669633865, + 0.030636966228485107, + 0.06760526448488235, + 0.038297709077596664, + 0.026940934360027313, + 0.028344644233584404, + 0.04265328496694565, + 0.03502411022782326, + 0.003860334400087595, + 0.0046511199325323105, + 0.0036173572298139334, + 0.003415984334424138, + 0.005302696023136377, + 0.003210838418453932, + 0.004361878149211407, + 0.0034398806747049093, + 0.006089882459491491, + 0.006441950798034668, + 0.004219837952405214, + 0.006018811836838722, + 0.00738925114274025, + 0.00552427489310503, + 0.007470198441296816, + 0.006921062245965004, + 0.012412150390446186, + 0.005452968645840883, + 0.009037589654326439, + 0.005645556841045618, + 0.009761122055351734, + 0.0037544178776443005, + 0.009336679242551327, + 0.005444517359137535, + 0.007867441512644291, + 0.013668662868440151, + 0.011167568154633045, + 0.00936929415911436, + 0.013611442409455776, + 0.011064339429140091, + 0.011612046509981155, + 0.01144298817962408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 9 + ], + "selected_reward": -0.008645604364573956 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "focused close-up begins on the hands of an elderly woman, her skin delicately wrinkled with age, but her grip steady and cautious, her fingers wrapped around the sleek, matte black, ergonomic handlebars of a sleek, pearly white premium mobility scooter. Then, the camera slowly and gently tilts upward, gliding over her arm, draped in a soft, cream-colored cashmere cardigan. The movement continues, smoothly bringing her face into the beautifully composed portrait frame. Her wispy, silver-white hair, styled in a neatly combed bob, gently frames her face, her kind eyes framed by gentle crow's feet, and a contented, soft smile graces her lips. The background is a warm, out-of-focus blur of a sun-drenched room, the soft green shapes of indoor plants creating a delightful bokeh effect. As she gracefully settles into the vehicle's comfortable seat, the soft, diffuse morning light streaming from an unseen window illuminates her serene expression, capturing a quiet, optimistic moment of newfound independence and peaceful anticipation before she prepares for action.", + "seed": 8268, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_7/index_7599_seed_8268_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.009255174919962883, + "vlm_logprob": -0.009255174919962883, + "vlm_yes_no_margin": 5.025, + "vlm_token_credit": -0.009255174919962883, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled skin and silver-white hair styled in a bob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding the handlebars of a pearly white mobility scooter with a matte black ergonomic grip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere cardigan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a close-up on her hands and slowly tilt upward to reveal her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a warm, out-of-focus blur of a sun-drenched room with green indoor plants creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005028695799410343 + ], + "opposite_logprob_tokens": [ + -6.13002872467041 + ], + "logprob_reward": -0.005028695799410343, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.005028695799410343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19539296627044678, + 0.11867406964302063, + 0.07894052565097809, + 0.06095511093735695, + 0.10963721573352814, + 0.05339423939585686, + 0.07065950334072113, + 0.07018839567899704, + 0.12833848595619202, + 0.11381951719522476 + ], + "credit_map": [ + 0.029578357934951782, + 0.020413266494870186, + 0.027454612776637077, + 0.03177176043391228, + 0.01746448129415512, + 0.024495897814631462, + 0.02746559865772724, + 0.01674899086356163, + 0.012425702065229416, + 0.015475858002901077, + 0.010152093134820461, + 0.00958137959241867, + 0.03641878068447113, + 0.017623336985707283, + 0.007785972207784653, + 0.009210949763655663, + 0.011702331714332104, + 0.011671774089336395, + 0.009409564547240734, + 0.00893772765994072, + 0.006932291202247143, + 0.010309714823961258, + 0.00904804840683937, + 0.010929068550467491, + 0.00845556240528822, + 0.008986218832433224, + 0.007041075266897678, + 0.010357826016843319, + 0.006546748802065849, + 0.006080276798456907, + 0.005517017096281052, + 0.007970382459461689, + 0.014075896702706814, + 0.015545203350484371, + 0.01024926919490099, + 0.011724503710865974, + 0.01883676089346409, + 0.012124747969210148, + 0.009713557548820972, + 0.017367269843816757, + 0.005709590390324593, + 0.008176861330866814, + 0.008518065325915813, + 0.007326112128794193, + 0.007157109212130308, + 0.007233420852571726, + 0.00439427001401782, + 0.004878811538219452, + 0.005901920143514872, + 0.013286118395626545, + 0.005620521027594805, + 0.009133106097579002, + 0.011369056068360806, + 0.00950926635414362, + 0.005188162438571453, + 0.01065135095268488, + 0.004410270601511002, + 0.008949637413024902, + 0.01009966991841793, + 0.012364695779979229, + 0.010897025465965271, + 0.006260580383241177, + 0.008982104249298573, + 0.008224410936236382, + 0.009547819383442402, + 0.025337256491184235, + 0.0135429035872221, + 0.020028194412589073, + 0.018356885761022568, + 0.011888920329511166, + 0.011409026570618153, + 0.018227484077215195, + 0.011731867678463459, + 0.018635760992765427, + 0.015991469845175743, + 0.01455561351031065, + 0.015172287821769714, + 0.01227503176778555, + 0.012821833603084087, + 0.012635649181902409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.005028695799410343 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled skin and silver-white hair styled in a bob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding the handlebars of a pearly white mobility scooter with a matte black ergonomic grip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere cardigan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a close-up on her hands and slowly tilt upward to reveal her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a warm, out-of-focus blur of a sun-drenched room with green indoor plants creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0206122025847435 + ], + "opposite_logprob_tokens": [ + -3.8956122398376465 + ], + "logprob_reward": -0.0206122025847435, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.0206122025847435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2644840478897095, + 0.2725018560886383, + 0.18189473450183868, + 0.045111823827028275, + 0.07443643361330032, + 0.011728718876838684, + 0.02720540203154087, + 0.04720112308859825, + 0.048764899373054504, + 0.026670977473258972 + ], + "credit_map": [ + 0.03482673689723015, + 0.030887003988027573, + 0.023008573800325394, + 0.02298498898744583, + 0.03276299312710762, + 0.038378261029720306, + 0.03195048123598099, + 0.04968499019742012, + 0.02686009556055069, + 0.01995743438601494, + 0.017080485820770264, + 0.01643468253314495, + 0.09971661120653152, + 0.05999494716525078, + 0.016456758603453636, + 0.016000855714082718, + 0.024877166375517845, + 0.018590854480862617, + 0.026517709717154503, + 0.029985718429088593, + 0.014515695162117481, + 0.022521857172250748, + 0.019480228424072266, + 0.025405503809452057, + 0.00934793334454298, + 0.006334594916552305, + 0.005093806888908148, + 0.006219632923603058, + 0.0038766846992075443, + 0.0044707865454256535, + 0.004885179456323385, + 0.004883207380771637, + 0.012172644957900047, + 0.011846852488815784, + 0.009591520763933659, + 0.01091568823903799, + 0.007426219526678324, + 0.006346592213958502, + 0.008206085301935673, + 0.007930824533104897, + 0.0015464595053344965, + 0.0015815549995750189, + 0.0016102478839457035, + 0.0014406695263460279, + 0.00148588209412992, + 0.0013076133327558637, + 0.0014791755238547921, + 0.0012771164765581489, + 0.0019734378438442945, + 0.004286563955247402, + 0.003832629881799221, + 0.003969133365899324, + 0.004619174636900425, + 0.0032212836667895317, + 0.002818137174472213, + 0.0024850424379110336, + 0.0031552519649267197, + 0.005793237127363682, + 0.007007843814790249, + 0.009747914038598537, + 0.006359624210745096, + 0.005134699400514364, + 0.005051106214523315, + 0.004951445385813713, + 0.004861160181462765, + 0.007072974927723408, + 0.00955585204064846, + 0.005808690097182989, + 0.005546174943447113, + 0.005620401352643967, + 0.005282209720462561, + 0.00501743471249938, + 0.002891051582992077, + 0.0040052421391010284, + 0.003088586963713169, + 0.002969257766380906, + 0.0037006400525569916, + 0.0030308524146676064, + 0.002944342326372862, + 0.004041003994643688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0206122025847435 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled skin and silver-white hair styled in a bob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding the handlebars of a pearly white mobility scooter with a matte black ergonomic grip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere cardigan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a close-up on her hands and slowly tilt upward to reveal her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a warm, out-of-focus blur of a sun-drenched room with green indoor plants creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009792149998247623 + ], + "opposite_logprob_tokens": [ + -4.634792327880859 + ], + "logprob_reward": -0.009792149998247623, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009792149998247623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19935248792171478, + 0.11260513961315155, + 0.07629451155662537, + 0.15025541186332703, + 0.11037957668304443, + 0.024263733997941017, + 0.10021749883890152, + 0.08222751319408417, + 0.07520254701375961, + 0.06920159608125687 + ], + "credit_map": [ + 0.03206251189112663, + 0.026064187288284302, + 0.024394284933805466, + 0.02705223299562931, + 0.018552664667367935, + 0.026353491470217705, + 0.02025253139436245, + 0.02462058886885643, + 0.01509005669504404, + 0.012973598204553127, + 0.010446189902722836, + 0.009229025803506374, + 0.034296948462724686, + 0.015111729502677917, + 0.008050012402236462, + 0.007407581899315119, + 0.011718891561031342, + 0.012015530839562416, + 0.011387686245143414, + 0.01031535491347313, + 0.006121024955064058, + 0.009149649180471897, + 0.007870450615882874, + 0.007715919520705938, + 0.029896171763539314, + 0.0168011337518692, + 0.018690558150410652, + 0.025917423889040947, + 0.010595470666885376, + 0.017170825973153114, + 0.013308873400092125, + 0.017874954268336296, + 0.016869280487298965, + 0.017191670835018158, + 0.014270536601543427, + 0.015583583153784275, + 0.010726125910878181, + 0.008772018365561962, + 0.012608730234205723, + 0.014357631094753742, + 0.0023541899863630533, + 0.0026225941255688667, + 0.002573024947196245, + 0.0024573130067437887, + 0.0034983570221811533, + 0.0047695450484752655, + 0.0018220931524410844, + 0.0041666170582175255, + 0.006934282835572958, + 0.011142768897116184, + 0.016299154609441757, + 0.02505045011639595, + 0.013810092583298683, + 0.00719075184315443, + 0.012680315412580967, + 0.0071096839383244514, + 0.005905706901103258, + 0.010623135603964329, + 0.011498411186039448, + 0.012066748924553394, + 0.011335359886288643, + 0.008666908368468285, + 0.013424587436020374, + 0.008706656284630299, + 0.008610502816736698, + 0.010631457902491093, + 0.009568382985889912, + 0.011400134302675724, + 0.010660065338015556, + 0.009472521021962166, + 0.006620532833039761, + 0.00823894888162613, + 0.0068515571765601635, + 0.012024003081023693, + 0.01017531007528305, + 0.007666125427931547, + 0.00852755643427372, + 0.00761778699234128, + 0.007276984862983227, + 0.00906227808445692 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.009792149998247623 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled skin and silver-white hair styled in a bob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding the handlebars of a pearly white mobility scooter with a matte black ergonomic grip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere cardigan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a close-up on her hands and slowly tilt upward to reveal her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a warm, out-of-focus blur of a sun-drenched room with green indoor plants creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097638186067343 + ], + "opposite_logprob_tokens": [ + -5.5040974617004395 + ], + "logprob_reward": -0.004097638186067343, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097638186067343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2512945830821991, + 0.10570430010557175, + 0.15088482201099396, + 0.10992903262376785, + 0.07505667209625244, + 0.05549924820661545, + 0.07119815796613693, + 0.035286206752061844, + 0.06623252481222153, + 0.07891441881656647 + ], + "credit_map": [ + 0.03051300346851349, + 0.032992083579301834, + 0.02345278300344944, + 0.02934507466852665, + 0.023441176861524582, + 0.04790070280432701, + 0.02602520026266575, + 0.0376245342195034, + 0.011601163074374199, + 0.010313126258552074, + 0.008351253345608711, + 0.008027198724448681, + 0.03595667704939842, + 0.01754213310778141, + 0.007503907661885023, + 0.006408847868442535, + 0.02842033840715885, + 0.029944494366645813, + 0.019293775781989098, + 0.021178463473916054, + 0.010701010935008526, + 0.02035192772746086, + 0.009453958831727505, + 0.011540863662958145, + 0.01934291422367096, + 0.012743296101689339, + 0.014743889681994915, + 0.01850934699177742, + 0.008905730210244656, + 0.012926142662763596, + 0.009368238970637321, + 0.013389470987021923, + 0.011592874303460121, + 0.010805748403072357, + 0.007436358369886875, + 0.008907424286007881, + 0.0108712837100029, + 0.012751495465636253, + 0.006367653142660856, + 0.006323833018541336, + 0.010636422783136368, + 0.007445898838341236, + 0.007924598641693592, + 0.005249937996268272, + 0.009296861477196217, + 0.006661624182015657, + 0.0036248236428946257, + 0.0046590836718678474, + 0.003775449236854911, + 0.010098873637616634, + 0.010885790921747684, + 0.012152720242738724, + 0.011551435105502605, + 0.010714237578213215, + 0.005207992624491453, + 0.00681166211143136, + 0.0036869694013148546, + 0.00464374665170908, + 0.0037110024131834507, + 0.0056763505563139915, + 0.004361002240329981, + 0.004879779182374477, + 0.0030199293978512287, + 0.005307424813508987, + 0.008136987686157227, + 0.00787152349948883, + 0.007705845404416323, + 0.008331888355314732, + 0.0063957530073821545, + 0.008849470876157284, + 0.006542125716805458, + 0.01239892840385437, + 0.007556051481515169, + 0.0088205486536026, + 0.012209572829306126, + 0.009267239831387997, + 0.009805387817323208, + 0.008742534555494785, + 0.008217322640120983, + 0.014295761473476887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.004097638186067343 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled skin and silver-white hair styled in a bob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding the handlebars of a pearly white mobility scooter with a matte black ergonomic grip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere cardigan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a close-up on her hands and slowly tilt upward to reveal her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a warm, out-of-focus blur of a sun-drenched room with green indoor plants creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006745188031345606 + ], + "opposite_logprob_tokens": [ + -5.006745338439941 + ], + "logprob_reward": -0.006745188031345606, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006745188031345606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10074764490127563, + 0.1903529167175293, + 0.06487226486206055, + 0.07875023037195206, + 0.0736350491642952, + 0.09611666947603226, + 0.07130417227745056, + 0.1699109673500061, + 0.09317778795957565, + 0.061132270842790604 + ], + "credit_map": [ + 0.012196703813970089, + 0.01526118628680706, + 0.012345833703875542, + 0.012939797714352608, + 0.008795295841991901, + 0.016775181517004967, + 0.010797915980219841, + 0.011635737493634224, + 0.02124219387769699, + 0.021062973886728287, + 0.013407363556325436, + 0.01389307714998722, + 0.06486786901950836, + 0.03371751680970192, + 0.01032670121639967, + 0.01183524914085865, + 0.012134658172726631, + 0.011539491824805737, + 0.009554929099977016, + 0.007702539674937725, + 0.004398511257022619, + 0.008452245034277439, + 0.006055969744920731, + 0.005033923778682947, + 0.009055324830114841, + 0.007704094052314758, + 0.008814360946416855, + 0.008413597010076046, + 0.01005096361041069, + 0.009849452413618565, + 0.010013546794652939, + 0.014848902821540833, + 0.007479431573301554, + 0.01065638568252325, + 0.008442829363048077, + 0.012073390185832977, + 0.008659781888127327, + 0.00702933594584465, + 0.009202892892062664, + 0.010091003961861134, + 0.01097038108855486, + 0.010123849846422672, + 0.014742493629455566, + 0.010154333896934986, + 0.015231532044708729, + 0.012208741158246994, + 0.006491539068520069, + 0.016193803399801254, + 0.005902685225009918, + 0.010613076388835907, + 0.00848884042352438, + 0.010662706568837166, + 0.009180719032883644, + 0.00977136380970478, + 0.008780170232057571, + 0.007904618978500366, + 0.014841544441878796, + 0.02143370360136032, + 0.018641207367181778, + 0.018460558727383614, + 0.022894171997904778, + 0.02782192826271057, + 0.02008444257080555, + 0.025733426213264465, + 0.008808397687971592, + 0.010339171625673771, + 0.012059731408953667, + 0.010067757219076157, + 0.010412427596747875, + 0.010156121104955673, + 0.012474368326365948, + 0.018859824165701866, + 0.004927538800984621, + 0.009283952414989471, + 0.008522659540176392, + 0.00826514232903719, + 0.00778605230152607, + 0.006232001818716526, + 0.008086609654128551, + 0.008028318174183369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.006745188031345606 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "focused close-up begins on the hands of an elderly woman, her skin delicately wrinkled with age, but her grip steady and cautious, her fingers wrapped around the sleek, matte black, ergonomic handlebars of a sleek, pearly white premium mobility scooter. Then, the camera slowly and gently tilts upward, gliding over her arm, draped in a soft, cream-colored cashmere cardigan. The movement continues, smoothly bringing her face into the beautifully composed portrait frame. Her wispy, silver-white hair, styled in a neatly combed bob, gently frames her face, her kind eyes framed by gentle crow's feet, and a contented, soft smile graces her lips. The background is a warm, out-of-focus blur of a sun-drenched room, the soft green shapes of indoor plants creating a delightful bokeh effect. As she gracefully settles into the vehicle's comfortable seat, the soft, diffuse morning light streaming from an unseen window illuminates her serene expression, capturing a quiet, optimistic moment of newfound independence and peaceful anticipation before she prepares for action.", + "seed": 8269, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_7/index_7599_seed_8269_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.008575621992349625, + "vlm_logprob": -0.008575621992349625, + "vlm_yes_no_margin": 5.05, + "vlm_token_credit": -0.008575621992349625, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled skin and silver-white hair styled in a bob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding the handlebars of a pearly white mobility scooter with a matte black ergonomic grip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere cardigan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a close-up on her hands and slowly tilt upward to reveal her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a warm, out-of-focus blur of a sun-drenched room with green indoor plants creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0044069793075323105 + ], + "opposite_logprob_tokens": [ + -6.129406929016113 + ], + "logprob_reward": -0.0044069793075323105, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0044069793075323105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14427325129508972, + 0.07633846998214722, + 0.20461884140968323, + 0.14848260581493378, + 0.04172644764184952, + 0.0663389340043068, + 0.05800829455256462, + 0.10481301695108414, + 0.07884630560874939, + 0.0765538290143013 + ], + "credit_map": [ + 0.028336988762021065, + 0.02377847209572792, + 0.024117404595017433, + 0.01145266741514206, + 0.013491670601069927, + 0.009910689666867256, + 0.022161152213811874, + 0.011024191975593567, + 0.007097186986356974, + 0.009584760293364525, + 0.012713203206658363, + 0.014688343740999699, + 0.009190576151013374, + 0.005015531089156866, + 0.008550790138542652, + 0.00949807744473219, + 0.027843965217471123, + 0.04175626114010811, + 0.033818092197179794, + 0.029238706454634666, + 0.022464364767074585, + 0.013395850546658039, + 0.020951874554157257, + 0.015149728395044804, + 0.01222219504415989, + 0.05549103394150734, + 0.00950449239462614, + 0.007677722256630659, + 0.010897569358348846, + 0.0342397540807724, + 0.008451851084828377, + 0.009997979737818241, + 0.005760528147220612, + 0.00797856692224741, + 0.005346815567463636, + 0.004480373580008745, + 0.006274163722991943, + 0.004920300096273422, + 0.003034838242456317, + 0.003930860199034214, + 0.00582543108612299, + 0.006658913567662239, + 0.008511028252542019, + 0.008784263394773006, + 0.010750637389719486, + 0.006182503886520863, + 0.012473982758820057, + 0.007152179721742868, + 0.003719805274158716, + 0.007038818672299385, + 0.007686438038945198, + 0.0077760168351233006, + 0.007048337254673243, + 0.008423756808042526, + 0.010191340930759907, + 0.006123778410255909, + 0.011268618516623974, + 0.01898396573960781, + 0.012459373101592064, + 0.012328834272921085, + 0.01767384260892868, + 0.008154261857271194, + 0.012705675326287746, + 0.011238442733883858, + 0.011959399096667767, + 0.01040063239634037, + 0.014001845382153988, + 0.007760236505419016, + 0.012697281315922737, + 0.007782879285514355, + 0.007754919584840536, + 0.006489116232842207, + 0.008378840051591396, + 0.009810841642320156, + 0.009222595952451229, + 0.011087652295827866, + 0.0162159726023674, + 0.006621821783483028, + 0.007550492417067289, + 0.007665608078241348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.0044069793075323105 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled skin and silver-white hair styled in a bob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding the handlebars of a pearly white mobility scooter with a matte black ergonomic grip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere cardigan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a close-up on her hands and slowly tilt upward to reveal her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a warm, out-of-focus blur of a sun-drenched room with green indoor plants creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014220098033547401 + ], + "opposite_logprob_tokens": [ + -4.264220237731934 + ], + "logprob_reward": -0.014220098033547401, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014220098033547401, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4621518850326538, + 0.041306182742118835, + 0.1931266039609909, + 0.04568905383348465, + 0.08437936007976532, + 0.03739100694656372, + 0.03470828756690025, + 0.05134119465947151, + 0.03135966137051582, + 0.018546700477600098 + ], + "credit_map": [ + 0.1270691156387329, + 0.06316306442022324, + 0.10468774288892746, + 0.03014533407986164, + 0.02721530757844448, + 0.021140415221452713, + 0.06026535481214523, + 0.028465552255511284, + 0.004038257524371147, + 0.007228015456348658, + 0.008541245013475418, + 0.00545648206025362, + 0.005471516866236925, + 0.0032018686179071665, + 0.0031830056104809046, + 0.004185788799077272, + 0.0242106132209301, + 0.04040969908237457, + 0.03473518043756485, + 0.02403365634381771, + 0.02088118903338909, + 0.01199117861688137, + 0.02161157689988613, + 0.01525350846350193, + 0.005206072703003883, + 0.01466325856745243, + 0.0036659210454672575, + 0.0031090041156858206, + 0.004179979674518108, + 0.008447726257145405, + 0.0034011495299637318, + 0.003015941707417369, + 0.017242977395653725, + 0.011083030141890049, + 0.008793625049293041, + 0.009009257890284061, + 0.011979907751083374, + 0.009410182945430279, + 0.0065147061832249165, + 0.010345668531954288, + 0.00345089309848845, + 0.003580512711778283, + 0.004500681534409523, + 0.00562063604593277, + 0.005655919201672077, + 0.006271916441619396, + 0.004663273226469755, + 0.003647175617516041, + 0.0028130009304732084, + 0.006288313306868076, + 0.0033072337973862886, + 0.005399967543780804, + 0.004045473877340555, + 0.004200730938464403, + 0.004737637005746365, + 0.003915929235517979, + 0.005967100150883198, + 0.011032101698219776, + 0.008565963245928288, + 0.00670118210837245, + 0.0055779884569346905, + 0.00400817533954978, + 0.0048926956951618195, + 0.004595988430082798, + 0.004215451888740063, + 0.0032661384902894497, + 0.004636657889932394, + 0.003496504155918956, + 0.005073840729892254, + 0.00278984266333282, + 0.004678385332226753, + 0.003202838357537985, + 0.0018530545057728887, + 0.002447917126119137, + 0.0021327026188373566, + 0.0036444710567593575, + 0.0024602487683296204, + 0.001587345963343978, + 0.002204084536060691, + 0.002216875785961747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.014220098033547401 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled skin and silver-white hair styled in a bob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding the handlebars of a pearly white mobility scooter with a matte black ergonomic grip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere cardigan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a close-up on her hands and slowly tilt upward to reveal her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a warm, out-of-focus blur of a sun-drenched room with green indoor plants creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014204349368810654 + ], + "opposite_logprob_tokens": [ + -4.264204502105713 + ], + "logprob_reward": -0.014204349368810654, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014204349368810654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10170525312423706, + 0.06389500200748444, + 0.08633831888437271, + 0.06491685658693314, + 0.12343684583902359, + 0.03810827061533928, + 0.16171103715896606, + 0.19993074238300323, + 0.11513889580965042, + 0.044818729162216187 + ], + "credit_map": [ + 0.014937673695385456, + 0.013934886083006859, + 0.01797894388437271, + 0.012297050096094608, + 0.009427937678992748, + 0.007618291769176722, + 0.015782169997692108, + 0.009728300385177135, + 0.007685263641178608, + 0.01220210362225771, + 0.009002875536680222, + 0.010652408935129642, + 0.00789234321564436, + 0.004632994998246431, + 0.0059545631520450115, + 0.005872451234608889, + 0.015117570757865906, + 0.01604722999036312, + 0.014882739633321762, + 0.012168665416538715, + 0.008670065551996231, + 0.005373961757868528, + 0.008692148141562939, + 0.005385936703532934, + 0.006629112642258406, + 0.02186594158411026, + 0.004564272239804268, + 0.0028166507836431265, + 0.0026878658682107925, + 0.019278697669506073, + 0.0026703220792114735, + 0.004403991624712944, + 0.0249676201492548, + 0.017600739374756813, + 0.014683173969388008, + 0.013963939622044563, + 0.016854140907526016, + 0.013126784935593605, + 0.01047706138342619, + 0.011763389222323895, + 0.0039653838612139225, + 0.0038577611558139324, + 0.004430779721587896, + 0.005053339526057243, + 0.007445022463798523, + 0.00502868602052331, + 0.004425265360623598, + 0.0039020320400595665, + 0.011744198389351368, + 0.028320834040641785, + 0.023195259273052216, + 0.02423868700861931, + 0.014300110749900341, + 0.01726311258971691, + 0.029691999778151512, + 0.01295683067291975, + 0.024342065677046776, + 0.028113922104239464, + 0.03838839754462242, + 0.0232528205960989, + 0.022603189572691917, + 0.023063721135258675, + 0.01948334462940693, + 0.02068328484892845, + 0.011834954842925072, + 0.010932276025414467, + 0.01937243528664112, + 0.016630059108138084, + 0.016638027504086494, + 0.00729558989405632, + 0.019093846902251244, + 0.013341709971427917, + 0.00413871044293046, + 0.006417263764888048, + 0.005331964232027531, + 0.01145007461309433, + 0.005058431066572666, + 0.0024969002697616816, + 0.006423516198992729, + 0.003501870669424534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 4 + ], + "selected_reward": -0.014204349368810654 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled skin and silver-white hair styled in a bob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding the handlebars of a pearly white mobility scooter with a matte black ergonomic grip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere cardigan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a close-up on her hands and slowly tilt upward to reveal her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a warm, out-of-focus blur of a sun-drenched room with green indoor plants creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004094195086508989 + ], + "opposite_logprob_tokens": [ + -5.504094123840332 + ], + "logprob_reward": -0.004094195086508989, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004094195086508989, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4530563950538635, + 0.04416084662079811, + 0.04696645215153694, + 0.04317884147167206, + 0.10959254205226898, + 0.15991529822349548, + 0.052261337637901306, + 0.03540698066353798, + 0.026518214493989944, + 0.028943104669451714 + ], + "credit_map": [ + 0.1361471265554428, + 0.05609869211912155, + 0.11692944914102554, + 0.027695488184690475, + 0.025761762633919716, + 0.014411685056984425, + 0.056805990636348724, + 0.01920619048178196, + 0.005495142191648483, + 0.007703425362706184, + 0.007783771958202124, + 0.007149888202548027, + 0.004002932924777269, + 0.0031964178197085857, + 0.004853567574173212, + 0.0039756991900503635, + 0.005676464643329382, + 0.006961122155189514, + 0.010612744837999344, + 0.005678035318851471, + 0.0049265106208622456, + 0.00230819801799953, + 0.006835554260760546, + 0.003967822063714266, + 0.005894750822335482, + 0.010570312850177288, + 0.004028912167996168, + 0.0030521471053361893, + 0.0023403246887028217, + 0.012100465595722198, + 0.003132502781227231, + 0.0020594208035618067, + 0.01762569695711136, + 0.0162836704403162, + 0.011426607146859169, + 0.012202776037156582, + 0.013329830951988697, + 0.012244205921888351, + 0.010516979731619358, + 0.01596277579665184, + 0.01525486633181572, + 0.015087728388607502, + 0.016577940434217453, + 0.01927591860294342, + 0.024749625474214554, + 0.03479326143860817, + 0.016542386263608932, + 0.01763356849551201, + 0.003043968928977847, + 0.006479846313595772, + 0.006646669935435057, + 0.009455915540456772, + 0.005276123993098736, + 0.005928055848926306, + 0.009859723038971424, + 0.005571033339947462, + 0.004671164322644472, + 0.005468134302645922, + 0.0053046573884785175, + 0.004877772647887468, + 0.004564964212477207, + 0.003772514406591654, + 0.0034820460714399815, + 0.0032657303381711245, + 0.0028896937146782875, + 0.003626164048910141, + 0.0032165406737476587, + 0.0027655763551592827, + 0.004663506057113409, + 0.0026162308640778065, + 0.0037942170165479183, + 0.0029462850652635098, + 0.0029844646342098713, + 0.003648066194728017, + 0.003601829055696726, + 0.005076705943793058, + 0.004625250585377216, + 0.002537054242566228, + 0.0032940630335360765, + 0.0031756702810525894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.004094195086508989 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled skin and silver-white hair styled in a bob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding the handlebars of a pearly white mobility scooter with a matte black ergonomic grip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere cardigan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a close-up on her hands and slowly tilt upward to reveal her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a warm, out-of-focus blur of a sun-drenched room with green indoor plants creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005952488165348768 + ], + "opposite_logprob_tokens": [ + -5.13095235824585 + ], + "logprob_reward": -0.005952488165348768, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005952488165348768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2920992076396942, + 0.045016542077064514, + 0.0895514041185379, + 0.0650629997253418, + 0.1162998229265213, + 0.11142896860837936, + 0.15524017810821533, + 0.048826538026332855, + 0.05405981093645096, + 0.022414555773139 + ], + "credit_map": [ + 0.08420469611883163, + 0.03338528797030449, + 0.08106192946434021, + 0.01799691841006279, + 0.015256579965353012, + 0.008695718832314014, + 0.03992725536227226, + 0.011570804752409458, + 0.004081642720848322, + 0.007057896815240383, + 0.006215150002390146, + 0.008972687646746635, + 0.006512071006000042, + 0.0025892318226397038, + 0.005621231626719236, + 0.00396663136780262, + 0.009792677126824856, + 0.018787788227200508, + 0.015047320164740086, + 0.013407682999968529, + 0.009935006499290466, + 0.005520322360098362, + 0.009630987420678139, + 0.007429616991430521, + 0.005770533345639706, + 0.021273203194141388, + 0.004387990105897188, + 0.003011460416018963, + 0.0035042064264416695, + 0.018306080251932144, + 0.0040841251611709595, + 0.0047254012897610664, + 0.023090608417987823, + 0.015673479065299034, + 0.013143337331712246, + 0.012797167524695396, + 0.013597054407000542, + 0.012395737692713737, + 0.00990153755992651, + 0.015700900927186012, + 0.009442022070288658, + 0.012750743888318539, + 0.01086641289293766, + 0.013265474699437618, + 0.02639014646410942, + 0.01934926025569439, + 0.009599518962204456, + 0.009765386581420898, + 0.00849002692848444, + 0.029266491532325745, + 0.015012478455901146, + 0.029977913945913315, + 0.013671452179551125, + 0.022046351805329323, + 0.020349593833088875, + 0.01642586663365364, + 0.005337078589946032, + 0.009011072106659412, + 0.006900876760482788, + 0.006662145722657442, + 0.007964247837662697, + 0.004130300134420395, + 0.004929184913635254, + 0.0038916333578526974, + 0.005442824214696884, + 0.006097720470279455, + 0.007440219633281231, + 0.008320340886712074, + 0.007314847316592932, + 0.004611878655850887, + 0.008577077649533749, + 0.006254899315536022, + 0.0022396037820726633, + 0.002436829498037696, + 0.002627958543598652, + 0.004720429424196482, + 0.0037160669453442097, + 0.0012878573033958673, + 0.0033121092710644007, + 0.0020737010054290295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.005952488165348768 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_8/index_tensor([1462, 1462])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_8/index_tensor([1462, 1462])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2709875786d8e1522f0a1a9ae6c013c1fa603fc3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_8/index_tensor([1462, 1462])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "cinematic, static, extreme close-up shot from a slightly elevated angle initially depicts a vintage monochrome ticket, its details rendered with the fidelity of a vintage photograph. Set against a stark, non-reflective black background, the ticket's thick, tan paper displays authentic wear: frayed edges, softened corners, and a prominent center fold. The faded, nearly illegible charcoal-gray font and the faint outline of the serial number are visible beneath a soft light source that sweeps across the surface, carving subtle shadows into the creases and highlighting the fiber texture. Suddenly, the stillness of the shot breaks, revealing a web of fine, spider-web-like cracks creeping across the ticket's face. Then, the camera begins a slow, steady zoom as the ticket inexorably fractures into countless jagged fragments. Rather than falling, these fragments delicately drift, separating and drifting into the surrounding black void, as if carried by a gentle, invisible current. Each piece gracefully tumbles along its own path, creating a poignant and deeply nostalgic atmosphere, the slow disintegration of a cherished memory.", + "seed": 13994, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_8/index_1462_seed_13994_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.049189230799674986, + "vlm_logprob": -0.049189230799674986, + "vlm_yes_no_margin": 3.675000023841858, + "vlm_token_credit": -0.049189230799674986, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vintage monochrome ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ticket display visible wear including frayed edges, softened corners, and a prominent center fold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine, spider-web-like cracks appear on the ticket's surface and does the ticket fracture into jagged fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady zoom while the fragments drift and tumble into the black void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024737132713198662 + ], + "opposite_logprob_tokens": [ + -4.149737358093262 + ], + "logprob_reward": -0.024737132713198662, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.024737132713198662, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16510377824306488, + 0.06712490320205688, + 0.24385644495487213, + 0.18090802431106567, + 0.1005728617310524, + 0.03885681927204132, + 0.033708807080984116, + 0.052884653210639954, + 0.04528176411986351, + 0.07170185446739197 + ], + "credit_map": [ + 0.01955356076359749, + 0.0217712614685297, + 0.01683097891509533, + 0.02313816361129284, + 0.025568852201104164, + 0.014853416942059994, + 0.029079139232635498, + 0.014308417215943336, + 0.009450226090848446, + 0.010767568834125996, + 0.005728533491492271, + 0.008629316464066505, + 0.010763400234282017, + 0.00700344517827034, + 0.008360474370419979, + 0.006421938538551331, + 0.018902592360973358, + 0.03086186572909355, + 0.037506576627492905, + 0.02850678376853466, + 0.06942057609558105, + 0.016084741801023483, + 0.029213331639766693, + 0.01336000207811594, + 0.018433917313814163, + 0.029833940789103508, + 0.025364456698298454, + 0.0386948361992836, + 0.018437223508954048, + 0.019850004464387894, + 0.02132934331893921, + 0.008964324370026588, + 0.009982145391404629, + 0.014318245463073254, + 0.008031117729842663, + 0.017665104940533638, + 0.014325981959700584, + 0.00847561378031969, + 0.01723342202603817, + 0.010541236028075218, + 0.0041084070689976215, + 0.00547264888882637, + 0.0032116917427629232, + 0.007960965856909752, + 0.006055692210793495, + 0.002393797505646944, + 0.006072700023651123, + 0.00358091969974339, + 0.004766472149640322, + 0.004775312263518572, + 0.002542741596698761, + 0.004329873714596033, + 0.00566409295424819, + 0.004057111218571663, + 0.0049842470325529575, + 0.0025889608077704906, + 0.006047520786523819, + 0.012643605470657349, + 0.0037463579792529345, + 0.005839163903146982, + 0.012609435245394707, + 0.0030386969447135925, + 0.005240641999989748, + 0.003719234373420477, + 0.004313143901526928, + 0.008098118007183075, + 0.0031449769157916307, + 0.006494409870356321, + 0.005317980423569679, + 0.004661076702177525, + 0.006323383655399084, + 0.006928678136318922, + 0.011442960239946842, + 0.00967346504330635, + 0.005514668766409159, + 0.008577256463468075, + 0.008392511866986752, + 0.011201860383152962, + 0.008448726497590542, + 0.008450410328805447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.024737132713198662 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vintage monochrome ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ticket display visible wear including frayed edges, softened corners, and a prominent center fold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine, spider-web-like cracks appear on the ticket's surface and does the ticket fracture into jagged fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady zoom while the fragments drift and tumble into the black void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14285863935947418 + ], + "opposite_logprob_tokens": [ + -2.0178587436676025 + ], + "logprob_reward": -0.14285863935947418, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14285863935947418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1297459751367569, + 0.06235402822494507, + 0.14232461154460907, + 0.12204597145318985, + 0.1507442742586136, + 0.06647630780935287, + 0.025909848511219025, + 0.09050939977169037, + 0.05465886369347572, + 0.15523076057434082 + ], + "credit_map": [ + 0.013188653625547886, + 0.018121009692549706, + 0.010016771033406258, + 0.02031318098306656, + 0.017879057675600052, + 0.016918696463108063, + 0.020961349830031395, + 0.012347262352705002, + 0.006994481198489666, + 0.01038672961294651, + 0.004018357954919338, + 0.007327950559556484, + 0.01162345614284277, + 0.006931663956493139, + 0.00819048099219799, + 0.0068809096701443195, + 0.013449481688439846, + 0.021049190312623978, + 0.02056705206632614, + 0.019509319216012955, + 0.03024217113852501, + 0.00905679352581501, + 0.01902882568538189, + 0.009421765804290771, + 0.01333297323435545, + 0.03213764354586601, + 0.011491981334984303, + 0.023915300145745277, + 0.007954958826303482, + 0.011096886359155178, + 0.015087166801095009, + 0.0070290593430399895, + 0.01715623401105404, + 0.01837068796157837, + 0.011944658122956753, + 0.02609369531273842, + 0.022717000916600227, + 0.01531890220940113, + 0.02264905907213688, + 0.016494033858180046, + 0.005283865612000227, + 0.00912919919937849, + 0.004484259989112616, + 0.015801386907696724, + 0.009139432571828365, + 0.005047947634011507, + 0.013062897138297558, + 0.0045273154973983765, + 0.0031159876380115747, + 0.0037893650587648153, + 0.0026229589711874723, + 0.0035888913553208113, + 0.0036005289293825626, + 0.002104763174429536, + 0.0042007300071418285, + 0.0028866231441497803, + 0.0076220897026360035, + 0.020822325721383095, + 0.0074293059296905994, + 0.011068280786275864, + 0.019956694915890694, + 0.0067919958382844925, + 0.010240200906991959, + 0.006578515283763409, + 0.0050992281176149845, + 0.008357543498277664, + 0.00443584518507123, + 0.006409706547856331, + 0.007154996506869793, + 0.00846575666218996, + 0.006536193657666445, + 0.008199593052268028, + 0.020698213949799538, + 0.022933930158615112, + 0.011434793472290039, + 0.019148539751768112, + 0.01987297274172306, + 0.021371547132730484, + 0.020557675510644913, + 0.019213100895285606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 2 + ], + "selected_reward": -0.14285863935947418 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vintage monochrome ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ticket display visible wear including frayed edges, softened corners, and a prominent center fold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine, spider-web-like cracks appear on the ticket's surface and does the ticket fracture into jagged fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady zoom while the fragments drift and tumble into the black void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00865529477596283 + ], + "opposite_logprob_tokens": [ + -4.758655071258545 + ], + "logprob_reward": -0.00865529477596283, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00865529477596283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12113526463508606, + 0.06196347996592522, + 0.12349829077720642, + 0.20590469241142273, + 0.07577064633369446, + 0.13213422894477844, + 0.03854338824748993, + 0.047248195856809616, + 0.07116296142339706, + 0.12263886630535126 + ], + "credit_map": [ + 0.013964608311653137, + 0.013318364508450031, + 0.011762458831071854, + 0.016083786264061928, + 0.03061865270137787, + 0.011768853291869164, + 0.014535642229020596, + 0.009082899428904057, + 0.005661619361490011, + 0.008118955418467522, + 0.004469275940209627, + 0.008490960113704205, + 0.012447146698832512, + 0.0048075588420033455, + 0.009815244935452938, + 0.008152720518410206, + 0.00961978081613779, + 0.021039411425590515, + 0.012363218702375889, + 0.0165383480489254, + 0.015017351135611534, + 0.011504756286740303, + 0.018057571724057198, + 0.019357850775122643, + 0.01316514890640974, + 0.06445219367742538, + 0.038253139704465866, + 0.016020040959119797, + 0.028813844546675682, + 0.014376267790794373, + 0.02173754759132862, + 0.009086515754461288, + 0.005735000595450401, + 0.01257339958101511, + 0.0047478401102125645, + 0.015378153882920742, + 0.009890575893223286, + 0.004102753475308418, + 0.01742011122405529, + 0.0059228138998150826, + 0.01650260202586651, + 0.022764114663004875, + 0.009035135619342327, + 0.024150483310222626, + 0.01638149283826351, + 0.011775895953178406, + 0.022719936445355415, + 0.008804576471447945, + 0.0038163140416145325, + 0.006260586436837912, + 0.0026948475278913975, + 0.0043724351562559605, + 0.005855361465364695, + 0.002708044368773699, + 0.007571582682430744, + 0.005264217499643564, + 0.006889721378684044, + 0.009218530729413033, + 0.0023893604520708323, + 0.004567072726786137, + 0.009762275032699108, + 0.004453378263860941, + 0.006089876871556044, + 0.0038779787719249725, + 0.006877517327666283, + 0.01348808128386736, + 0.006035507190972567, + 0.00976562686264515, + 0.008739125914871693, + 0.006223862059414387, + 0.010033167898654938, + 0.010000072419643402, + 0.010182451456785202, + 0.016011035069823265, + 0.00947592593729496, + 0.014610614627599716, + 0.017530538141727448, + 0.014243642799556255, + 0.02613465115427971, + 0.014450008049607277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.00865529477596283 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vintage monochrome ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ticket display visible wear including frayed edges, softened corners, and a prominent center fold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine, spider-web-like cracks appear on the ticket's surface and does the ticket fracture into jagged fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady zoom while the fragments drift and tumble into the black void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007638176903128624 + ], + "opposite_logprob_tokens": [ + -4.882637977600098 + ], + "logprob_reward": -0.007638176903128624, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007638176903128624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10271807760000229, + 0.047623567283153534, + 0.10403743386268616, + 0.09447045624256134, + 0.05476784333586693, + 0.09201353788375854, + 0.023666588589549065, + 0.08812525868415833, + 0.11943241208791733, + 0.2731448709964752 + ], + "credit_map": [ + 0.010545764118432999, + 0.012168414890766144, + 0.008536377921700478, + 0.02140871435403824, + 0.018799830228090286, + 0.008277514949440956, + 0.016835708171129227, + 0.006145750172436237, + 0.005405737087130547, + 0.006435635965317488, + 0.0028930988628417253, + 0.005109673365950584, + 0.011559358797967434, + 0.004302606917917728, + 0.00808761827647686, + 0.0038298359140753746, + 0.009662345983088017, + 0.01991433836519718, + 0.008360976353287697, + 0.012973376549780369, + 0.022450732067227364, + 0.007350913714617491, + 0.014354786835610867, + 0.00896996445953846, + 0.007435196079313755, + 0.028624648228287697, + 0.006789406295865774, + 0.008292405866086483, + 0.02736426703631878, + 0.003700535511597991, + 0.0074564507231116295, + 0.004807556048035622, + 0.004825490061193705, + 0.007628899998962879, + 0.003864184021949768, + 0.011602049693465233, + 0.007214099168777466, + 0.003281238256022334, + 0.01131492666900158, + 0.005036956164985895, + 0.00875263474881649, + 0.014134855940937996, + 0.005339475814253092, + 0.02046998217701912, + 0.012231596745550632, + 0.007153827231377363, + 0.01654919609427452, + 0.0073819588869810104, + 0.0026205384638160467, + 0.003842609701678157, + 0.001620869035832584, + 0.0036599754821509123, + 0.003670299192890525, + 0.0017338877078145742, + 0.0039337435737252235, + 0.0025846639182418585, + 0.01111600361764431, + 0.014081168919801712, + 0.006478790193796158, + 0.0101295942440629, + 0.016274597495794296, + 0.006652118172496557, + 0.012945456430315971, + 0.010447523556649685, + 0.01675357297062874, + 0.02052769809961319, + 0.009363371878862381, + 0.014920465648174286, + 0.01355920359492302, + 0.014988421462476254, + 0.013448815792798996, + 0.015870871022343636, + 0.023517055436968803, + 0.03792325779795647, + 0.020823052152991295, + 0.034574251621961594, + 0.03672539070248604, + 0.03490781411528587, + 0.048685699701309204, + 0.035988353192806244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.007638176903128624 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vintage monochrome ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ticket display visible wear including frayed edges, softened corners, and a prominent center fold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine, spider-web-like cracks appear on the ticket's surface and does the ticket fracture into jagged fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady zoom while the fragments drift and tumble into the black void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205691024661064 + ], + "opposite_logprob_tokens": [ + -2.8120570182800293 + ], + "logprob_reward": -0.06205691024661064, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205691024661064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11168433725833893, + 0.0649220421910286, + 0.11734098941087723, + 0.1244882345199585, + 0.10719933360815048, + 0.04934447258710861, + 0.04578712582588196, + 0.07022253423929214, + 0.12384002655744553, + 0.18517082929611206 + ], + "credit_map": [ + 0.01263989694416523, + 0.011995779350399971, + 0.008120009675621986, + 0.01782122626900673, + 0.013837045058608055, + 0.017956791445612907, + 0.017113424837589264, + 0.012200179509818554, + 0.009271812625229359, + 0.00894959270954132, + 0.005770922172814608, + 0.009452925994992256, + 0.008222242817282677, + 0.006616456434130669, + 0.008689282462000847, + 0.00794881209731102, + 0.010535524226725101, + 0.02016363851726055, + 0.01426986139267683, + 0.014528235420584679, + 0.02640356495976448, + 0.0072333840653300285, + 0.015727758407592773, + 0.008479036390781403, + 0.007771609351038933, + 0.03436890244483948, + 0.011152273043990135, + 0.01615668460726738, + 0.024780400097370148, + 0.007071253377944231, + 0.014626272022724152, + 0.0085608484223485, + 0.011532812379300594, + 0.013656185939908028, + 0.008308257907629013, + 0.018275493755936623, + 0.015807004645466805, + 0.011384112760424614, + 0.016756290569901466, + 0.011479191482067108, + 0.007000347133725882, + 0.008728216402232647, + 0.003066392382606864, + 0.008964543230831623, + 0.0060113584622740746, + 0.002885024528950453, + 0.008768620900809765, + 0.003919973503798246, + 0.0049904282204806805, + 0.007075165864080191, + 0.004052178002893925, + 0.007976310327649117, + 0.006500803399831057, + 0.0026914163026958704, + 0.008397174067795277, + 0.004103659652173519, + 0.006615092512220144, + 0.012139609083533287, + 0.005099012050777674, + 0.009826019406318665, + 0.012547464109957218, + 0.007079873699694872, + 0.009904569946229458, + 0.007010899484157562, + 0.010341423563659191, + 0.02140744775533676, + 0.00980105809867382, + 0.015643756836652756, + 0.01649155654013157, + 0.01635124161839485, + 0.015219444409012794, + 0.018584102392196655, + 0.02414417639374733, + 0.026662105694413185, + 0.013631157577037811, + 0.021345866844058037, + 0.02482544630765915, + 0.023528488352894783, + 0.029524268582463264, + 0.02150934934616089 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.06205691024661064 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, static, extreme close-up shot from a slightly elevated angle initially depicts a vintage monochrome ticket, its details rendered with the fidelity of a vintage photograph. Set against a stark, non-reflective black background, the ticket's thick, tan paper displays authentic wear: frayed edges, softened corners, and a prominent center fold. The faded, nearly illegible charcoal-gray font and the faint outline of the serial number are visible beneath a soft light source that sweeps across the surface, carving subtle shadows into the creases and highlighting the fiber texture. Suddenly, the stillness of the shot breaks, revealing a web of fine, spider-web-like cracks creeping across the ticket's face. Then, the camera begins a slow, steady zoom as the ticket inexorably fractures into countless jagged fragments. Rather than falling, these fragments delicately drift, separating and drifting into the surrounding black void, as if carried by a gentle, invisible current. Each piece gracefully tumbles along its own path, creating a poignant and deeply nostalgic atmosphere, the slow disintegration of a cherished memory.", + "seed": 13995, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_8/index_1462_seed_13995_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03163137808442116, + "vlm_logprob": -0.03163137808442116, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.03163137808442116, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vintage monochrome ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ticket display visible wear including frayed edges, softened corners, and a prominent center fold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine, spider-web-like cracks appear on the ticket's surface and does the ticket fracture into jagged fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady zoom while the fragments drift and tumble into the black void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027170466259121895 + ], + "opposite_logprob_tokens": [ + -4.152170658111572 + ], + "logprob_reward": -0.027170466259121895, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.027170466259121895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09864720702171326, + 0.43345919251441956, + 0.030755162239074707, + 0.03539117053151131, + 0.1289059817790985, + 0.08783803135156631, + 0.07171620428562164, + 0.032095834612846375, + 0.03981945291161537, + 0.04137176647782326 + ], + "credit_map": [ + 0.013950706459581852, + 0.018157431855797768, + 0.009167395532131195, + 0.015155068598687649, + 0.009579118341207504, + 0.006451068911701441, + 0.015510596334934235, + 0.01067582331597805, + 0.05362725257873535, + 0.09142790734767914, + 0.03328759968280792, + 0.05686125531792641, + 0.03374364227056503, + 0.03236683830618858, + 0.06137973070144653, + 0.0707649439573288, + 0.004171430598944426, + 0.005632000043988228, + 0.003055831417441368, + 0.004275568760931492, + 0.002699953503906727, + 0.0020133282523602247, + 0.005308603867888451, + 0.003598446724936366, + 0.004549908917397261, + 0.00588618777692318, + 0.003510013921186328, + 0.004668783862143755, + 0.003228459507226944, + 0.0032684418838471174, + 0.005476746708154678, + 0.00480262516066432, + 0.010356070473790169, + 0.014009462669491768, + 0.00620430801063776, + 0.011436226777732372, + 0.010722269304096699, + 0.022384334355592728, + 0.04305638000369072, + 0.010736932046711445, + 0.009722397662699223, + 0.013683265075087547, + 0.006638529244810343, + 0.007313874084502459, + 0.006390744354575872, + 0.006212909705936909, + 0.025952348485589027, + 0.011923969723284245, + 0.007925937883555889, + 0.01036467682570219, + 0.0067296102643013, + 0.0062992931343615055, + 0.005989967379719019, + 0.005493032280355692, + 0.009960867464542389, + 0.018952812999486923, + 0.0034241031389683485, + 0.002812711289152503, + 0.002381822094321251, + 0.003894239664077759, + 0.002532175974920392, + 0.0037456853315234184, + 0.004882904700934887, + 0.008422191254794598, + 0.007224563043564558, + 0.005441244691610336, + 0.0038132253102958202, + 0.0037990182172507048, + 0.003703065449371934, + 0.004791940096765757, + 0.0053032999858260155, + 0.005743098445236683, + 0.005428621545433998, + 0.004222446586936712, + 0.0031863064505159855, + 0.004213212989270687, + 0.0035822419449687004, + 0.009131278842687607, + 0.005458890460431576, + 0.006148769520223141 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.027170466259121895 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vintage monochrome ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ticket display visible wear including frayed edges, softened corners, and a prominent center fold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine, spider-web-like cracks appear on the ticket's surface and does the ticket fracture into jagged fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady zoom while the fragments drift and tumble into the black void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900816947221756 + ], + "opposite_logprob_tokens": [ + -2.579008102416992 + ], + "logprob_reward": -0.07900816947221756, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900816947221756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39093559980392456, + 0.30658644437789917, + 0.06284473836421967, + 0.024128127843141556, + 0.06663435697555542, + 0.017309973016381264, + 0.005504948552697897, + 0.012071298435330391, + 0.03919044882059097, + 0.07479403913021088 + ], + "credit_map": [ + 0.051680732518434525, + 0.06821811199188232, + 0.03233296051621437, + 0.05279455706477165, + 0.03039587289094925, + 0.0315987765789032, + 0.05893426015973091, + 0.06498035043478012, + 0.03830352798104286, + 0.06538738310337067, + 0.023630039766430855, + 0.04063522443175316, + 0.023966437205672264, + 0.022464526817202568, + 0.04285087808966637, + 0.04934844747185707, + 0.008478141389787197, + 0.010598085820674896, + 0.005904824007302523, + 0.007472676690667868, + 0.005832558963447809, + 0.0058683669194579124, + 0.011061332188546658, + 0.007628759369254112, + 0.003180781612172723, + 0.0032973415218293667, + 0.0028653317131102085, + 0.002698881784453988, + 0.0024983338080346584, + 0.0026782737113535404, + 0.003823287785053253, + 0.003085898235440254, + 0.005292909219861031, + 0.00831175222992897, + 0.004094960168004036, + 0.0053168535232543945, + 0.006088437978178263, + 0.011293336749076843, + 0.01939229480922222, + 0.0068438174203038216, + 0.002180190524086356, + 0.0018971518147736788, + 0.001487996894866228, + 0.00151621806435287, + 0.0016177860088646412, + 0.00289978738874197, + 0.0030788977164775133, + 0.002631946699693799, + 0.0006876512197777629, + 0.0005849523586221039, + 0.0006093527190387249, + 0.0005830131121911108, + 0.0006443666643463075, + 0.000529035460203886, + 0.0007328861975111067, + 0.0011336914030835032, + 0.001562356948852539, + 0.0006623808294534683, + 0.0006021705339662731, + 0.0006976713193580508, + 0.0006050007650628686, + 0.0016128929564729333, + 0.0015536127611994743, + 0.004775213543325663, + 0.009494162164628506, + 0.003660526592284441, + 0.0029254627879709005, + 0.0020891421008855104, + 0.003619489027187228, + 0.0028604252729564905, + 0.0030880174599587917, + 0.01145322248339653, + 0.024671899154782295, + 0.006015429738909006, + 0.003427303396165371, + 0.0017979285912588239, + 0.005888462066650391, + 0.02434687130153179, + 0.004326195456087589, + 0.004319954197853804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.07900816947221756 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vintage monochrome ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ticket display visible wear including frayed edges, softened corners, and a prominent center fold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine, spider-web-like cracks appear on the ticket's surface and does the ticket fracture into jagged fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady zoom while the fragments drift and tumble into the black void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011090775020420551 + ], + "opposite_logprob_tokens": [ + -4.5110907554626465 + ], + "logprob_reward": -0.011090775020420551, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011090775020420551, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1940561980009079, + 0.16897229850292206, + 0.040202558040618896, + 0.021055415272712708, + 0.08424482494592667, + 0.030928470194339752, + 0.037147797644138336, + 0.05263484641909599, + 0.029404114931821823, + 0.3413534462451935 + ], + "credit_map": [ + 0.024957867339253426, + 0.0374966561794281, + 0.016183149069547653, + 0.025382405146956444, + 0.015576028265058994, + 0.015722408890724182, + 0.026490593329072, + 0.03224710375070572, + 0.019861748442053795, + 0.03730195760726929, + 0.012962684966623783, + 0.022233303636312485, + 0.013096077367663383, + 0.012402055785059929, + 0.02410902827978134, + 0.027005450800061226, + 0.005662642884999514, + 0.006309656426310539, + 0.0038600130937993526, + 0.005423889961093664, + 0.0039378684014081955, + 0.0035956886131316423, + 0.0058936732821166515, + 0.005519131198525429, + 0.0022535882890224457, + 0.0031516116578131914, + 0.002571799559518695, + 0.001982403453439474, + 0.0018425036687403917, + 0.00226828595623374, + 0.00446972856298089, + 0.0025154943577945232, + 0.007067590486258268, + 0.010015692561864853, + 0.005531050264835358, + 0.006933466996997595, + 0.008013811893761158, + 0.014139982871711254, + 0.02342628687620163, + 0.009116943925619125, + 0.0036712426226586103, + 0.004942163359373808, + 0.0027827033773064613, + 0.0027810160536319017, + 0.003535209223628044, + 0.004248298704624176, + 0.004985387437045574, + 0.003982447553426027, + 0.0036616302095353603, + 0.005929578095674515, + 0.00432007247582078, + 0.00415042182430625, + 0.0032366716768592596, + 0.003495582612231374, + 0.006117779761552811, + 0.006236060988157988, + 0.005345879588276148, + 0.002957930089905858, + 0.0026199566200375557, + 0.003537587821483612, + 0.0029554804787039757, + 0.009140264242887497, + 0.006494887173175812, + 0.019582858309149742, + 0.003948138561099768, + 0.003453282406553626, + 0.0025299936532974243, + 0.0030809345189481974, + 0.004257603082805872, + 0.0034020377788692713, + 0.0038266589399427176, + 0.004905464127659798, + 0.10504675656557083, + 0.03022579848766327, + 0.016080882400274277, + 0.012966995127499104, + 0.026505131274461746, + 0.10875629633665085, + 0.022189542651176453, + 0.019582059234380722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.011090775020420551 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vintage monochrome ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ticket display visible wear including frayed edges, softened corners, and a prominent center fold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine, spider-web-like cracks appear on the ticket's surface and does the ticket fracture into jagged fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady zoom while the fragments drift and tumble into the black void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011084645055234432 + ], + "opposite_logprob_tokens": [ + -4.51108455657959 + ], + "logprob_reward": -0.011084645055234432, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011084645055234432, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09575040638446808, + 0.021983522921800613, + 0.0684041827917099, + 0.045462701469659805, + 0.08303298056125641, + 0.04957510158419609, + 0.05247931182384491, + 0.08886631578207016, + 0.061331864446401596, + 0.43311363458633423 + ], + "credit_map": [ + 0.011300990357995033, + 0.01790539175271988, + 0.007878364995121956, + 0.01290628407150507, + 0.007103513926267624, + 0.007493435405194759, + 0.014982831664383411, + 0.016179589554667473, + 0.0030032971408218145, + 0.004220100119709969, + 0.0021483099553734064, + 0.003024045843631029, + 0.0017750102560967207, + 0.001885654404759407, + 0.003078462788835168, + 0.0028486414812505245, + 0.009404872544109821, + 0.01206173375248909, + 0.00587468734011054, + 0.0090278759598732, + 0.006283466704189777, + 0.005321166943758726, + 0.010518493130803108, + 0.009911879897117615, + 0.00671039754524827, + 0.006327021401375532, + 0.004994601011276245, + 0.005673330742865801, + 0.004209086764603853, + 0.0042488244362175465, + 0.0071371388621628284, + 0.006162299774587154, + 0.007727616932243109, + 0.010568822734057903, + 0.006127787288278341, + 0.0072256517596542835, + 0.007913823239505291, + 0.01323702186346054, + 0.020791595801711082, + 0.009440657682716846, + 0.004919091239571571, + 0.007263593375682831, + 0.004147022031247616, + 0.004379778169095516, + 0.005428342614322901, + 0.00883423537015915, + 0.007780723739415407, + 0.006822317373007536, + 0.008009426295757294, + 0.00971875712275505, + 0.0046324348077178, + 0.005038103554397821, + 0.004475321155041456, + 0.00444610882550478, + 0.006081108935177326, + 0.010078050196170807, + 0.008096936158835888, + 0.005257922690361738, + 0.004523457959294319, + 0.006098750978708267, + 0.005516095086932182, + 0.016808513551950455, + 0.01103341393172741, + 0.03153122588992119, + 0.011330111883580685, + 0.006371431518346071, + 0.006174975540488958, + 0.005160416942089796, + 0.0077201807871460915, + 0.006537525448948145, + 0.006683679297566414, + 0.011353542096912861, + 0.12262044847011566, + 0.038994669914245605, + 0.02164401486515999, + 0.021639840677380562, + 0.0342886745929718, + 0.1311338245868683, + 0.031126849353313446, + 0.03166535124182701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.011084645055234432 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vintage monochrome ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ticket display visible wear including frayed edges, softened corners, and a prominent center fold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine, spider-web-like cracks appear on the ticket's surface and does the ticket fracture into jagged fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady zoom while the fragments drift and tumble into the black void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980283461511135 + ], + "opposite_logprob_tokens": [ + -3.5298027992248535 + ], + "logprob_reward": -0.02980283461511135, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980283461511135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03218143805861473, + 0.15048889815807343, + 0.07234486937522888, + 0.04928691312670708, + 0.13728517293930054, + 0.1779954880475998, + 0.038582030683755875, + 0.0710049644112587, + 0.07991743832826614, + 0.19091273844242096 + ], + "credit_map": [ + 0.005041162017732859, + 0.004930543247610331, + 0.0033517442643642426, + 0.004354110918939114, + 0.0027480574790388346, + 0.0026348468381911516, + 0.004993284586817026, + 0.004127688240259886, + 0.01896894909441471, + 0.031728994101285934, + 0.011715945787727833, + 0.019930075854063034, + 0.011752303689718246, + 0.01126063335686922, + 0.0210456270724535, + 0.02408636175096035, + 0.010253483429551125, + 0.012121549807488918, + 0.006970684044063091, + 0.008654915727674961, + 0.006801478564739227, + 0.006115863565355539, + 0.01151993591338396, + 0.009906956925988197, + 0.006532028317451477, + 0.007545527536422014, + 0.005707235541194677, + 0.0056368447840213776, + 0.005137329455465078, + 0.004345472436398268, + 0.007903283461928368, + 0.006479186471551657, + 0.01202398631721735, + 0.018268220126628876, + 0.010089077055454254, + 0.011573521420359612, + 0.012526996433734894, + 0.01999487541615963, + 0.03728272020816803, + 0.01552576944231987, + 0.024009399116039276, + 0.021696072071790695, + 0.01540261972695589, + 0.014746326953172684, + 0.01773505099117756, + 0.023931602016091347, + 0.03345242515206337, + 0.02702198177576065, + 0.0033638221211731434, + 0.0047372509725391865, + 0.0036004476714879274, + 0.004762087017297745, + 0.0031862142495810986, + 0.0060797217302024364, + 0.006251671817153692, + 0.00660081347450614, + 0.005905637983232737, + 0.006465803366154432, + 0.006226839032024145, + 0.01138669066131115, + 0.007315217051655054, + 0.0074910447001457214, + 0.010000389069318771, + 0.016213340684771538, + 0.01737091690301895, + 0.006461167708039284, + 0.00484024453908205, + 0.0047421022318303585, + 0.006534229498356581, + 0.0065161376260221004, + 0.007772472687065601, + 0.025680169463157654, + 0.053519245237112045, + 0.014808286912739277, + 0.010823625139892101, + 0.012543834745883942, + 0.015589455142617226, + 0.05231250450015068, + 0.012354789301753044, + 0.01896098256111145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 1 + ], + "selected_reward": -0.02980283461511135 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, static, extreme close-up shot from a slightly elevated angle initially depicts a vintage monochrome ticket, its details rendered with the fidelity of a vintage photograph. Set against a stark, non-reflective black background, the ticket's thick, tan paper displays authentic wear: frayed edges, softened corners, and a prominent center fold. The faded, nearly illegible charcoal-gray font and the faint outline of the serial number are visible beneath a soft light source that sweeps across the surface, carving subtle shadows into the creases and highlighting the fiber texture. Suddenly, the stillness of the shot breaks, revealing a web of fine, spider-web-like cracks creeping across the ticket's face. Then, the camera begins a slow, steady zoom as the ticket inexorably fractures into countless jagged fragments. Rather than falling, these fragments delicately drift, separating and drifting into the surrounding black void, as if carried by a gentle, invisible current. Each piece gracefully tumbles along its own path, creating a poignant and deeply nostalgic atmosphere, the slow disintegration of a cherished memory.", + "seed": 13996, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_8/index_1462_seed_13996_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.048309298045933244, + "vlm_logprob": -0.048309298045933244, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.048309298045933244, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vintage monochrome ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ticket display visible wear including frayed edges, softened corners, and a prominent center fold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine, spider-web-like cracks appear on the ticket's surface and does the ticket fracture into jagged fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady zoom while the fragments drift and tumble into the black void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1061653345823288 + ], + "opposite_logprob_tokens": [ + -2.3561654090881348 + ], + "logprob_reward": -0.1061653345823288, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1061653345823288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025393350049853325, + 0.08534104377031326, + 0.23584479093551636, + 0.3366757333278656, + 0.10559682548046112, + 0.04058779031038284, + 0.011377773247659206, + 0.028316680341959, + 0.11653362959623337, + 0.01433238573372364 + ], + "credit_map": [ + 0.00291497097350657, + 0.0036353503819555044, + 0.003344789147377014, + 0.0038717479910701513, + 0.0038116909563541412, + 0.0019743498414754868, + 0.0033250560518354177, + 0.002515395637601614, + 0.018763121217489243, + 0.005021469667553902, + 0.008044670335948467, + 0.0047491854056715965, + 0.005769456271082163, + 0.007084911223500967, + 0.0066128941252827644, + 0.029295334592461586, + 0.010733338072896004, + 0.02199667878448963, + 0.013957202434539795, + 0.08933179825544357, + 0.019256429746747017, + 0.010212631896138191, + 0.017807211726903915, + 0.05254951864480972, + 0.02349342778325081, + 0.11121310293674469, + 0.019193200394511223, + 0.014527034945786, + 0.09266971051692963, + 0.018878823146224022, + 0.014681401662528515, + 0.04201904311776161, + 0.012291764840483665, + 0.014482516795396805, + 0.020317919552326202, + 0.011440173722803593, + 0.0147511912509799, + 0.012107866816222668, + 0.008666479028761387, + 0.011538906022906303, + 0.008211511187255383, + 0.006825875025242567, + 0.004225559998303652, + 0.002941218437626958, + 0.006529039703309536, + 0.0037187309935688972, + 0.003526917891576886, + 0.004608938470482826, + 0.001049238839186728, + 0.0021786813158541918, + 0.0011674893321469426, + 0.000959861499723047, + 0.002438280498608947, + 0.001184936030767858, + 0.0012399574043229222, + 0.001159328268840909, + 0.002257678424939513, + 0.006291951518505812, + 0.0032950020395219326, + 0.0022578092757612467, + 0.007056008093059063, + 0.0026363616343587637, + 0.002214757725596428, + 0.0023071132600307465, + 0.013973833993077278, + 0.021158333867788315, + 0.013111315667629242, + 0.0070863342843949795, + 0.030875831842422485, + 0.012450923211872578, + 0.007864336483180523, + 0.010012723505496979, + 0.0017837958876043558, + 0.0021776037756353617, + 0.0013805823400616646, + 0.0018551775719970465, + 0.0020894668996334076, + 0.0013723380398005247, + 0.0018632152350619435, + 0.0018102056346833706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 8 + ], + "selected_reward": -0.1061653345823288 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vintage monochrome ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ticket display visible wear including frayed edges, softened corners, and a prominent center fold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine, spider-web-like cracks appear on the ticket's surface and does the ticket fracture into jagged fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady zoom while the fragments drift and tumble into the black void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10035204887390137 + ], + "opposite_logprob_tokens": [ + -2.3503520488739014 + ], + "logprob_reward": -0.10035204887390137, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10035204887390137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04440360143780708, + 0.062270376831293106, + 0.05222850665450096, + 0.2660602629184723, + 0.09040515869855881, + 0.03452521190047264, + 0.11311833560466766, + 0.12767581641674042, + 0.06784234195947647, + 0.14147043228149414 + ], + "credit_map": [ + 0.005100051872432232, + 0.00562877906486392, + 0.005303056910634041, + 0.0049984510987997055, + 0.007215667981654406, + 0.003206202294677496, + 0.007054140791296959, + 0.005897254217416048, + 0.012437483295798302, + 0.005835803225636482, + 0.005248222965747118, + 0.004231835249811411, + 0.006490013096481562, + 0.004945471882820129, + 0.0050789122469723225, + 0.01800263673067093, + 0.005204863380640745, + 0.004180495627224445, + 0.0062712980434298515, + 0.013318712823092937, + 0.004497934132814407, + 0.005052655469626188, + 0.006162215489894152, + 0.007540328893810511, + 0.018086889758706093, + 0.08814085274934769, + 0.017326274886727333, + 0.012179199606180191, + 0.0736560970544815, + 0.01538585964590311, + 0.010487311519682407, + 0.0307978093624115, + 0.01031574234366417, + 0.013657544739544392, + 0.009766272269189358, + 0.006704077590256929, + 0.016961682587862015, + 0.01215916033834219, + 0.007135528605431318, + 0.013705149292945862, + 0.005613403394818306, + 0.004903942346572876, + 0.004859167616814375, + 0.0038537937216460705, + 0.004488162696361542, + 0.003683011746034026, + 0.003160925582051277, + 0.003962801769375801, + 0.009924832731485367, + 0.03418638929724693, + 0.008805394172668457, + 0.006975341122597456, + 0.030751479789614677, + 0.007459561340510845, + 0.00582291092723608, + 0.009192429482936859, + 0.007107860408723354, + 0.0432463176548481, + 0.006579224951565266, + 0.005861850455403328, + 0.04729297384619713, + 0.0058441986329853535, + 0.005446454044431448, + 0.0062969280406832695, + 0.006073134485632181, + 0.011322489008307457, + 0.008038809522986412, + 0.006532716564834118, + 0.011996809393167496, + 0.008521671406924725, + 0.008439457975327969, + 0.006917258724570274, + 0.019032174721360207, + 0.02049385942518711, + 0.014059479348361492, + 0.01803639717400074, + 0.02080029621720314, + 0.014030560851097107, + 0.02025533840060234, + 0.014762327075004578 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 7 + ], + "selected_reward": -0.10035204887390137 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vintage monochrome ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ticket display visible wear including frayed edges, softened corners, and a prominent center fold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine, spider-web-like cracks appear on the ticket's surface and does the ticket fracture into jagged fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady zoom while the fragments drift and tumble into the black void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009797343984246254 + ], + "opposite_logprob_tokens": [ + -4.6347975730896 + ], + "logprob_reward": -0.009797343984246254, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009797343984246254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025981716811656952, + 0.055228833109140396, + 0.17931996285915375, + 0.33431753516197205, + 0.09365016967058182, + 0.01814117096364498, + 0.01713576726615429, + 0.028370214626193047, + 0.1884668469429016, + 0.05938781052827835 + ], + "credit_map": [ + 0.0030275778844952583, + 0.0033362586982548237, + 0.002840085420757532, + 0.004110848996788263, + 0.0033165202476084232, + 0.0023917723447084427, + 0.004384409636259079, + 0.002574242651462555, + 0.010945207439363003, + 0.0036496520042419434, + 0.004956195596605539, + 0.0037296570371836424, + 0.004249824211001396, + 0.004581736866384745, + 0.005751667078584433, + 0.01736489124596119, + 0.007798306178301573, + 0.016617588698863983, + 0.010431451722979546, + 0.06847862899303436, + 0.014741919003427029, + 0.00755379069596529, + 0.01326671987771988, + 0.04043154790997505, + 0.023005351424217224, + 0.11099880933761597, + 0.01963822916150093, + 0.014567367732524872, + 0.09333068877458572, + 0.01917087659239769, + 0.013494424521923065, + 0.040111757814884186, + 0.011186587624251842, + 0.012225688435137272, + 0.015915926545858383, + 0.009893031790852547, + 0.014587553218007088, + 0.011255682446062565, + 0.007906968705356121, + 0.010678725317120552, + 0.0016628392040729523, + 0.002494428539648652, + 0.0016487203538417816, + 0.0016884899232536554, + 0.003425188362598419, + 0.003608727129176259, + 0.0020707251969724894, + 0.0015420509735122323, + 0.0016156117198988795, + 0.003828238695859909, + 0.0013674955116584897, + 0.0015267423586919904, + 0.0039730616845190525, + 0.001794644515030086, + 0.0011858365032821894, + 0.0018441352294757962, + 0.0016311267390847206, + 0.0074269138276577, + 0.00281285447999835, + 0.0018122834153473377, + 0.008420104160904884, + 0.0030803359113633633, + 0.001459797378629446, + 0.0017267974326387048, + 0.023117240518331528, + 0.03551864996552467, + 0.020493794232606888, + 0.010736234486103058, + 0.04700319841504097, + 0.021631522104144096, + 0.013394795358181, + 0.01657138764858246, + 0.011694117449223995, + 0.010625133290886879, + 0.00432464899495244, + 0.005697320215404034, + 0.014966940507292747, + 0.004574732854962349, + 0.003124152310192585, + 0.004380763042718172 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 2 + ], + "selected_reward": -0.009797343984246254 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vintage monochrome ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ticket display visible wear including frayed edges, softened corners, and a prominent center fold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine, spider-web-like cracks appear on the ticket's surface and does the ticket fracture into jagged fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady zoom while the fragments drift and tumble into the black void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046440837904810905 + ], + "opposite_logprob_tokens": [ + -5.37964391708374 + ], + "logprob_reward": -0.0046440837904810905, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046440837904810905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.022338315844535828, + 0.033525340259075165, + 0.1268693506717682, + 0.2851353585720062, + 0.08583433926105499, + 0.027091579511761665, + 0.023698031902313232, + 0.0530046708881855, + 0.2360060214996338, + 0.10649700462818146 + ], + "credit_map": [ + 0.002552130725234747, + 0.003153827739879489, + 0.0029466329142451286, + 0.003606732003390789, + 0.003174673765897751, + 0.0017311995616182685, + 0.003385059302672744, + 0.0017880592495203018, + 0.0064809927716851234, + 0.002622086089104414, + 0.0033584092743694782, + 0.002339877188205719, + 0.0024518563877791166, + 0.0031843879260122776, + 0.002654067240655422, + 0.010433662682771683, + 0.0053492398001253605, + 0.01188010722398758, + 0.007168699987232685, + 0.04881206527352333, + 0.010324288159608841, + 0.005192742682993412, + 0.009317387826740742, + 0.02882480062544346, + 0.020495973527431488, + 0.09328781068325043, + 0.01776048168540001, + 0.012654651887714863, + 0.07722831517457962, + 0.016795897856354713, + 0.011295104399323463, + 0.035617124289274216, + 0.010744315572082996, + 0.011219321750104427, + 0.015493040904402733, + 0.00892459787428379, + 0.012393631041049957, + 0.010593947023153305, + 0.006874480284750462, + 0.009591002948582172, + 0.004887938033789396, + 0.0038624536246061325, + 0.0037549585103988647, + 0.0027764905244112015, + 0.0034603970125317574, + 0.0025132140144705772, + 0.0021762705873697996, + 0.0036598569713532925, + 0.0015053109964355826, + 0.004763714037835598, + 0.002113234717398882, + 0.0017707252409309149, + 0.006429016590118408, + 0.0024193348363041878, + 0.002110853558406234, + 0.002585840877145529, + 0.002345328452065587, + 0.01828964799642563, + 0.0022801535669714212, + 0.0029781300108879805, + 0.020250903442502022, + 0.003039721865206957, + 0.001822406891733408, + 0.0019983812235295773, + 0.029859036207199097, + 0.04460969567298889, + 0.02473868802189827, + 0.013118784874677658, + 0.05957157909870148, + 0.026795847341418266, + 0.016698498278856277, + 0.020613903179764748, + 0.020083174109458923, + 0.018905149772763252, + 0.006885307375341654, + 0.010065045207738876, + 0.02605508640408516, + 0.008867369964718819, + 0.008981534279882908, + 0.006654338911175728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 2 + ], + "selected_reward": -0.0046440837904810905 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vintage monochrome ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ticket display visible wear including frayed edges, softened corners, and a prominent center fold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine, spider-web-like cracks appear on the ticket's surface and does the ticket fracture into jagged fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady zoom while the fragments drift and tumble into the black void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020587678998708725 + ], + "opposite_logprob_tokens": [ + -3.895587682723999 + ], + "logprob_reward": -0.020587678998708725, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020587678998708725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05189624801278114, + 0.040952783077955246, + 0.25887390971183777, + 0.3345942497253418, + 0.07076843082904816, + 0.029123779386281967, + 0.05476126819849014, + 0.0530979223549366, + 0.03091089241206646, + 0.07502049952745438 + ], + "credit_map": [ + 0.006790000945329666, + 0.006211237981915474, + 0.007104557007551193, + 0.007188309449702501, + 0.00772274611517787, + 0.004271557554602623, + 0.006832887418568134, + 0.005774949677288532, + 0.0076957265846431255, + 0.003271248424425721, + 0.004016739781945944, + 0.0024452731013298035, + 0.0033144273329526186, + 0.004760625772178173, + 0.0034817506093531847, + 0.011966992169618607, + 0.011356817558407784, + 0.023479893803596497, + 0.014468769542872906, + 0.10029388219118118, + 0.01981339417397976, + 0.0107731306925416, + 0.019247397780418396, + 0.059440601617097855, + 0.023456033319234848, + 0.11142858862876892, + 0.02055012248456478, + 0.01415977906435728, + 0.09132373332977295, + 0.017553359270095825, + 0.013045886531472206, + 0.04307674244046211, + 0.007591534871608019, + 0.009711487218737602, + 0.008292770013213158, + 0.006551194004714489, + 0.013437903486192226, + 0.009715832769870758, + 0.0062346868216991425, + 0.009233017452061176, + 0.0035764609929174185, + 0.0038244121242314577, + 0.003264286322519183, + 0.0030107435304671526, + 0.005806315690279007, + 0.003505311906337738, + 0.0037147984839975834, + 0.0024214524310082197, + 0.0025613135658204556, + 0.017590723931789398, + 0.003968776669353247, + 0.0023934459313750267, + 0.018355369567871094, + 0.0032736104913055897, + 0.0030644219368696213, + 0.0035536091309040785, + 0.0027047544717788696, + 0.01558070257306099, + 0.004035996738821268, + 0.0039862701669335365, + 0.01652972400188446, + 0.0034303031861782074, + 0.003960344009101391, + 0.00286982674151659, + 0.003921404015272856, + 0.0056082820519804955, + 0.003414342412725091, + 0.003011597553268075, + 0.006405308842658997, + 0.003197154263034463, + 0.002825097180902958, + 0.0025277084205299616, + 0.008110057562589645, + 0.010553832165896893, + 0.008782296441495419, + 0.008788186125457287, + 0.014226649887859821, + 0.006126272026449442, + 0.010673836804926395, + 0.007759375497698784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 9 + ], + "selected_reward": -0.020587678998708725 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, static, extreme close-up shot from a slightly elevated angle initially depicts a vintage monochrome ticket, its details rendered with the fidelity of a vintage photograph. Set against a stark, non-reflective black background, the ticket's thick, tan paper displays authentic wear: frayed edges, softened corners, and a prominent center fold. The faded, nearly illegible charcoal-gray font and the faint outline of the serial number are visible beneath a soft light source that sweeps across the surface, carving subtle shadows into the creases and highlighting the fiber texture. Suddenly, the stillness of the shot breaks, revealing a web of fine, spider-web-like cracks creeping across the ticket's face. Then, the camera begins a slow, steady zoom as the ticket inexorably fractures into countless jagged fragments. Rather than falling, these fragments delicately drift, separating and drifting into the surrounding black void, as if carried by a gentle, invisible current. Each piece gracefully tumbles along its own path, creating a poignant and deeply nostalgic atmosphere, the slow disintegration of a cherished memory.", + "seed": 13997, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_8/index_1462_seed_13997_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.029084039852023125, + "vlm_logprob": -0.029084039852023125, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.029084039852023125, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vintage monochrome ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ticket display visible wear including frayed edges, softened corners, and a prominent center fold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine, spider-web-like cracks appear on the ticket's surface and does the ticket fracture into jagged fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady zoom while the fragments drift and tumble into the black void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0319492444396019 + ], + "opposite_logprob_tokens": [ + -3.781949281692505 + ], + "logprob_reward": -0.0319492444396019, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0319492444396019, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12017909437417984, + 0.02710968442261219, + 0.18492797017097473, + 0.10717462003231049, + 0.08049394190311432, + 0.12465830147266388, + 0.13395583629608154, + 0.06875686347484589, + 0.0757574513554573, + 0.07698626071214676 + ], + "credit_map": [ + 0.014916970394551754, + 0.022189941257238388, + 0.019071782007813454, + 0.015891943126916885, + 0.019897233694791794, + 0.0077748047187924385, + 0.011403487995266914, + 0.009032939560711384, + 0.0029185183811932802, + 0.005459756124764681, + 0.002116062678396702, + 0.0029180250130593777, + 0.005802710074931383, + 0.0025782259181141853, + 0.002583630383014679, + 0.002732757944613695, + 0.009846213273704052, + 0.07151949405670166, + 0.017905663698911667, + 0.011745522730052471, + 0.051169514656066895, + 0.005462110973894596, + 0.011053857393562794, + 0.00622559105977416, + 0.009218631312251091, + 0.017057038843631744, + 0.006453660316765308, + 0.012946795672178268, + 0.016951628029346466, + 0.007778954226523638, + 0.025082100182771683, + 0.011685813777148724, + 0.008666492998600006, + 0.014704108238220215, + 0.005754127632826567, + 0.008118011057376862, + 0.015372222289443016, + 0.010268012061715126, + 0.008308442309498787, + 0.009302517399191856, + 0.012242360971868038, + 0.017416417598724365, + 0.010827711783349514, + 0.01640065759420395, + 0.01989736407995224, + 0.01104626152664423, + 0.022739555686712265, + 0.014087970368564129, + 0.014058568514883518, + 0.023324891924858093, + 0.016880834475159645, + 0.02091353014111519, + 0.02005116641521454, + 0.013005499728024006, + 0.015626780688762665, + 0.010094573721289635, + 0.008568552322685719, + 0.00906445924192667, + 0.006070015951991081, + 0.009095022454857826, + 0.011029728688299656, + 0.009665458463132381, + 0.007242691703140736, + 0.008020931854844093, + 0.0063271005637943745, + 0.009014477953314781, + 0.00812411867082119, + 0.010646353475749493, + 0.015943247824907303, + 0.008415212854743004, + 0.0074423314072191715, + 0.009844607673585415, + 0.005479126237332821, + 0.009778753854334354, + 0.004171623382717371, + 0.015643756836652756, + 0.010899130254983902, + 0.011995856650173664, + 0.009079347364604473, + 0.009938663803040981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 5 + ], + "selected_reward": -0.0319492444396019 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vintage monochrome ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ticket display visible wear including frayed edges, softened corners, and a prominent center fold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine, spider-web-like cracks appear on the ticket's surface and does the ticket fracture into jagged fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady zoom while the fragments drift and tumble into the black void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043092746287584305 + ], + "opposite_logprob_tokens": [ + -3.168092727661133 + ], + "logprob_reward": -0.043092746287584305, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043092746287584305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10846468061208725, + 0.02962150238454342, + 0.25140610337257385, + 0.06825242191553116, + 0.19017982482910156, + 0.17606684565544128, + 0.04835495352745056, + 0.0250017698854208, + 0.058853402733802795, + 0.043798502534627914 + ], + "credit_map": [ + 0.011855478398501873, + 0.020301884040236473, + 0.01369541697204113, + 0.014590744860470295, + 0.01989476941525936, + 0.009576049633324146, + 0.008499677293002605, + 0.010050659999251366, + 0.004498491995036602, + 0.003342436859384179, + 0.0031353309750556946, + 0.00455767847597599, + 0.004239254631102085, + 0.003360111266374588, + 0.0027526747435331345, + 0.003735525067895651, + 0.010048841126263142, + 0.10468153655529022, + 0.022540107369422913, + 0.015148588456213474, + 0.07834469527006149, + 0.006922087632119656, + 0.0067842332646250725, + 0.006936030928045511, + 0.0065194531343877316, + 0.012039531953632832, + 0.004389483015984297, + 0.008717871271073818, + 0.010511325672268867, + 0.006480813026428223, + 0.013704902492463589, + 0.005889042280614376, + 0.012788362801074982, + 0.047944575548172, + 0.013842492364346981, + 0.02490176633000374, + 0.04149806126952171, + 0.018800923600792885, + 0.015581740066409111, + 0.014821895398199558, + 0.007544610183686018, + 0.037306055426597595, + 0.0061221010982990265, + 0.02007671631872654, + 0.03923720866441727, + 0.01355909463018179, + 0.03575611859560013, + 0.016464948654174805, + 0.004658407531678677, + 0.007047249004244804, + 0.00885059218853712, + 0.004688185639679432, + 0.007550443056970835, + 0.004751031752675772, + 0.00610945513471961, + 0.0046995943412184715, + 0.003135658334940672, + 0.0032067704014480114, + 0.0022586537525057793, + 0.0026484064292162657, + 0.004176396876573563, + 0.002679029945284128, + 0.003203017171472311, + 0.003693836275488138, + 0.006408230401575565, + 0.005853189621120691, + 0.011053262278437614, + 0.007136208936572075, + 0.008075387217104435, + 0.005489709321409464, + 0.006583813112229109, + 0.00825359858572483, + 0.003018640447407961, + 0.008585774339735508, + 0.003188682720065117, + 0.005555631592869759, + 0.008311262354254723, + 0.00560290040448308, + 0.004181383177638054, + 0.005354230757802725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 5 + ], + "selected_reward": -0.043092746287584305 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vintage monochrome ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ticket display visible wear including frayed edges, softened corners, and a prominent center fold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine, spider-web-like cracks appear on the ticket's surface and does the ticket fracture into jagged fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady zoom while the fragments drift and tumble into the black void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607787050306797 + ], + "opposite_logprob_tokens": [ + -4.141077995300293 + ], + "logprob_reward": -0.01607787050306797, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607787050306797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06209808215498924, + 0.021454302594065666, + 0.21069863438606262, + 0.08138339221477509, + 0.12559406459331512, + 0.28381478786468506, + 0.08569961041212082, + 0.03291912004351616, + 0.026190131902694702, + 0.07014782726764679 + ], + "credit_map": [ + 0.006969672627747059, + 0.011053799651563168, + 0.004556236322969198, + 0.009027877822518349, + 0.010202756151556969, + 0.0044583422131836414, + 0.01008075475692749, + 0.005748644005507231, + 0.0022362670861184597, + 0.0025461772456765175, + 0.0022551878355443478, + 0.0035412751603871584, + 0.002879480365663767, + 0.002641161670908332, + 0.002585953800007701, + 0.0027688017580658197, + 0.007547416724264622, + 0.08936741948127747, + 0.017126424238085747, + 0.01044325903058052, + 0.06927124410867691, + 0.005913270637392998, + 0.005712260026484728, + 0.0053173513151705265, + 0.008501367643475533, + 0.012968899682164192, + 0.005743003915995359, + 0.010482481680810452, + 0.013968081213533878, + 0.00803668238222599, + 0.01483472716063261, + 0.0068481480702757835, + 0.006825190968811512, + 0.033379845321178436, + 0.0065886592492461205, + 0.01169176958501339, + 0.03431910276412964, + 0.013384099118411541, + 0.010062385350465775, + 0.009343024343252182, + 0.020152416080236435, + 0.04010707512497902, + 0.019871540367603302, + 0.03843053802847862, + 0.04495280236005783, + 0.0240264143794775, + 0.06847686320543289, + 0.0277971513569355, + 0.007301082368940115, + 0.012174617499113083, + 0.012529516592621803, + 0.009526072070002556, + 0.015006446279585361, + 0.007128500379621983, + 0.013326308690011501, + 0.008707068860530853, + 0.003627232275903225, + 0.005026680417358875, + 0.0024476125836372375, + 0.003087664023041725, + 0.007646380923688412, + 0.0038455557078123093, + 0.0040140338242053986, + 0.0032239635474979877, + 0.0030141633469611406, + 0.0030156432185322046, + 0.005408681463450193, + 0.0032053105533123016, + 0.0030767640564590693, + 0.003344916505739093, + 0.0021744437981396914, + 0.002950211986899376, + 0.004536984022706747, + 0.013644399121403694, + 0.003526912769302726, + 0.011599021963775158, + 0.01383580919355154, + 0.008009119890630245, + 0.007817540317773819, + 0.007178041618317366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 4 + ], + "selected_reward": -0.01607787050306797 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vintage monochrome ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ticket display visible wear including frayed edges, softened corners, and a prominent center fold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine, spider-web-like cracks appear on the ticket's surface and does the ticket fracture into jagged fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady zoom while the fragments drift and tumble into the black void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371873125433922 + ], + "opposite_logprob_tokens": [ + -3.4087188243865967 + ], + "logprob_reward": -0.03371873125433922, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371873125433922, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1532379537820816, + 0.01057063601911068, + 0.3057497441768646, + 0.07051214575767517, + 0.100145123898983, + 0.2453698366880417, + 0.046597812324762344, + 0.027617739513516426, + 0.01732460968196392, + 0.022874385118484497 + ], + "credit_map": [ + 0.016349220648407936, + 0.030452635139226913, + 0.017944462597370148, + 0.020524946972727776, + 0.03097611851990223, + 0.00997837819159031, + 0.016268206760287285, + 0.010743984021246433, + 0.001131890807300806, + 0.0014697308652102947, + 0.0017052211333066225, + 0.0014487507287412882, + 0.0014524556463584304, + 0.0008652310352772474, + 0.0011849626898765564, + 0.0013123935787007213, + 0.01246650144457817, + 0.12504322826862335, + 0.029361657798290253, + 0.015430117957293987, + 0.09665174037218094, + 0.00806517619639635, + 0.009124851785600185, + 0.009606482461094856, + 0.006284819450229406, + 0.011577201075851917, + 0.004568892531096935, + 0.010676506906747818, + 0.011531875468790531, + 0.00733029842376709, + 0.012504533864557743, + 0.006038018502295017, + 0.005184453912079334, + 0.03172678500413895, + 0.004387701861560345, + 0.00699609937146306, + 0.032090336084365845, + 0.0094619020819664, + 0.005256993230432272, + 0.005040852352976799, + 0.014405887573957443, + 0.034423261880874634, + 0.014240643009543419, + 0.03379659727215767, + 0.041285961866378784, + 0.019000805914402008, + 0.0631277859210968, + 0.02508891001343727, + 0.004602305591106415, + 0.006186621263623238, + 0.009667756967246532, + 0.0052056340500712395, + 0.006266396027058363, + 0.0040689436718821526, + 0.00580619415268302, + 0.004793962929397821, + 0.0029886059928685427, + 0.003723020665347576, + 0.0019455659203231335, + 0.0030613811686635017, + 0.004691274370998144, + 0.0030883285216987133, + 0.0034972422290593386, + 0.004622322507202625, + 0.0019467188976705074, + 0.001883647171780467, + 0.0033247373066842556, + 0.002777799963951111, + 0.002303660847246647, + 0.0019617637153714895, + 0.0013498686021193862, + 0.0017764117801561952, + 0.0018968658987432718, + 0.003716333070769906, + 0.0030055472161620855, + 0.0025789979845285416, + 0.003959436435252428, + 0.0026865224353969097, + 0.0019808353390544653, + 0.003049845341593027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.03371873125433922 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vintage monochrome ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ticket display visible wear including frayed edges, softened corners, and a prominent center fold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine, spider-web-like cracks appear on the ticket's surface and does the ticket fracture into jagged fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady zoom while the fragments drift and tumble into the black void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581606775522232 + ], + "opposite_logprob_tokens": [ + -3.8955817222595215 + ], + "logprob_reward": -0.020581606775522232, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581606775522232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11854879558086395, + 0.027447106316685677, + 0.278085857629776, + 0.07237110286951065, + 0.09972424805164337, + 0.1394651085138321, + 0.05760907381772995, + 0.07855269312858582, + 0.046977248042821884, + 0.08121871203184128 + ], + "credit_map": [ + 0.011035327799618244, + 0.025227153673768044, + 0.013136999681591988, + 0.015615210868418217, + 0.027230245992541313, + 0.007395587395876646, + 0.010569047182798386, + 0.008339223451912403, + 0.0032821842469274998, + 0.005016182083636522, + 0.0025811013765633106, + 0.0034159247297793627, + 0.004559468477964401, + 0.0020105321891605854, + 0.0027338252402842045, + 0.0038478896021842957, + 0.009621397592127323, + 0.11744322627782822, + 0.021809613332152367, + 0.012276941910386086, + 0.09382057189941406, + 0.007430671248584986, + 0.0078110527247190475, + 0.007872371934354305, + 0.006709035020321608, + 0.013291656970977783, + 0.004175973124802113, + 0.009351651184260845, + 0.011375865899026394, + 0.0055764215067029, + 0.01565314456820488, + 0.006237352732568979, + 0.009475271217525005, + 0.019495097920298576, + 0.006323071662336588, + 0.014917226508259773, + 0.018314892426133156, + 0.010127215646207333, + 0.01128623727709055, + 0.009785238653421402, + 0.012812335044145584, + 0.02943846583366394, + 0.01379362866282463, + 0.013233721256256104, + 0.02805284969508648, + 0.013814518228173256, + 0.01757965423166752, + 0.01073993556201458, + 0.004467786755412817, + 0.009025359526276588, + 0.007710807491093874, + 0.005754619836807251, + 0.010525131598114967, + 0.005230533424764872, + 0.010042066685855389, + 0.004852768033742905, + 0.007493411656469107, + 0.010766208171844482, + 0.007527424022555351, + 0.012790767475962639, + 0.010771548375487328, + 0.011750947684049606, + 0.009110510349273682, + 0.00834187027066946, + 0.004796851426362991, + 0.005454177036881447, + 0.006620210129767656, + 0.005361729767173529, + 0.009382179006934166, + 0.004671435337513685, + 0.004891717340797186, + 0.005798953119665384, + 0.0058580725453794, + 0.014352086931467056, + 0.006085362751036882, + 0.009008517488837242, + 0.015272313728928566, + 0.009683742187917233, + 0.008568438701331615, + 0.012390182353556156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.020581606775522232 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_9/index_tensor([2026, 2026])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_9/index_tensor([2026, 2026])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a246bdecd89ebdf8edcb5a97c83246543fd9e2c8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_9/index_tensor([2026, 2026])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle tracking shot follows a skeleton draped in a tattered black cloak as it walks across a cracked, barren, and parched landscape with a slow, steady, and ominous gait. The remarkably lifelike skeleton's yellowed bones are riddled with tiny cracks and weathering marks, and its hollow eye sockets seem to burn with an invisible energy. Its thick, rough black cloak, frayed to wisps at the edges, sways feebly in the dry wind. A bony hand grips a massive scythe taller than itself. The blade gleams a cold metallic sheen against the gloomy sky, the tip dragged steadily behind it, carving a shallow, continuous trail across the cracked earth. In the background, a leaden sky and a barren horizon stand out, with only a few dead, twisted tree roots reaching out from the ground like desperate claws. An even and oppressive diffuse light envelopes the entire scene, highlighting the pale texture of the bones and the deep black of the cloak, creating a cinematic atmosphere full of death, end and eternal loneliness.", + "seed": 298, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_9/index_2026_seed_298_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.030027109337970613, + "vlm_logprob": -0.030027109337970613, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.030027109337970613, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a skeleton wearing a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skeleton holding a massive scythe that is taller than itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a cracked, barren, and parched landscape with dead tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse, highlighting the pale bones against a gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006844290532171726 + ], + "opposite_logprob_tokens": [ + -5.381844520568848 + ], + "logprob_reward": -0.006844290532171726, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.006844290532171726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06030222773551941, + 0.06494160741567612, + 0.3472587466239929, + 0.13657373189926147, + 0.03307326138019562, + 0.020937921479344368, + 0.23183609545230865, + 0.039658986032009125, + 0.047106824815273285, + 0.018310658633708954 + ], + "credit_map": [ + 0.003984763752669096, + 0.00901823304593563, + 0.010603507980704308, + 0.009437290020287037, + 0.005893554538488388, + 0.0046009523794054985, + 0.011256496421992779, + 0.005507430527359247, + 0.0060961609706282616, + 0.006754695903509855, + 0.00746936397626996, + 0.010163220576941967, + 0.00514080049470067, + 0.013962428085505962, + 0.0061318096704781055, + 0.009223127737641335, + 0.05249100923538208, + 0.016081955283880234, + 0.12189637124538422, + 0.07558989524841309, + 0.007812122348695993, + 0.03011765517294407, + 0.0354660227894783, + 0.007803706917911768, + 0.020619189366698265, + 0.019917674362659454, + 0.025519685819745064, + 0.012924917973577976, + 0.009418748319149017, + 0.018552681431174278, + 0.01830606907606125, + 0.011314761824905872, + 0.004155376926064491, + 0.003959338646382093, + 0.005126587115228176, + 0.0037620330695062876, + 0.003972670529037714, + 0.003790718037635088, + 0.003916863817721605, + 0.004389673005789518, + 0.0023543282877653837, + 0.0030251776333898306, + 0.0021651703864336014, + 0.0026569149922579527, + 0.002377732889726758, + 0.002641326282173395, + 0.0028191774617880583, + 0.0028980921488255262, + 0.07248134911060333, + 0.011187360621988773, + 0.021154465153813362, + 0.008578956127166748, + 0.012000610120594501, + 0.0825207382440567, + 0.015241092070937157, + 0.008671536110341549, + 0.007104591000825167, + 0.003832563292235136, + 0.004584114532917738, + 0.002700252691283822, + 0.004560508299618959, + 0.009677832946181297, + 0.0036439644172787666, + 0.0035551590844988823, + 0.005019825417548418, + 0.005733989179134369, + 0.0042642815969884396, + 0.005166671238839626, + 0.007729667704552412, + 0.0046345870941877365, + 0.007370491977781057, + 0.007187304552644491, + 0.002990024397149682, + 0.0022556802723556757, + 0.0030356980860233307, + 0.0016352247912436724, + 0.0024493515957146883, + 0.0028893693815916777, + 0.0015885126776993275, + 0.0014667983632534742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 3 + ], + "selected_reward": -0.006844290532171726 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a skeleton wearing a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skeleton holding a massive scythe that is taller than itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a cracked, barren, and parched landscape with dead tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse, highlighting the pale bones against a gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295619606971741 + ], + "opposite_logprob_tokens": [ + -2.2379562854766846 + ], + "logprob_reward": -0.11295619606971741, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295619606971741, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13178761303424835, + 0.09607191383838654, + 0.48163700103759766, + 0.06566054373979568, + 0.05697275325655937, + 0.029868753626942635, + 0.039982348680496216, + 0.04252684861421585, + 0.029619012027978897, + 0.02587318606674671 + ], + "credit_map": [ + 0.022457363083958626, + 0.015198541805148125, + 0.028264205902814865, + 0.013939659111201763, + 0.01002274826169014, + 0.01764814183115959, + 0.013478088192641735, + 0.010778860189020634, + 0.00871314201503992, + 0.014913170598447323, + 0.011123072355985641, + 0.013152789324522018, + 0.00903978943824768, + 0.014942210167646408, + 0.012459845282137394, + 0.011727897450327873, + 0.08570275455713272, + 0.024840589612722397, + 0.1544891595840454, + 0.06971167027950287, + 0.022679001092910767, + 0.04902353137731552, + 0.03993260860443115, + 0.03525768965482712, + 0.007614600006490946, + 0.010347525589168072, + 0.010364118032157421, + 0.005181305576115847, + 0.005207805894315243, + 0.011287983506917953, + 0.00875958800315857, + 0.006897619925439358, + 0.009063889272511005, + 0.006398764438927174, + 0.011711660772562027, + 0.0052429139614105225, + 0.005652598571032286, + 0.006276227999478579, + 0.005923604127019644, + 0.006703093647956848, + 0.0036522296722978354, + 0.004749651532620192, + 0.005075441673398018, + 0.0030033423099666834, + 0.0023417791817337275, + 0.00422537000849843, + 0.0036834070924669504, + 0.0031375307589769363, + 0.009182075969874859, + 0.005170639604330063, + 0.004231854807585478, + 0.002747463993728161, + 0.0030637227464467287, + 0.009723001159727573, + 0.0031996474135667086, + 0.0026639404240995646, + 0.0062330481596291065, + 0.0054101357236504555, + 0.007359093986451626, + 0.0052925217896699905, + 0.0032804831862449646, + 0.005874822847545147, + 0.005287762265652418, + 0.0037889813538640738, + 0.0039246282540261745, + 0.0032889151480048895, + 0.004406181164085865, + 0.003102643880993128, + 0.004011454991996288, + 0.003210442140698433, + 0.003853647271171212, + 0.0038210987113416195, + 0.004814834799617529, + 0.0030004456639289856, + 0.004411696456372738, + 0.0018203181680291891, + 0.002886729082092643, + 0.004330747295171022, + 0.0022654093336313963, + 0.002343005733564496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.11295619606971741 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a skeleton wearing a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skeleton holding a massive scythe that is taller than itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a cracked, barren, and parched landscape with dead tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse, highlighting the pale bones against a gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536294758319855 + ], + "opposite_logprob_tokens": [ + -4.387536525726318 + ], + "logprob_reward": -0.012536294758319855, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536294758319855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12459923326969147, + 0.12398457527160645, + 0.14780566096305847, + 0.1260751485824585, + 0.12633293867111206, + 0.056959327310323715, + 0.13721144199371338, + 0.017689719796180725, + 0.06693154573440552, + 0.07241041213274002 + ], + "credit_map": [ + 0.021189739927649498, + 0.018945330753922462, + 0.019391125068068504, + 0.017031364142894745, + 0.009829223155975342, + 0.014938979409635067, + 0.015325377695262432, + 0.007948094047605991, + 0.0106069790199399, + 0.016956888139247894, + 0.01297492254525423, + 0.018940716981887817, + 0.009537206031382084, + 0.021785279735922813, + 0.018292484804987907, + 0.014890101738274097, + 0.019938422366976738, + 0.008995210751891136, + 0.04668593779206276, + 0.03273039683699608, + 0.004672071430832148, + 0.012200537137687206, + 0.016827499493956566, + 0.005755568388849497, + 0.012845158576965332, + 0.013440747745335102, + 0.017297402024269104, + 0.014768128283321857, + 0.01449397299438715, + 0.013974931091070175, + 0.021409163251519203, + 0.017845645546913147, + 0.01041362900286913, + 0.0232821237295866, + 0.015375964343547821, + 0.015079877339303493, + 0.014583488926291466, + 0.013566490262746811, + 0.01739390753209591, + 0.016637463122606277, + 0.005350819788873196, + 0.010033481754362583, + 0.006839141249656677, + 0.00545729324221611, + 0.005721108056604862, + 0.004387719091027975, + 0.00808093324303627, + 0.011088830418884754, + 0.044000331312417984, + 0.005820584949105978, + 0.009208195842802525, + 0.004145731683820486, + 0.007755640894174576, + 0.05149584263563156, + 0.009341825731098652, + 0.0054432787001132965, + 0.0027088490314781666, + 0.0023379004560410976, + 0.0036736158654093742, + 0.0016592135652899742, + 0.0013074127491563559, + 0.002174869878217578, + 0.0022892660927027464, + 0.0015385919250547886, + 0.008309664204716682, + 0.006124717649072409, + 0.0062193069607019424, + 0.007431782782077789, + 0.010081500746309757, + 0.007159073371440172, + 0.011576492339372635, + 0.010029005818068981, + 0.013144796714186668, + 0.012722872197628021, + 0.008855068124830723, + 0.004780413582921028, + 0.006851636338979006, + 0.012714993208646774, + 0.00789972860366106, + 0.005440902896225452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 4 + ], + "selected_reward": -0.012536294758319855 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a skeleton wearing a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skeleton holding a massive scythe that is taller than itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a cracked, barren, and parched landscape with dead tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse, highlighting the pale bones against a gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036100472789257765 + ], + "opposite_logprob_tokens": [ + -5.628610134124756 + ], + "logprob_reward": -0.0036100472789257765, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036100472789257765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12430289387702942, + 0.10175972431898117, + 0.06832010298967361, + 0.15779617428779602, + 0.20499062538146973, + 0.08740350604057312, + 0.04371343553066254, + 0.14671270549297333, + 0.03217857703566551, + 0.03282230347394943 + ], + "credit_map": [ + 0.019265811890363693, + 0.01482920628041029, + 0.026812560856342316, + 0.013828079216182232, + 0.009707020595669746, + 0.015606517903506756, + 0.013011996634304523, + 0.011241694912314415, + 0.011363455094397068, + 0.009729113429784775, + 0.014169836416840553, + 0.016808530315756798, + 0.006598432548344135, + 0.02353796176612377, + 0.010447923094034195, + 0.009104471653699875, + 0.012378084473311901, + 0.005545210093259811, + 0.0209808386862278, + 0.008502289652824402, + 0.0032472205348312855, + 0.00747669767588377, + 0.005922895390540361, + 0.004266866948455572, + 0.0394255593419075, + 0.01321424450725317, + 0.03137442469596863, + 0.009791070595383644, + 0.012689311057329178, + 0.019294051453471184, + 0.019459674134850502, + 0.012547844089567661, + 0.043202489614486694, + 0.014597970992326736, + 0.04092773422598839, + 0.015611174516379833, + 0.019129620864987373, + 0.029150929301977158, + 0.018545087426900864, + 0.023825621232390404, + 0.016127021983265877, + 0.011280430480837822, + 0.018018299713730812, + 0.006561963818967342, + 0.008500968106091022, + 0.00970293115824461, + 0.009344366379082203, + 0.007867520675063133, + 0.01135902851819992, + 0.002518540946766734, + 0.005246592685580254, + 0.002671865513548255, + 0.003041572403162718, + 0.012800001539289951, + 0.0035017768386751413, + 0.0025740554556250572, + 0.026495402678847313, + 0.013569640927016735, + 0.027712376788258553, + 0.015350919216871262, + 0.011016769334673882, + 0.021190950646996498, + 0.01941256783902645, + 0.011964076198637486, + 0.004517748486250639, + 0.004924972541630268, + 0.005921193398535252, + 0.0026298335287719965, + 0.0036962530575692654, + 0.004235065076500177, + 0.0029863794334232807, + 0.003267129184678197, + 0.0058105322532355785, + 0.003709790762513876, + 0.006161723285913467, + 0.002782749244943261, + 0.004852630198001862, + 0.004678267519921064, + 0.0023450693115592003, + 0.002481540199369192 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 7 + ], + "selected_reward": -0.0036100472789257765 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a skeleton wearing a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skeleton holding a massive scythe that is taller than itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a cracked, barren, and parched landscape with dead tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse, highlighting the pale bones against a gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014188718050718307 + ], + "opposite_logprob_tokens": [ + -4.264188766479492 + ], + "logprob_reward": -0.014188718050718307, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014188718050718307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12834115326404572, + 0.05086808279156685, + 0.19707562029361725, + 0.10148512572050095, + 0.09083456546068192, + 0.1101057231426239, + 0.0779557153582573, + 0.04754834994673729, + 0.04005725309252739, + 0.15572836995124817 + ], + "credit_map": [ + 0.011771908961236477, + 0.013265718705952168, + 0.03093971498310566, + 0.01619970239698887, + 0.012284111231565475, + 0.014577929861843586, + 0.014609625563025475, + 0.014692450873553753, + 0.006220820359885693, + 0.005999986082315445, + 0.006490715313702822, + 0.005616838578134775, + 0.004538626875728369, + 0.008582682348787785, + 0.005853876005858183, + 0.007564537227153778, + 0.021904272958636284, + 0.018811702728271484, + 0.045843835920095444, + 0.05010911822319031, + 0.0074346186593174934, + 0.012627377174794674, + 0.02891269139945507, + 0.011432033963501453, + 0.015110662207007408, + 0.00877078901976347, + 0.02102782018482685, + 0.009656442329287529, + 0.008039036765694618, + 0.016934586688876152, + 0.012578782625496387, + 0.009367015212774277, + 0.014569038525223732, + 0.008056160062551498, + 0.022667106240987778, + 0.007085240911692381, + 0.01102245133370161, + 0.011310170404613018, + 0.008715628646314144, + 0.007408776320517063, + 0.018917134031653404, + 0.015091354958713055, + 0.021212978288531303, + 0.008781901560723782, + 0.00908812414854765, + 0.016306012868881226, + 0.01201909501105547, + 0.008689126931130886, + 0.020488806068897247, + 0.007351001258939505, + 0.009512988850474358, + 0.00443101255223155, + 0.00490967882797122, + 0.02239428274333477, + 0.005354047752916813, + 0.003513903357088566, + 0.00676655862480402, + 0.006197055336087942, + 0.010909589007496834, + 0.004897393751889467, + 0.0033024174626916647, + 0.0068616378121078014, + 0.00477020675316453, + 0.003843495389446616, + 0.006340802647173405, + 0.006300339009612799, + 0.008707203902304173, + 0.0029504457488656044, + 0.003815258154645562, + 0.005886506289243698, + 0.002971806563436985, + 0.003084893338382244, + 0.031559694558382034, + 0.02097379043698311, + 0.02380184456706047, + 0.011259427294135094, + 0.015307210385799408, + 0.028144439682364464, + 0.013823427259922028, + 0.010858551599085331 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.014188718050718307 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot follows a skeleton draped in a tattered black cloak as it walks across a cracked, barren, and parched landscape with a slow, steady, and ominous gait. The remarkably lifelike skeleton's yellowed bones are riddled with tiny cracks and weathering marks, and its hollow eye sockets seem to burn with an invisible energy. Its thick, rough black cloak, frayed to wisps at the edges, sways feebly in the dry wind. A bony hand grips a massive scythe taller than itself. The blade gleams a cold metallic sheen against the gloomy sky, the tip dragged steadily behind it, carving a shallow, continuous trail across the cracked earth. In the background, a leaden sky and a barren horizon stand out, with only a few dead, twisted tree roots reaching out from the ground like desperate claws. An even and oppressive diffuse light envelopes the entire scene, highlighting the pale texture of the bones and the deep black of the cloak, creating a cinematic atmosphere full of death, end and eternal loneliness.", + "seed": 299, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_9/index_2026_seed_299_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04572062585502863, + "vlm_logprob": -0.04572062585502863, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.04572062585502863, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a skeleton wearing a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skeleton holding a massive scythe that is taller than itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a cracked, barren, and parched landscape with dead tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse, highlighting the pale bones against a gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013906137086451054 + ], + "opposite_logprob_tokens": [ + -4.513906002044678 + ], + "logprob_reward": -0.013906137086451054, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.013906137086451054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2549610137939453, + 0.14971263706684113, + 0.13879483938217163, + 0.050413791090250015, + 0.1003158763051033, + 0.021368030458688736, + 0.09656652808189392, + 0.05055122822523117, + 0.09893536567687988, + 0.038380734622478485 + ], + "credit_map": [ + 0.02208496630191803, + 0.027788879349827766, + 0.03144193813204765, + 0.02439725026488304, + 0.053236860781908035, + 0.03672018274664879, + 0.020221438258886337, + 0.03906949982047081, + 0.014822966419160366, + 0.01218855008482933, + 0.012737305834889412, + 0.01529103983193636, + 0.044893793761730194, + 0.014526781626045704, + 0.017608702182769775, + 0.017643490806221962, + 0.012579523958265781, + 0.013715233653783798, + 0.01338653452694416, + 0.012034165672957897, + 0.044120922684669495, + 0.012780647724866867, + 0.01664711721241474, + 0.013530691154301167, + 0.005526898894459009, + 0.005349030718207359, + 0.004883571993559599, + 0.003827120875939727, + 0.012342024594545364, + 0.004718523006886244, + 0.007416846230626106, + 0.0063497754745185375, + 0.010447785258293152, + 0.012061900459229946, + 0.013996470719575882, + 0.00904533825814724, + 0.01944010704755783, + 0.011484541930258274, + 0.012054753489792347, + 0.011784967966377735, + 0.0024677312467247248, + 0.002062638057395816, + 0.001985932234674692, + 0.0017453497275710106, + 0.006652267184108496, + 0.002448367653414607, + 0.001763535081408918, + 0.0022422093898057938, + 0.010425589047372341, + 0.01147201843559742, + 0.011110867373645306, + 0.008903197944164276, + 0.02280341647565365, + 0.011036544106900692, + 0.007285105064511299, + 0.013529795221984386, + 0.005803375504910946, + 0.006254622712731361, + 0.004957878962159157, + 0.006098686717450619, + 0.009138948284089565, + 0.0064003183506429195, + 0.003803415223956108, + 0.008093980140984058, + 0.014031060971319675, + 0.009497769176959991, + 0.013596736826002598, + 0.008722423575818539, + 0.014803613536059856, + 0.012780632823705673, + 0.007315200753509998, + 0.01818792335689068, + 0.004679504781961441, + 0.0036389378365129232, + 0.0054442137479782104, + 0.004163031931966543, + 0.006324050482362509, + 0.004671932198107243, + 0.00297326548025012, + 0.006485796067863703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.013906137086451054 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a skeleton wearing a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skeleton holding a massive scythe that is taller than itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a cracked, barren, and parched landscape with dead tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse, highlighting the pale bones against a gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297227442264557 + ], + "opposite_logprob_tokens": [ + -2.2379722595214844 + ], + "logprob_reward": -0.11297227442264557, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297227442264557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08843735605478287, + 0.07433321326971054, + 0.05052543431520462, + 0.170893132686615, + 0.038793329149484634, + 0.053221672773361206, + 0.03024914674460888, + 0.12335789203643799, + 0.32902878522872925, + 0.041160039603710175 + ], + "credit_map": [ + 0.007887792773544788, + 0.009079783223569393, + 0.01404145359992981, + 0.008670536801218987, + 0.017008239403367043, + 0.011056452989578247, + 0.009743441827595234, + 0.010949657298624516, + 0.00983364973217249, + 0.00698284013196826, + 0.01149679720401764, + 0.00876898318529129, + 0.009357397444546223, + 0.008324493654072285, + 0.008502592332661152, + 0.011066458187997341, + 0.005655722226947546, + 0.004750873427838087, + 0.005916888359934092, + 0.004874177742749453, + 0.009156687185168266, + 0.006025942508131266, + 0.006080924067646265, + 0.008064217865467072, + 0.01593022607266903, + 0.018672678619623184, + 0.013786911964416504, + 0.011558571830391884, + 0.053547658026218414, + 0.01962980628013611, + 0.01726207323372364, + 0.020505201071500778, + 0.003447851864621043, + 0.003252988215535879, + 0.0037620598450303078, + 0.004650660324841738, + 0.0042847683653235435, + 0.00395013764500618, + 0.00805542804300785, + 0.0073894355446100235, + 0.005600234027951956, + 0.004852863494306803, + 0.008138845674693584, + 0.007352548185735941, + 0.01080322265625, + 0.005793684162199497, + 0.004116957541555166, + 0.006563317961990833, + 0.004226857330650091, + 0.0029556616209447384, + 0.003333860542625189, + 0.002755494322627783, + 0.005673124920576811, + 0.004169380757957697, + 0.002540897810831666, + 0.004593867342919111, + 0.012915224768221378, + 0.014578307047486305, + 0.011405139230191708, + 0.016503561288118362, + 0.031048640608787537, + 0.011634108610451221, + 0.0064313155598938465, + 0.01884160377085209, + 0.04852290451526642, + 0.040220197290182114, + 0.055651817470788956, + 0.02415815368294716, + 0.04727049171924591, + 0.047528836876153946, + 0.017682362347841263, + 0.04799402132630348, + 0.007373353932052851, + 0.003061749506741762, + 0.007454166654497385, + 0.0037134098820388317, + 0.004416837822645903, + 0.006155933253467083, + 0.002316951984539628, + 0.006667635403573513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.11297227442264557 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a skeleton wearing a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skeleton holding a massive scythe that is taller than itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a cracked, barren, and parched landscape with dead tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse, highlighting the pale bones against a gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306066036224365 + ], + "opposite_logprob_tokens": [ + -3.168060779571533 + ], + "logprob_reward": -0.04306066036224365, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306066036224365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17890675365924835, + 0.07872284948825836, + 0.08428195118904114, + 0.09255896508693695, + 0.16943685710430145, + 0.046286799013614655, + 0.15429115295410156, + 0.08354339748620987, + 0.07262219488620758, + 0.039349112659692764 + ], + "credit_map": [ + 0.01908625289797783, + 0.02258162945508957, + 0.019896021112799644, + 0.017164211720228195, + 0.03260046988725662, + 0.023038430139422417, + 0.02077646180987358, + 0.023763282224535942, + 0.013124896213412285, + 0.006759960670024157, + 0.0062834760174155235, + 0.009923288598656654, + 0.01661819964647293, + 0.009896136820316315, + 0.009646777994930744, + 0.006470113527029753, + 0.010186715051531792, + 0.00783067848533392, + 0.01360420323908329, + 0.008950447663664818, + 0.01029854454100132, + 0.010781766846776009, + 0.00986432284116745, + 0.012765267863869667, + 0.008266911841928959, + 0.010454723611474037, + 0.006577573250979185, + 0.00802202895283699, + 0.026042496785521507, + 0.008930552750825882, + 0.015465732663869858, + 0.008798941038548946, + 0.01568959280848503, + 0.016841020435094833, + 0.018081694841384888, + 0.017945433035492897, + 0.04252214729785919, + 0.018485087901353836, + 0.02095675840973854, + 0.01891510933637619, + 0.004807054530829191, + 0.005006649997085333, + 0.005402617622166872, + 0.0046457950957119465, + 0.013081690296530724, + 0.005094744730740786, + 0.0036902348510921, + 0.004558012820780277, + 0.01574295200407505, + 0.017857437953352928, + 0.011869166046380997, + 0.016316024586558342, + 0.03295597434043884, + 0.01849064975976944, + 0.014076354913413525, + 0.02698259800672531, + 0.01002556923776865, + 0.00720969820395112, + 0.016059212386608124, + 0.00795267429202795, + 0.009110018610954285, + 0.008773209527134895, + 0.00513574481010437, + 0.01927727460861206, + 0.014339755289256573, + 0.005637246649712324, + 0.010325889103114605, + 0.006369675509631634, + 0.00925785768777132, + 0.010885358788073063, + 0.004814979620277882, + 0.010991431772708893, + 0.007750994525849819, + 0.0037880998570472, + 0.005505397450178862, + 0.003552999347448349, + 0.00420528557151556, + 0.006173137109726667, + 0.003465454326942563, + 0.0049077426083385944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.04306066036224365 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a skeleton wearing a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skeleton holding a massive scythe that is taller than itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a cracked, barren, and parched landscape with dead tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse, highlighting the pale bones against a gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057226374745369 + ], + "opposite_logprob_tokens": [ + -3.8955721855163574 + ], + "logprob_reward": -0.02057226374745369, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057226374745369, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1348850280046463, + 0.09664910286664963, + 0.18664678931236267, + 0.17089718580245972, + 0.15082886815071106, + 0.08331582695245743, + 0.03388764336705208, + 0.02092985063791275, + 0.04684598371386528, + 0.07511366903781891 + ], + "credit_map": [ + 0.014111832715570927, + 0.014213384129106998, + 0.017415499314665794, + 0.011906993575394154, + 0.024137921631336212, + 0.017536887899041176, + 0.01469176635146141, + 0.020870763808488846, + 0.01104377955198288, + 0.007290287408977747, + 0.011314763687551022, + 0.010623492300510406, + 0.025721272453665733, + 0.008692780509591103, + 0.011551468633115292, + 0.010411262512207031, + 0.017221717163920403, + 0.017910761758685112, + 0.014731617644429207, + 0.012886553071439266, + 0.06772919744253159, + 0.017355037853121758, + 0.02165014110505581, + 0.01716177724301815, + 0.014599236659705639, + 0.018441608175635338, + 0.014663055539131165, + 0.013249795883893967, + 0.05238089710474014, + 0.016401827335357666, + 0.023550689220428467, + 0.0176100991666317, + 0.014509012922644615, + 0.010105381719768047, + 0.01767192780971527, + 0.017352867871522903, + 0.01676066964864731, + 0.015271967276930809, + 0.027527591213583946, + 0.03162946552038193, + 0.007386039476841688, + 0.010250316001474857, + 0.007690384518355131, + 0.0077926116064190865, + 0.02380465902388096, + 0.00980562437325716, + 0.007557318080216646, + 0.009028883650898933, + 0.004008650779724121, + 0.004236757289618254, + 0.00331508438102901, + 0.004481744486838579, + 0.006374482996761799, + 0.004050263669341803, + 0.003032831009477377, + 0.004387829918414354, + 0.0023013183381408453, + 0.0021028982009738684, + 0.0021295961923897266, + 0.002239856170490384, + 0.003127409378066659, + 0.002574566751718521, + 0.001963829854503274, + 0.004490375984460115, + 0.00786042120307684, + 0.004734619986265898, + 0.008018361404538155, + 0.004248636309057474, + 0.0059507619589567184, + 0.006231359206140041, + 0.002854675520211458, + 0.006947149522602558, + 0.010011672042310238, + 0.008192379027605057, + 0.013777898624539375, + 0.006918010767549276, + 0.011122017167508602, + 0.010965727269649506, + 0.004171330016106367, + 0.009954642504453659 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 4 + ], + "selected_reward": -0.02057226374745369 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a skeleton wearing a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skeleton holding a massive scythe that is taller than itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a cracked, barren, and parched landscape with dead tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse, highlighting the pale bones against a gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809179365634918 + ], + "opposite_logprob_tokens": [ + -3.2880918979644775 + ], + "logprob_reward": -0.03809179365634918, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809179365634918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22058288753032684, + 0.051688626408576965, + 0.04498961567878723, + 0.0615464486181736, + 0.09103018045425415, + 0.07390930503606796, + 0.05621769651770592, + 0.11269606649875641, + 0.04519046097993851, + 0.24214865267276764 + ], + "credit_map": [ + 0.023677580058574677, + 0.0281083881855011, + 0.018179655075073242, + 0.02045617625117302, + 0.04395224153995514, + 0.029994424432516098, + 0.020428631454706192, + 0.03578580170869827, + 0.007908592000603676, + 0.0037202348466962576, + 0.004168909043073654, + 0.006164821330457926, + 0.009876111522316933, + 0.006457765586674213, + 0.005807339213788509, + 0.007584854960441589, + 0.005977401044219732, + 0.0038053507450968027, + 0.004432528745383024, + 0.004618437495082617, + 0.007437885273247957, + 0.0056890444830060005, + 0.006409011781215668, + 0.006619954947382212, + 0.006381659302860498, + 0.008098811842501163, + 0.00677106948569417, + 0.006243202835321426, + 0.0132652772590518, + 0.005837535485625267, + 0.0068115065805613995, + 0.008137385360896587, + 0.009482510387897491, + 0.007010190282016993, + 0.010747105814516544, + 0.008311509154736996, + 0.01593138463795185, + 0.010556399822235107, + 0.016741793602705002, + 0.012249285355210304, + 0.007095456589013338, + 0.009945093654096127, + 0.00731950206682086, + 0.006560574285686016, + 0.022824639454483986, + 0.0081958482041955, + 0.006113537587225437, + 0.005854655522853136, + 0.0065451632253825665, + 0.006795038469135761, + 0.006058939266949892, + 0.00753889512270689, + 0.00853564403951168, + 0.007035862188786268, + 0.005938137881457806, + 0.007770019583404064, + 0.018773982301354408, + 0.011928415857255459, + 0.013187555596232414, + 0.009659785777330399, + 0.017706748098134995, + 0.01563253067433834, + 0.00825550127774477, + 0.017551548779010773, + 0.009515187703073025, + 0.0047173332422971725, + 0.005856919102370739, + 0.004035158082842827, + 0.005216252990067005, + 0.006708947475999594, + 0.0028922988567501307, + 0.00624836515635252, + 0.06498468667268753, + 0.014499424025416374, + 0.038678303360939026, + 0.015806982293725014, + 0.01718015968799591, + 0.04668867960572243, + 0.010358252562582493, + 0.033952176570892334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.03809179365634918 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot follows a skeleton draped in a tattered black cloak as it walks across a cracked, barren, and parched landscape with a slow, steady, and ominous gait. The remarkably lifelike skeleton's yellowed bones are riddled with tiny cracks and weathering marks, and its hollow eye sockets seem to burn with an invisible energy. Its thick, rough black cloak, frayed to wisps at the edges, sways feebly in the dry wind. A bony hand grips a massive scythe taller than itself. The blade gleams a cold metallic sheen against the gloomy sky, the tip dragged steadily behind it, carving a shallow, continuous trail across the cracked earth. In the background, a leaden sky and a barren horizon stand out, with only a few dead, twisted tree roots reaching out from the ground like desperate claws. An even and oppressive diffuse light envelopes the entire scene, highlighting the pale texture of the bones and the deep black of the cloak, creating a cinematic atmosphere full of death, end and eternal loneliness.", + "seed": 300, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_9/index_2026_seed_300_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03266104701906443, + "vlm_logprob": -0.03266104701906443, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.03266104701906443, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a skeleton wearing a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skeleton holding a massive scythe that is taller than itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a cracked, barren, and parched landscape with dead tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse, highlighting the pale bones against a gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0053092907182872295 + ], + "opposite_logprob_tokens": [ + -6.005309104919434 + ], + "logprob_reward": -0.0053092907182872295, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0053092907182872295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13430921733379364, + 0.08885359764099121, + 0.028650997206568718, + 0.12976370751857758, + 0.0780194103717804, + 0.0740061029791832, + 0.20238909125328064, + 0.16099479794502258, + 0.05460332706570625, + 0.04840974882245064 + ], + "credit_map": [ + 0.01832493394613266, + 0.04303066432476044, + 0.012497213669121265, + 0.012775045819580555, + 0.011190989054739475, + 0.011662484146654606, + 0.011193417944014072, + 0.01363446656614542, + 0.009786481969058514, + 0.011826325207948685, + 0.01209984626621008, + 0.011302225291728973, + 0.010999123565852642, + 0.011215499602258205, + 0.010141302831470966, + 0.011482792906463146, + 0.0029721688479185104, + 0.007775329053401947, + 0.0027863862924277782, + 0.002117049414664507, + 0.002665014937520027, + 0.003253599861636758, + 0.0019053241703659296, + 0.0051761227659881115, + 0.011987926438450813, + 0.028463061898946762, + 0.016356123611330986, + 0.01614939607679844, + 0.008652259595692158, + 0.014724868349730968, + 0.01574924774467945, + 0.017680823802947998, + 0.007640144322067499, + 0.01385711133480072, + 0.00928028579801321, + 0.009362085722386837, + 0.005934709217399359, + 0.009692885912954807, + 0.01146842259913683, + 0.010783763602375984, + 0.005469423718750477, + 0.01342050638049841, + 0.013280992396175861, + 0.00983114168047905, + 0.0070582106709480286, + 0.007956178858876228, + 0.006963294465094805, + 0.010026359930634499, + 0.026116790249943733, + 0.026821797713637352, + 0.02754135988652706, + 0.014840219169855118, + 0.012746528722345829, + 0.028163697570562363, + 0.01981527730822563, + 0.04634341225028038, + 0.018568139523267746, + 0.04357513040304184, + 0.017332427203655243, + 0.01624683476984501, + 0.018920566886663437, + 0.01692459173500538, + 0.014347296208143234, + 0.015079810284078121, + 0.007012161426246166, + 0.004600293934345245, + 0.009272590279579163, + 0.004820330999791622, + 0.0050054327584803104, + 0.007224166765809059, + 0.006180292926728725, + 0.0104880565777421, + 0.006791082676500082, + 0.005491575691848993, + 0.005135442595928907, + 0.00503455800935626, + 0.0043665324337780476, + 0.006631424650549889, + 0.006127283442765474, + 0.008831847459077835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.0053092907182872295 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a skeleton wearing a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skeleton holding a massive scythe that is taller than itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a cracked, barren, and parched landscape with dead tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse, highlighting the pale bones against a gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295513063669205 + ], + "opposite_logprob_tokens": [ + -2.237955093383789 + ], + "logprob_reward": -0.11295513063669205, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295513063669205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11820980161428452, + 0.07864363491535187, + 0.06150124967098236, + 0.10108242928981781, + 0.10086899995803833, + 0.10035213828086853, + 0.15728020668029785, + 0.12361408770084381, + 0.043515585362911224, + 0.11493191868066788 + ], + "credit_map": [ + 0.02124880813062191, + 0.014273243956267834, + 0.01307189092040062, + 0.008862759917974472, + 0.014401044696569443, + 0.014268332161009312, + 0.01304601226001978, + 0.019037703052163124, + 0.007959077134728432, + 0.012253187596797943, + 0.011129055172204971, + 0.00871575903147459, + 0.0063655986450612545, + 0.012222912162542343, + 0.007183562498539686, + 0.012814481742680073, + 0.006506838835775852, + 0.012932893820106983, + 0.005356139969080687, + 0.0067562926560640335, + 0.010113613680005074, + 0.007452488411217928, + 0.005423898808658123, + 0.006959076505154371, + 0.009304867126047611, + 0.018332364037632942, + 0.00806983932852745, + 0.009064232930541039, + 0.0075784181244671345, + 0.013064965605735779, + 0.01650838367640972, + 0.0191593449562788, + 0.009406981989741325, + 0.025678202509880066, + 0.0089315976947546, + 0.01035214215517044, + 0.006250922568142414, + 0.012287299148738384, + 0.017707103863358498, + 0.010254746302962303, + 0.010597337037324905, + 0.020475322380661964, + 0.015522442758083344, + 0.01139784138649702, + 0.011341736651957035, + 0.011412074789404869, + 0.007514783646911383, + 0.012090591713786125, + 0.01476831641048193, + 0.04229827970266342, + 0.014100638218224049, + 0.011468108743429184, + 0.013593195006251335, + 0.01715928688645363, + 0.01622130535542965, + 0.027671055868268013, + 0.01113592553883791, + 0.030915584415197372, + 0.011188022792339325, + 0.012593917548656464, + 0.012787277810275555, + 0.013786316849291325, + 0.012418100610375404, + 0.018788933753967285, + 0.005042736418545246, + 0.004194946959614754, + 0.005800346843898296, + 0.0035480489023029804, + 0.0034404913894832134, + 0.006851966958492994, + 0.005288982763886452, + 0.009348057210445404, + 0.012767674401402473, + 0.010212162509560585, + 0.009134858846664429, + 0.011903821490705013, + 0.015081275254487991, + 0.01585082896053791, + 0.018825016915798187, + 0.021156279370188713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.11295513063669205 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a skeleton wearing a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skeleton holding a massive scythe that is taller than itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a cracked, barren, and parched landscape with dead tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse, highlighting the pale bones against a gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007626583334058523 + ], + "opposite_logprob_tokens": [ + -4.882626533508301 + ], + "logprob_reward": -0.007626583334058523, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007626583334058523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15330438315868378, + 0.12864316999912262, + 0.06514594703912735, + 0.08747842162847519, + 0.12046639621257782, + 0.05744299292564392, + 0.1521686613559723, + 0.10347606986761093, + 0.0780305564403534, + 0.0538434199988842 + ], + "credit_map": [ + 0.020684728398919106, + 0.030225783586502075, + 0.018810538575053215, + 0.014692762866616249, + 0.016835030168294907, + 0.01767083816230297, + 0.01643413119018078, + 0.01795056462287903, + 0.01784219779074192, + 0.02357310988008976, + 0.014040768146514893, + 0.012994997203350067, + 0.01477726735174656, + 0.017392614856362343, + 0.01066606119275093, + 0.017356140539050102, + 0.007938746362924576, + 0.013231419026851654, + 0.008238048292696476, + 0.005776205100119114, + 0.007887382060289383, + 0.007459782995283604, + 0.005293795373290777, + 0.00932056363672018, + 0.011288771405816078, + 0.02292381040751934, + 0.010092377662658691, + 0.009959862567484379, + 0.005147777032107115, + 0.008795480243861675, + 0.007166889030486345, + 0.012103443033993244, + 0.01152463536709547, + 0.02276962623000145, + 0.007853688672184944, + 0.009825145825743675, + 0.007607280742377043, + 0.017346570268273354, + 0.024621056392788887, + 0.01891838200390339, + 0.00682218000292778, + 0.01801433227956295, + 0.0060826377011835575, + 0.006619795225560665, + 0.003831727895885706, + 0.005483598448336124, + 0.004381257575005293, + 0.006207460071891546, + 0.02019369602203369, + 0.021021461114287376, + 0.022750580683350563, + 0.01186324842274189, + 0.012118518352508545, + 0.018636144697666168, + 0.016286062076687813, + 0.029298949986696243, + 0.011906959116458893, + 0.02491569332778454, + 0.010541939176619053, + 0.009852792136371136, + 0.010782552883028984, + 0.012137653306126595, + 0.009075382724404335, + 0.014263088814914227, + 0.009780107997357845, + 0.0063031516037881374, + 0.018024710938334465, + 0.008794141001999378, + 0.005541935563087463, + 0.011380936950445175, + 0.009482445195317268, + 0.008723122999072075, + 0.006264642812311649, + 0.0046051498502492905, + 0.004589705262333155, + 0.0065575260668993, + 0.0058733331970870495, + 0.006996144540607929, + 0.00901535339653492, + 0.00994155928492546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.007626583334058523 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a skeleton wearing a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skeleton holding a massive scythe that is taller than itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a cracked, barren, and parched landscape with dead tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse, highlighting the pale bones against a gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007619958370923996 + ], + "opposite_logprob_tokens": [ + -4.882619857788086 + ], + "logprob_reward": -0.007619958370923996, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007619958370923996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22279240190982819, + 0.07682861387729645, + 0.035631123930215836, + 0.1126381978392601, + 0.12286671996116638, + 0.07758650928735733, + 0.12994417548179626, + 0.07006875425577164, + 0.08583415299654007, + 0.06580935418605804 + ], + "credit_map": [ + 0.039134904742240906, + 0.03522954508662224, + 0.032264236360788345, + 0.013970494270324707, + 0.020606057718396187, + 0.03017912618815899, + 0.018851501867175102, + 0.03255653753876686, + 0.008527558296918869, + 0.009059950709342957, + 0.009987064637243748, + 0.008959713391959667, + 0.010427494533360004, + 0.009235629811882973, + 0.009595193900167942, + 0.011036010459065437, + 0.0046054585836827755, + 0.007195057813078165, + 0.0034540740307420492, + 0.002959094475954771, + 0.003384844632819295, + 0.004646861460059881, + 0.0025726582389324903, + 0.00681307353079319, + 0.013334096409380436, + 0.017687290906906128, + 0.011859369464218616, + 0.00970546342432499, + 0.0064658066257834435, + 0.017161646857857704, + 0.010752302594482899, + 0.025672225281596184, + 0.010863867588341236, + 0.01709938608109951, + 0.009488639421761036, + 0.008108747191727161, + 0.007857839576900005, + 0.020734675228595734, + 0.01812751218676567, + 0.030586056411266327, + 0.0062454091385006905, + 0.012709930539131165, + 0.009867757558822632, + 0.007876647636294365, + 0.00672211404889822, + 0.01080165896564722, + 0.005620492156594992, + 0.017742503434419632, + 0.015324234031140804, + 0.02150561474263668, + 0.01727846823632717, + 0.010208030231297016, + 0.009486200287938118, + 0.016922250390052795, + 0.013369954191148281, + 0.02584943361580372, + 0.00799709651619196, + 0.01858547516167164, + 0.005637814290821552, + 0.006308127660304308, + 0.007034203503280878, + 0.007992886938154697, + 0.006201041862368584, + 0.01031210646033287, + 0.012258784845471382, + 0.006697701755911112, + 0.020028920844197273, + 0.0074087209068238735, + 0.004836612846702337, + 0.012811519205570221, + 0.00807597953826189, + 0.013715914450585842, + 0.009960023686289787, + 0.008093150332570076, + 0.006324698682874441, + 0.006481146905571222, + 0.005856458097696304, + 0.008742619305849075, + 0.007778624072670937, + 0.012572629377245903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.007619958370923996 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a skeleton wearing a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skeleton holding a massive scythe that is taller than itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a cracked, barren, and parched landscape with dead tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse, highlighting the pale bones against a gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029794272035360336 + ], + "opposite_logprob_tokens": [ + -3.529794216156006 + ], + "logprob_reward": -0.029794272035360336, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029794272035360336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2011314481496811, + 0.09643212705850601, + 0.03302885219454765, + 0.0720471739768982, + 0.10960034281015396, + 0.05259481817483902, + 0.186321422457695, + 0.17124594748020172, + 0.04567248001694679, + 0.03192535787820816 + ], + "credit_map": [ + 0.026612481102347374, + 0.044459786266088486, + 0.023266078904271126, + 0.019991662353277206, + 0.018260011449456215, + 0.020644012838602066, + 0.01712823659181595, + 0.030769167467951775, + 0.011910957284271717, + 0.017091823741793633, + 0.01410756167024374, + 0.012887179851531982, + 0.010430353693664074, + 0.009738393127918243, + 0.008951793424785137, + 0.011314070783555508, + 0.003870407585054636, + 0.0049246568232774734, + 0.004097742959856987, + 0.0036692593712359667, + 0.003517488017678261, + 0.004353740252554417, + 0.0031174493487924337, + 0.00547810597345233, + 0.007388380356132984, + 0.014413205906748772, + 0.00996509287506342, + 0.008461826480925083, + 0.004293115343898535, + 0.00834247563034296, + 0.00829530693590641, + 0.010887776501476765, + 0.011740045621991158, + 0.02511131949722767, + 0.009760881774127483, + 0.010152559727430344, + 0.009701605886220932, + 0.01389395073056221, + 0.016421249136328697, + 0.012818724848330021, + 0.005065359175205231, + 0.010071045719087124, + 0.009574458003044128, + 0.007671831641346216, + 0.004415243864059448, + 0.005115111358463764, + 0.004136306699365377, + 0.0065454584546387196, + 0.02172057330608368, + 0.03322562947869301, + 0.022871607914566994, + 0.013415644876658916, + 0.01179499737918377, + 0.02495160698890686, + 0.01645032875239849, + 0.04189102724194527, + 0.019351398572325706, + 0.04886632412672043, + 0.015885548666119576, + 0.016748109832406044, + 0.02172108180820942, + 0.017588119953870773, + 0.014260347001254559, + 0.016825025901198387, + 0.005626581609249115, + 0.003390141762793064, + 0.00616616802290082, + 0.00478614866733551, + 0.003638423513621092, + 0.005770145449787378, + 0.007460395339876413, + 0.008834476582705975, + 0.004053989890962839, + 0.003517058677971363, + 0.004454628098756075, + 0.0037879892624914646, + 0.00362351443618536, + 0.00446435809135437, + 0.0036740724463015795, + 0.0043497453443706036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.029794272035360336 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot follows a skeleton draped in a tattered black cloak as it walks across a cracked, barren, and parched landscape with a slow, steady, and ominous gait. The remarkably lifelike skeleton's yellowed bones are riddled with tiny cracks and weathering marks, and its hollow eye sockets seem to burn with an invisible energy. Its thick, rough black cloak, frayed to wisps at the edges, sways feebly in the dry wind. A bony hand grips a massive scythe taller than itself. The blade gleams a cold metallic sheen against the gloomy sky, the tip dragged steadily behind it, carving a shallow, continuous trail across the cracked earth. In the background, a leaden sky and a barren horizon stand out, with only a few dead, twisted tree roots reaching out from the ground like desperate claws. An even and oppressive diffuse light envelopes the entire scene, highlighting the pale texture of the bones and the deep black of the cloak, creating a cinematic atmosphere full of death, end and eternal loneliness.", + "seed": 301, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000000/group_9/index_2026_seed_301_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.036618994153104725, + "vlm_logprob": -0.036618994153104725, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.036618994153104725, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a skeleton wearing a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skeleton holding a massive scythe that is taller than itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a cracked, barren, and parched landscape with dead tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse, highlighting the pale bones against a gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008083367720246315 + ], + "opposite_logprob_tokens": [ + -5.258083343505859 + ], + "logprob_reward": -0.008083367720246315, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.008083367720246315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1812509298324585, + 0.06776551902294159, + 0.14881771802902222, + 0.1256306767463684, + 0.051956627517938614, + 0.03004773147404194, + 0.029940007254481316, + 0.26055440306663513, + 0.03449781611561775, + 0.06953857094049454 + ], + "credit_map": [ + 0.03243209421634674, + 0.020754823461174965, + 0.02293117344379425, + 0.02263951674103737, + 0.014771311543881893, + 0.03946724906563759, + 0.011168367229402065, + 0.017086394131183624, + 0.007283491548150778, + 0.005937236826866865, + 0.014514986425638199, + 0.00939868949353695, + 0.010879375971853733, + 0.00804644450545311, + 0.005164134316146374, + 0.006541165988892317, + 0.016895312815904617, + 0.012077916413545609, + 0.025534046813845634, + 0.016893770545721054, + 0.03351113200187683, + 0.019554220139980316, + 0.015775734558701515, + 0.008575588464736938, + 0.014432160183787346, + 0.010929521173238754, + 0.013743134215474129, + 0.023527875542640686, + 0.019333980977535248, + 0.018259143456816673, + 0.01265723630785942, + 0.012747623957693577, + 0.00663984939455986, + 0.01115391030907631, + 0.005115613341331482, + 0.00517541216686368, + 0.006618776824325323, + 0.007079079747200012, + 0.004206595476716757, + 0.005967386066913605, + 0.0033324221149086952, + 0.0032262790482491255, + 0.004611556883901358, + 0.004228890873491764, + 0.0036788207944482565, + 0.003932856023311615, + 0.0032181739807128906, + 0.0038187317550182343, + 0.0040366267785429955, + 0.002544577233493328, + 0.004669656045734882, + 0.0030212411656975746, + 0.004208013415336609, + 0.004742599558085203, + 0.0031967225950211287, + 0.0035205695312470198, + 0.029538840055465698, + 0.035859305411577225, + 0.030141856521368027, + 0.033399734646081924, + 0.03439699858427048, + 0.03536970540881157, + 0.02423061802983284, + 0.03761734440922737, + 0.0037747924216091633, + 0.00446367496624589, + 0.004177901428192854, + 0.003987634088844061, + 0.006619323510676622, + 0.004319806583225727, + 0.004073309246450663, + 0.0030813715420663357, + 0.008916521444916725, + 0.01213584654033184, + 0.008724040351808071, + 0.006893886718899012, + 0.010813024826347828, + 0.009831567294895649, + 0.00469208275899291, + 0.007531601469963789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.008083367720246315 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a skeleton wearing a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skeleton holding a massive scythe that is taller than itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a cracked, barren, and parched landscape with dead tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse, highlighting the pale bones against a gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12707188725471497 + ], + "opposite_logprob_tokens": [ + -2.1270718574523926 + ], + "logprob_reward": -0.12707188725471497, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12707188725471497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1368442326784134, + 0.04856831952929497, + 0.05823304131627083, + 0.06925772875547409, + 0.09767434746026993, + 0.06367325037717819, + 0.06764387339353561, + 0.2649233639240265, + 0.04500627517700195, + 0.14817556738853455 + ], + "credit_map": [ + 0.016877735033631325, + 0.012235134840011597, + 0.02148621715605259, + 0.011435562744736671, + 0.014668856747448444, + 0.02201472781598568, + 0.011439988389611244, + 0.026686014607548714, + 0.005497236270457506, + 0.004680606536567211, + 0.006995663046836853, + 0.008081316947937012, + 0.005376484710723162, + 0.005112822633236647, + 0.004398389253765345, + 0.008425801061093807, + 0.0078880675137043, + 0.005955853965133429, + 0.010988494381308556, + 0.006844191811978817, + 0.0069571202620863914, + 0.009276185184717178, + 0.005379106383770704, + 0.004944021347910166, + 0.008831786923110485, + 0.00595548190176487, + 0.008322243578732014, + 0.0067116194404661655, + 0.013181715272367, + 0.011323151178658009, + 0.007438621949404478, + 0.007493106182664633, + 0.015658551827073097, + 0.012982768006622791, + 0.010749158449470997, + 0.006658016704022884, + 0.011111673898994923, + 0.019562818109989166, + 0.008956526406109333, + 0.01199483871459961, + 0.006912751588970423, + 0.006997276563197374, + 0.009043144062161446, + 0.007590665016323328, + 0.01056577917188406, + 0.007870623841881752, + 0.00741789024323225, + 0.0072751170955598354, + 0.011132649146020412, + 0.003745954716578126, + 0.014751013368368149, + 0.004784011747688055, + 0.005693178158253431, + 0.014604421332478523, + 0.004694759380072355, + 0.008237885311245918, + 0.030767960473895073, + 0.03521629795432091, + 0.03470010682940483, + 0.031681545078754425, + 0.03050754964351654, + 0.03595734387636185, + 0.0277535580098629, + 0.03833901137113571, + 0.005955162458121777, + 0.004658957477658987, + 0.007989213801920414, + 0.005368820391595364, + 0.004372742958366871, + 0.007422460708767176, + 0.003975308034569025, + 0.005263608880341053, + 0.022957753390073776, + 0.01239786297082901, + 0.022182850167155266, + 0.02403058297932148, + 0.014374153688549995, + 0.020610148087143898, + 0.011884603649377823, + 0.019737619906663895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.12707188725471497 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a skeleton wearing a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skeleton holding a massive scythe that is taller than itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a cracked, barren, and parched landscape with dead tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse, highlighting the pale bones against a gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537825852632523 + ], + "opposite_logprob_tokens": [ + -4.387537956237793 + ], + "logprob_reward": -0.012537825852632523, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537825852632523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12662449479103088, + 0.13718082010746002, + 0.10251767933368683, + 0.0523262619972229, + 0.11690692603588104, + 0.04116798937320709, + 0.043363187462091446, + 0.14580471813678741, + 0.05849180743098259, + 0.17561613023281097 + ], + "credit_map": [ + 0.02471054531633854, + 0.01319730281829834, + 0.016415223479270935, + 0.01198375690728426, + 0.014568118378520012, + 0.02672603167593479, + 0.009128153324127197, + 0.009895365685224533, + 0.01575573906302452, + 0.013566263020038605, + 0.02785140462219715, + 0.022159835323691368, + 0.024567516520619392, + 0.010216710157692432, + 0.014416809193789959, + 0.008646536618471146, + 0.007707921322435141, + 0.0074707032181322575, + 0.01562991365790367, + 0.01044792402535677, + 0.027563007548451424, + 0.010375866666436195, + 0.012155959382653236, + 0.011166376993060112, + 0.00590120255947113, + 0.004823591560125351, + 0.007569576613605022, + 0.005686663556843996, + 0.010914728045463562, + 0.007776083890348673, + 0.004828984383493662, + 0.004825431853532791, + 0.020888714119791985, + 0.024777673184871674, + 0.00891660526394844, + 0.008907010778784752, + 0.013469254598021507, + 0.02146221324801445, + 0.008258113637566566, + 0.010227344930171967, + 0.005416477099061012, + 0.004503621719777584, + 0.004266041796654463, + 0.005008706357330084, + 0.0059382859617471695, + 0.007076808717101812, + 0.0038868950214236975, + 0.005071150604635477, + 0.00590195506811142, + 0.0043032048270106316, + 0.00676318584010005, + 0.00484442338347435, + 0.004702272359281778, + 0.007982774637639523, + 0.0040672882460057735, + 0.004798082169145346, + 0.01554783433675766, + 0.018611401319503784, + 0.020178575068712234, + 0.0220150426030159, + 0.015579374507069588, + 0.018126200884580612, + 0.01494161132723093, + 0.020804675295948982, + 0.006138119380921125, + 0.006717078387737274, + 0.007830433547496796, + 0.00875041913241148, + 0.0077615343034267426, + 0.0072923642583191395, + 0.008024047128856182, + 0.005977812688797712, + 0.02242421731352806, + 0.028542060405015945, + 0.02356666326522827, + 0.020992273464798927, + 0.02373754419386387, + 0.022301984950900078, + 0.017287714406847954, + 0.016763679683208466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 1 + ], + "selected_reward": -0.012537825852632523 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a skeleton wearing a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skeleton holding a massive scythe that is taller than itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a cracked, barren, and parched landscape with dead tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse, highlighting the pale bones against a gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001709667849354446 + ], + "opposite_logprob_tokens": [ + -6.376709461212158 + ], + "logprob_reward": -0.001709667849354446, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001709667849354446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1806592047214508, + 0.08875580132007599, + 0.19235169887542725, + 0.1509266048669815, + 0.07798409461975098, + 0.05272163078188896, + 0.05384440720081329, + 0.06201443821191788, + 0.05133877322077751, + 0.08940333127975464 + ], + "credit_map": [ + 0.024103550240397453, + 0.01756993681192398, + 0.025416238233447075, + 0.028248533606529236, + 0.018030211329460144, + 0.030219636857509613, + 0.013430655002593994, + 0.023640448227524757, + 0.011092990636825562, + 0.007933983579277992, + 0.02094569057226181, + 0.01428905688226223, + 0.010760216042399406, + 0.011517741717398167, + 0.005676906555891037, + 0.006539221853017807, + 0.024439120665192604, + 0.012853042222559452, + 0.05354919284582138, + 0.017417283728718758, + 0.01664683222770691, + 0.026917874813079834, + 0.014949733391404152, + 0.025578627362847328, + 0.018868282437324524, + 0.009863191284239292, + 0.04338838905096054, + 0.017208587378263474, + 0.011814101599156857, + 0.02633253112435341, + 0.0072141122072935104, + 0.016237422823905945, + 0.012706681154668331, + 0.012421523220837116, + 0.010624674148857594, + 0.005811741575598717, + 0.008043728768825531, + 0.011748970486223698, + 0.005332397297024727, + 0.011294379830360413, + 0.006203028839081526, + 0.0044631813652813435, + 0.010424097068607807, + 0.006415973417460918, + 0.005562033038586378, + 0.007000560872256756, + 0.00449147867038846, + 0.008161276578903198, + 0.008373027667403221, + 0.0036675487644970417, + 0.011298999190330505, + 0.005515896249562502, + 0.0051544117741286755, + 0.009246411733329296, + 0.0040511335246264935, + 0.006536979228258133, + 0.006930775474756956, + 0.007220352999866009, + 0.007221372798085213, + 0.007375937886536121, + 0.009281506761908531, + 0.007833601906895638, + 0.008530805818736553, + 0.007620091084390879, + 0.006828079465776682, + 0.0034434772096574306, + 0.013977236114442348, + 0.00604952173307538, + 0.0035245949402451515, + 0.00481213815510273, + 0.004034232348203659, + 0.008669495582580566, + 0.012640447355806828, + 0.00738947931677103, + 0.015055998228490353, + 0.013044138438999653, + 0.010857578366994858, + 0.012615258805453777, + 0.009287332184612751, + 0.008513106033205986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.001709667849354446 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a skeleton wearing a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skeleton holding a massive scythe that is taller than itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a cracked, barren, and parched landscape with dead tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse, highlighting the pale bones against a gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369222208857536 + ], + "opposite_logprob_tokens": [ + -3.4086921215057373 + ], + "logprob_reward": -0.03369222208857536, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369222208857536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06237482279539108, + 0.0320722758769989, + 0.02498393878340721, + 0.033874597400426865, + 0.5365240573883057, + 0.009522167965769768, + 0.012370513752102852, + 0.23751020431518555, + 0.02064870297908783, + 0.030118724331259727 + ], + "credit_map": [ + 0.006167241372168064, + 0.006526279728859663, + 0.006950344890356064, + 0.01089741475880146, + 0.007980727590620518, + 0.007033144123852253, + 0.005651634652167559, + 0.011168035678565502, + 0.003973896149545908, + 0.002098633674904704, + 0.008050905540585518, + 0.0038364410866051912, + 0.003311030101031065, + 0.004215204622596502, + 0.0022495496086776257, + 0.004336617421358824, + 0.0032453362364321947, + 0.0020698930602520704, + 0.006477998103946447, + 0.002586176386103034, + 0.0033844527788460255, + 0.0036408535670489073, + 0.0018230562563985586, + 0.0017561721615493298, + 0.003775696735829115, + 0.00328950397670269, + 0.004043637774884701, + 0.006409045774489641, + 0.004472218919545412, + 0.005120869260281324, + 0.0030741416849195957, + 0.003689482342451811, + 0.08693041652441025, + 0.12856200337409973, + 0.03571426868438721, + 0.04368289187550545, + 0.07160001993179321, + 0.0947878435254097, + 0.0408172681927681, + 0.034429337829351425, + 0.0011453237384557724, + 0.0008533945656381547, + 0.0014646191848441958, + 0.001175141311250627, + 0.0014922096161171794, + 0.0011357939802110195, + 0.001225665328092873, + 0.0010300202993676066, + 0.0020048802252858877, + 0.0010431907139718533, + 0.0019445347134023905, + 0.0009881971636787057, + 0.0015736661152914166, + 0.0022210353054106236, + 0.0012771418550983071, + 0.0013178682420402765, + 0.029079562053084373, + 0.03459903225302696, + 0.026171915233135223, + 0.025900445878505707, + 0.029060788452625275, + 0.03643132001161575, + 0.02151440642774105, + 0.034752752631902695, + 0.0020615807734429836, + 0.0014219904551282525, + 0.004380141384899616, + 0.004121570847928524, + 0.0014396350597962737, + 0.0015437022084370255, + 0.0017925762804225087, + 0.003887504804879427, + 0.0033122750464826822, + 0.003091837977990508, + 0.004382831044495106, + 0.005354728549718857, + 0.0030846388544887304, + 0.004096140619367361, + 0.0024053258821368217, + 0.004390944726765156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.03369222208857536 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_1/index_tensor([3928, 3928])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_1/index_tensor([3928, 3928])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..45bf374e7672be13e31ec33541bafcbe88d567ad --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_1/index_tensor([3928, 3928])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, medium-close-up shot from a level angle focuses on a man immersed in musical composition. He has short, dark brown hair and wears a pair of professional black headphones. He rhythmically and immersively nods his head up and down to the rhythmic music, his eyes slightly closed, his expression intensely focused. He is in a dimly lit recording studio, facing a complex professional mixing console, its surface densely packed with faders and illuminated buttons. His right hand's fingers move delicately and precisely across the control panel, carefully gripping a silver EQ knob between his thumb and index finger, steadily and deliberately adjusting it to find the perfect tonal balance. The only light comes from the multicolored LEDs flashing on the mixing console and the screens of the various devices in the background. These dynamic dots of light cast shifting shadows on his face, emphasizing his simple black T-shirt and his focused profile. The whole picture is based on deep dark tones, dotted with blue, red and green light emitted by the equipment, creating an intimate, passionate and creative atmosphere.", + "seed": 17714, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_1/index_3928_seed_17714_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03074890123680234, + "vlm_logprob": -0.03074890123680234, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.03074890123680234, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium-close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing professional black headphones and a simple black T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man rhythmically nodding his head with his eyes slightly closed and expression focused?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's right hand gripping a silver EQ knob between his thumb and index finger to adjust it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit only by multicolored LEDs from the mixing console and background screens, creating deep dark tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06458137929439545 + ], + "opposite_logprob_tokens": [ + -4.189581394195557 + ], + "logprob_reward": -0.06458137929439545, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.06458137929439545, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23706457018852234, + 0.10697363317012787, + 0.1295354813337326, + 0.07224484533071518, + 0.07312548905611038, + 0.06373870372772217, + 0.05506474897265434, + 0.0751713290810585, + 0.08545484393835068, + 0.10162640362977982 + ], + "credit_map": [ + 0.03922437131404877, + 0.022083448246121407, + 0.02579919807612896, + 0.033594127744436264, + 0.030507367104291916, + 0.02540780045092106, + 0.031623780727386475, + 0.028824448585510254, + 0.012936800718307495, + 0.00797253381460905, + 0.0208887942135334, + 0.011258899234235287, + 0.014299028553068638, + 0.008195530623197556, + 0.016270659863948822, + 0.01515137404203415, + 0.01641428843140602, + 0.01291376817971468, + 0.013471711426973343, + 0.01492841262370348, + 0.01703646592795849, + 0.018694475293159485, + 0.016282206401228905, + 0.01979413814842701, + 0.01070051547139883, + 0.011229629628360271, + 0.010136631317436695, + 0.0063515654765069485, + 0.00715098949149251, + 0.008663768880069256, + 0.005835231859236956, + 0.012176507152616978, + 0.010564994066953659, + 0.0110607435926795, + 0.007122962735593319, + 0.007039899006485939, + 0.010942823253571987, + 0.009458453394472599, + 0.0066989716142416, + 0.010236639529466629, + 0.006587598938494921, + 0.005224287509918213, + 0.006696347612887621, + 0.008838662877678871, + 0.008157116360962391, + 0.009311304427683353, + 0.007413122337311506, + 0.011510264128446579, + 0.007412348408252001, + 0.006616906262934208, + 0.010586750693619251, + 0.005563775543123484, + 0.006276211701333523, + 0.005822394974529743, + 0.005593497306108475, + 0.007192862220108509, + 0.01068994402885437, + 0.007882281206548214, + 0.006661257706582546, + 0.004401769023388624, + 0.009072521701455116, + 0.010717010125517845, + 0.008413299918174744, + 0.017333244904875755, + 0.008671652525663376, + 0.009335936047136784, + 0.010010365396738052, + 0.008256391622126102, + 0.01065708789974451, + 0.014707876369357109, + 0.008448504842817783, + 0.015367014333605766, + 0.01447698287665844, + 0.011607206426560879, + 0.01891148090362549, + 0.009124386124312878, + 0.010598128661513329, + 0.013847789727151394, + 0.008710716851055622, + 0.014349712990224361 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.06458137929439545 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium-close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing professional black headphones and a simple black T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man rhythmically nodding his head with his eyes slightly closed and expression focused?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's right hand gripping a silver EQ knob between his thumb and index finger to adjust it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit only by multicolored LEDs from the mixing console and background screens, creating deep dark tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012570434249937534 + ], + "opposite_logprob_tokens": [ + -4.387570381164551 + ], + "logprob_reward": -0.012570434249937534, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012570434249937534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14183653891086578, + 0.08687225729227066, + 0.06293687969446182, + 0.08650125563144684, + 0.3902989327907562, + 0.04572972655296326, + 0.04429430514574051, + 0.039854634553194046, + 0.06233655661344528, + 0.03933890163898468 + ], + "credit_map": [ + 0.01306323241442442, + 0.013589422218501568, + 0.024803876876831055, + 0.018012993037700653, + 0.01767278090119362, + 0.01904914155602455, + 0.01547845359891653, + 0.02016664668917656, + 0.01226353831589222, + 0.009297954849898815, + 0.010007760487496853, + 0.0072509171441197395, + 0.011126245371997356, + 0.014249955303966999, + 0.01155510451644659, + 0.011120781302452087, + 0.006502758711576462, + 0.0065986234694719315, + 0.004593184683471918, + 0.0052005029283463955, + 0.009529606439173222, + 0.010594693943858147, + 0.005948279518634081, + 0.013969230465590954, + 0.013900524005293846, + 0.013578740879893303, + 0.005773421376943588, + 0.009946622885763645, + 0.011798009276390076, + 0.009774199686944485, + 0.008766860701143742, + 0.012962879613041878, + 0.04942375421524048, + 0.06596238166093826, + 0.030185578390955925, + 0.03362787142395973, + 0.05392345413565636, + 0.05271681025624275, + 0.05459290370345116, + 0.04986618086695671, + 0.0069429874420166016, + 0.005630014464259148, + 0.00683118961751461, + 0.005911379586905241, + 0.0042035458609461784, + 0.006198016460984945, + 0.003550400957465172, + 0.006462190765887499, + 0.006034582853317261, + 0.005682942923158407, + 0.0070932786911726, + 0.003950919024646282, + 0.006181356962770224, + 0.005646878387778997, + 0.0031450018286705017, + 0.006559345405548811, + 0.005619872361421585, + 0.003713778220117092, + 0.005010501481592655, + 0.003970240708440542, + 0.004690802656114101, + 0.005145466886460781, + 0.0045724608935415745, + 0.007131511345505714, + 0.0054376861080527306, + 0.005673428066074848, + 0.004114925395697355, + 0.007865033112466335, + 0.006246843840926886, + 0.01094104815274477, + 0.010262561030685902, + 0.01179502997547388, + 0.004962052684277296, + 0.006106798071414232, + 0.002676506992429495, + 0.004282927140593529, + 0.0059133595786988735, + 0.006035344209522009, + 0.004374046344310045, + 0.004987867083400488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.012570434249937534 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium-close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing professional black headphones and a simple black T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man rhythmically nodding his head with his eyes slightly closed and expression focused?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's right hand gripping a silver EQ knob between his thumb and index finger to adjust it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit only by multicolored LEDs from the mixing console and background screens, creating deep dark tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0232975073158741 + ], + "opposite_logprob_tokens": [ + -3.7732975482940674 + ], + "logprob_reward": -0.0232975073158741, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0232975073158741, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21259918808937073, + 0.16716356575489044, + 0.12572309374809265, + 0.06999282538890839, + 0.10707908123731613, + 0.05775245279073715, + 0.09976332634687424, + 0.07065251469612122, + 0.05691816285252571, + 0.03235582262277603 + ], + "credit_map": [ + 0.02088213339447975, + 0.019358497112989426, + 0.03863445296883583, + 0.02701283060014248, + 0.024849364534020424, + 0.02400585263967514, + 0.03187834098935127, + 0.02597769722342491, + 0.02241593971848488, + 0.018176358193159103, + 0.01611810177564621, + 0.021769380196928978, + 0.029310142621397972, + 0.012795633636415005, + 0.02537914551794529, + 0.021198851987719536, + 0.017451399937272072, + 0.013007677160203457, + 0.014928597956895828, + 0.010275453329086304, + 0.012033035978674889, + 0.01869201846420765, + 0.015444952063262463, + 0.023889947682619095, + 0.00863250158727169, + 0.007570165675133467, + 0.00736596854403615, + 0.008416028693318367, + 0.011485381983220577, + 0.007473918609321117, + 0.007960529066622257, + 0.011088327504694462, + 0.019470546394586563, + 0.021015997976064682, + 0.010194100439548492, + 0.007683597970753908, + 0.014341214671730995, + 0.013731597922742367, + 0.009405755437910557, + 0.011236254125833511, + 0.006185502745211124, + 0.005758066661655903, + 0.004139966797083616, + 0.006103997118771076, + 0.005869023036211729, + 0.009071935899555683, + 0.006229689810425043, + 0.014394266530871391, + 0.015788407996296883, + 0.010533224791288376, + 0.011970890685915947, + 0.01018448919057846, + 0.014445565640926361, + 0.01217974629253149, + 0.013545256108045578, + 0.011115734465420246, + 0.007665933575481176, + 0.006926335860043764, + 0.005004378035664558, + 0.006701618432998657, + 0.009977198205888271, + 0.010569137521088123, + 0.007090246304869652, + 0.016717661172151566, + 0.007211057003587484, + 0.009918042458593845, + 0.0045144204050302505, + 0.00463199894875288, + 0.00758689921349287, + 0.009467474184930325, + 0.004777663387358189, + 0.008810603991150856, + 0.004625432658940554, + 0.002824074123054743, + 0.003624475561082363, + 0.0033536208793520927, + 0.004053076263517141, + 0.004746907856315374, + 0.004229455254971981, + 0.004898775368928909 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0232975073158741 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium-close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing professional black headphones and a simple black T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man rhythmically nodding his head with his eyes slightly closed and expression focused?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's right hand gripping a silver EQ knob between his thumb and index finger to adjust it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit only by multicolored LEDs from the mixing console and background screens, creating deep dark tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865477979183197 + ], + "opposite_logprob_tokens": [ + -3.048654794692993 + ], + "logprob_reward": -0.04865477979183197, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865477979183197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12283162772655487, + 0.24577949941158295, + 0.10754551738500595, + 0.08297049254179001, + 0.14443539083003998, + 0.041017260402441025, + 0.09836888313293457, + 0.06762239336967468, + 0.047320712357759476, + 0.04210824891924858 + ], + "credit_map": [ + 0.015032180584967136, + 0.013204495422542095, + 0.012425357475876808, + 0.01671161875128746, + 0.019769756123423576, + 0.013256978243589401, + 0.017460240051150322, + 0.0149709926918149, + 0.03739152103662491, + 0.022523583844304085, + 0.02146969921886921, + 0.02908642217516899, + 0.03589022904634476, + 0.024020960554480553, + 0.04622891545295715, + 0.029168155044317245, + 0.017622439190745354, + 0.010306501761078835, + 0.015447530895471573, + 0.007019110023975372, + 0.011013131588697433, + 0.01642591878771782, + 0.007603731006383896, + 0.022107137367129326, + 0.012707599438726902, + 0.01269053015857935, + 0.011217199265956879, + 0.00823948159813881, + 0.009200948290526867, + 0.008938119746744633, + 0.006618019193410873, + 0.013358582742512226, + 0.013477431610226631, + 0.012566452845931053, + 0.012025369331240654, + 0.012428530491888523, + 0.021517764776945114, + 0.02137453854084015, + 0.02925364300608635, + 0.02179165929555893, + 0.005399980582296848, + 0.004971341695636511, + 0.005166411865502596, + 0.0038762979675084352, + 0.004985758103430271, + 0.006107356399297714, + 0.004226278979331255, + 0.006283830385655165, + 0.014682356268167496, + 0.010679002851247787, + 0.018020611256361008, + 0.009104968048632145, + 0.010366734117269516, + 0.012641320005059242, + 0.0075308275409042835, + 0.015343047678470612, + 0.007095402106642723, + 0.008168649859726429, + 0.004378397949039936, + 0.005565653555095196, + 0.00939918216317892, + 0.009076219983398914, + 0.00754974689334631, + 0.016389135271310806, + 0.006602155044674873, + 0.006987063214182854, + 0.004583589266985655, + 0.002791579579934478, + 0.004178481642156839, + 0.009044497273862362, + 0.004013015888631344, + 0.009120327420532703, + 0.00728306220844388, + 0.004914223216474056, + 0.002808766206726432, + 0.0029230602085590363, + 0.0073246401734650135, + 0.008008117787539959, + 0.004312610253691673, + 0.004533764440566301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.04865477979183197 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium-close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing professional black headphones and a simple black T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man rhythmically nodding his head with his eyes slightly closed and expression focused?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's right hand gripping a silver EQ knob between his thumb and index finger to adjust it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit only by multicolored LEDs from the mixing console and background screens, creating deep dark tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004640405531972647 + ], + "opposite_logprob_tokens": [ + -5.379640579223633 + ], + "logprob_reward": -0.004640405531972647, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004640405531972647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14869052171707153, + 0.056479208171367645, + 0.05565481260418892, + 0.048641789704561234, + 0.08495485782623291, + 0.1294305920600891, + 0.03535778075456619, + 0.08868008852005005, + 0.17203804850578308, + 0.1800723373889923 + ], + "credit_map": [ + 0.016222333535552025, + 0.017959896475076675, + 0.015626316890120506, + 0.02174457535147667, + 0.017565034329891205, + 0.014917527325451374, + 0.025820568203926086, + 0.018834274262189865, + 0.009929807856678963, + 0.007750062737613916, + 0.0088348975405097, + 0.006766101345419884, + 0.005267549771815538, + 0.004697045311331749, + 0.00572989834472537, + 0.007503840606659651, + 0.008889549411833286, + 0.006735718343406916, + 0.010576515458524227, + 0.004231233615428209, + 0.006352378986775875, + 0.007919618859887123, + 0.004557163454592228, + 0.006392631679773331, + 0.0073296562768518925, + 0.00766985397785902, + 0.0036918276455253363, + 0.005862848833203316, + 0.006110094022005796, + 0.00615357793867588, + 0.005374261178076267, + 0.0064496686682105064, + 0.009659566916525364, + 0.007489167153835297, + 0.0092772813513875, + 0.007588841021060944, + 0.017062930390238762, + 0.01333708968013525, + 0.009759701788425446, + 0.010780282318592072, + 0.017416996881365776, + 0.010986440815031528, + 0.01103740744292736, + 0.01438032928854227, + 0.016740113496780396, + 0.01978950947523117, + 0.015069176442921162, + 0.02401062287390232, + 0.0049413638189435005, + 0.005958702880889177, + 0.0032626385800540447, + 0.0037484148051589727, + 0.0032068791333585978, + 0.004073513206094503, + 0.005500807426869869, + 0.004665459506213665, + 0.011864949949085712, + 0.011773237027227879, + 0.013420200906693935, + 0.007159959990531206, + 0.008407993242144585, + 0.011424393393099308, + 0.010120481252670288, + 0.014508876018226147, + 0.018286490812897682, + 0.019546985626220703, + 0.00936761125922203, + 0.019879836589097977, + 0.01811583898961544, + 0.032871048897504807, + 0.016166646033525467, + 0.037803590297698975, + 0.027402380481362343, + 0.018277626484632492, + 0.013451118022203445, + 0.026961931958794594, + 0.02050905115902424, + 0.025886178016662598, + 0.0176899041980505, + 0.029894154518842697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.004640405531972647 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium-close-up shot from a level angle focuses on a man immersed in musical composition. He has short, dark brown hair and wears a pair of professional black headphones. He rhythmically and immersively nods his head up and down to the rhythmic music, his eyes slightly closed, his expression intensely focused. He is in a dimly lit recording studio, facing a complex professional mixing console, its surface densely packed with faders and illuminated buttons. His right hand's fingers move delicately and precisely across the control panel, carefully gripping a silver EQ knob between his thumb and index finger, steadily and deliberately adjusting it to find the perfect tonal balance. The only light comes from the multicolored LEDs flashing on the mixing console and the screens of the various devices in the background. These dynamic dots of light cast shifting shadows on his face, emphasizing his simple black T-shirt and his focused profile. The whole picture is based on deep dark tones, dotted with blue, red and green light emitted by the equipment, creating an intimate, passionate and creative atmosphere.", + "seed": 17715, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_1/index_3928_seed_17715_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.030846931226551532, + "vlm_logprob": -0.030846931226551532, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.030846931226551532, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium-close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing professional black headphones and a simple black T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man rhythmically nodding his head with his eyes slightly closed and expression focused?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's right hand gripping a silver EQ knob between his thumb and index finger to adjust it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit only by multicolored LEDs from the mixing console and background screens, creating deep dark tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0423099547624588 + ], + "opposite_logprob_tokens": [ + -4.792309761047363 + ], + "logprob_reward": -0.0423099547624588, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.0423099547624588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19100187718868256, + 0.16224420070648193, + 0.11044438928365707, + 0.05178263783454895, + 0.08276952058076859, + 0.035333991050720215, + 0.04966573417186737, + 0.06478393822908401, + 0.08713307976722717, + 0.16484060883522034 + ], + "credit_map": [ + 0.03641882538795471, + 0.028949182480573654, + 0.01680978760123253, + 0.023148993030190468, + 0.01744609884917736, + 0.021050801500678062, + 0.02390363998711109, + 0.02327457070350647, + 0.03819387033581734, + 0.027969924733042717, + 0.012167275883257389, + 0.013874796219170094, + 0.015752499923110008, + 0.022366108372807503, + 0.011990451253950596, + 0.0199293065816164, + 0.02141491323709488, + 0.017162155359983444, + 0.017542604357004166, + 0.009044075384736061, + 0.01210874505341053, + 0.012483028694987297, + 0.01092915702611208, + 0.009759721346199512, + 0.007967649027705193, + 0.00878246407955885, + 0.005625500809401274, + 0.005431252531707287, + 0.004543627146631479, + 0.00930432602763176, + 0.0037803216837346554, + 0.006347499322146177, + 0.01971128210425377, + 0.014910445548593998, + 0.007012928836047649, + 0.005756840575486422, + 0.009336004965007305, + 0.008977429009974003, + 0.0058102901093661785, + 0.011254305951297283, + 0.0079939104616642, + 0.00623363396152854, + 0.003213301533833146, + 0.004674657713621855, + 0.003221504855901003, + 0.0031375957187265158, + 0.003434589831158519, + 0.0034247995354235172, + 0.008971265517175198, + 0.007959119975566864, + 0.0035970835015177727, + 0.0052747176960110664, + 0.006258971057832241, + 0.008200816810131073, + 0.004751783329993486, + 0.004651982337236404, + 0.009762941859662533, + 0.009964886121451855, + 0.005283074919134378, + 0.006082421634346247, + 0.009484670124948025, + 0.009154478088021278, + 0.00769759900867939, + 0.00735386973246932, + 0.015343927778303623, + 0.012307942844927311, + 0.012183931656181812, + 0.011927957646548748, + 0.008664767257869244, + 0.009120266884565353, + 0.007664439268410206, + 0.009919848293066025, + 0.044740933924913406, + 0.02434302307665348, + 0.01078362762928009, + 0.01409977674484253, + 0.013137629255652428, + 0.020582864060997963, + 0.017663434147834778, + 0.019489334896206856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.0423099547624588 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium-close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing professional black headphones and a simple black T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man rhythmically nodding his head with his eyes slightly closed and expression focused?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's right hand gripping a silver EQ knob between his thumb and index finger to adjust it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit only by multicolored LEDs from the mixing console and background screens, creating deep dark tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982470393180847 + ], + "opposite_logprob_tokens": [ + -3.529824733734131 + ], + "logprob_reward": -0.02982470393180847, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982470393180847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2598933279514313, + 0.15430380403995514, + 0.13036943972110748, + 0.03781575337052345, + 0.06540856510400772, + 0.10360781848430634, + 0.05245213583111763, + 0.07708102464675903, + 0.03904416039586067, + 0.08002393692731857 + ], + "credit_map": [ + 0.03381224721670151, + 0.03892875462770462, + 0.018591251224279404, + 0.025400757789611816, + 0.0307396799325943, + 0.04028122127056122, + 0.03239670395851135, + 0.039742738008499146, + 0.02892877347767353, + 0.024817967787384987, + 0.015978246927261353, + 0.012760872021317482, + 0.013048171065747738, + 0.024695390835404396, + 0.016329916194081306, + 0.01774447411298752, + 0.022428223863244057, + 0.02011014334857464, + 0.010662036947906017, + 0.013425814919173717, + 0.013908282853662968, + 0.016753874719142914, + 0.015545270405709743, + 0.017535805702209473, + 0.00874313898384571, + 0.005692888051271439, + 0.003011284861713648, + 0.0031777662225067616, + 0.003904759418219328, + 0.007146706338971853, + 0.0030311329755932093, + 0.0031080786138772964, + 0.00589896971359849, + 0.012703556567430496, + 0.003702547401189804, + 0.0056009008549153805, + 0.005297381896525621, + 0.012800613418221474, + 0.00396969448775053, + 0.015434901230037212, + 0.033857572823762894, + 0.01948287896811962, + 0.008926727809011936, + 0.0067688762210309505, + 0.009243876673281193, + 0.008421414531767368, + 0.007731251418590546, + 0.009175232611596584, + 0.014928805641829967, + 0.008887935429811478, + 0.006136253941804171, + 0.005366455763578415, + 0.0030667658429592848, + 0.005447203293442726, + 0.004501991905272007, + 0.004116724710911512, + 0.012522307224571705, + 0.010734678246080875, + 0.006618060637265444, + 0.008011418394744396, + 0.009342421777546406, + 0.011886181309819221, + 0.0076405941508710384, + 0.010325374081730843, + 0.008176838047802448, + 0.005635665263980627, + 0.0028962523210793734, + 0.0036758172791451216, + 0.005485170520842075, + 0.004820797126740217, + 0.0028874638956040144, + 0.005466159898787737, + 0.012295547872781754, + 0.010398028418421745, + 0.006258880719542503, + 0.007848074659705162, + 0.010927191004157066, + 0.012288223952054977, + 0.01260318886488676, + 0.007404808886349201 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.02982470393180847 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium-close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing professional black headphones and a simple black T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man rhythmically nodding his head with his eyes slightly closed and expression focused?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's right hand gripping a silver EQ knob between his thumb and index finger to adjust it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit only by multicolored LEDs from the mixing console and background screens, creating deep dark tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016073763370513916 + ], + "opposite_logprob_tokens": [ + -4.141073703765869 + ], + "logprob_reward": -0.016073763370513916, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016073763370513916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17082980275154114, + 0.2853512167930603, + 0.08305478096008301, + 0.0457693450152874, + 0.07955830544233322, + 0.05533401668071747, + 0.08918972313404083, + 0.043043382465839386, + 0.0651371106505394, + 0.08273237943649292 + ], + "credit_map": [ + 0.03024236671626568, + 0.028750712051987648, + 0.021053455770015717, + 0.014077863655984402, + 0.01918267458677292, + 0.023955365642905235, + 0.018697168678045273, + 0.014870194718241692, + 0.10075321793556213, + 0.05513555929064751, + 0.0295884869992733, + 0.02029123716056347, + 0.019414659589529037, + 0.021422699093818665, + 0.017026137560606003, + 0.021719185635447502, + 0.015532154589891434, + 0.011460550129413605, + 0.006218602880835533, + 0.010400787927210331, + 0.011806312017142773, + 0.011069408617913723, + 0.007573505863547325, + 0.008993461728096008, + 0.014435711316764355, + 0.008503196761012077, + 0.0026555226650089025, + 0.0025684256106615067, + 0.003574715694412589, + 0.006053496617823839, + 0.0036361725069582462, + 0.004342107102274895, + 0.011531637981534004, + 0.01199353113770485, + 0.0034732073545455933, + 0.006155794020742178, + 0.011019602417945862, + 0.012582988478243351, + 0.007676071021705866, + 0.015125473029911518, + 0.013822084292769432, + 0.008305750787258148, + 0.004866462200880051, + 0.008003181777894497, + 0.0036468650214374065, + 0.007084506563842297, + 0.00508873350918293, + 0.004516432527452707, + 0.024207476526498795, + 0.01473189890384674, + 0.006809171289205551, + 0.007639419753104448, + 0.008046247996389866, + 0.010013832710683346, + 0.008137725293636322, + 0.009603951126337051, + 0.006148544605821371, + 0.006816938519477844, + 0.0032280900049954653, + 0.00625804765149951, + 0.005873532500118017, + 0.00511360727250576, + 0.00539588974788785, + 0.004208734259009361, + 0.019005870446562767, + 0.010773098096251488, + 0.0058150929398834705, + 0.004278876818716526, + 0.005990883335471153, + 0.007017672061920166, + 0.004599093925207853, + 0.007656528148800135, + 0.022202808409929276, + 0.011614300310611725, + 0.005857942625880241, + 0.0064842491410672665, + 0.00872853398323059, + 0.011975809000432491, + 0.007130302954465151, + 0.008738433942198753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.016073763370513916 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium-close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing professional black headphones and a simple black T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man rhythmically nodding his head with his eyes slightly closed and expression focused?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's right hand gripping a silver EQ knob between his thumb and index finger to adjust it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit only by multicolored LEDs from the mixing console and background screens, creating deep dark tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495113879442215 + ], + "opposite_logprob_tokens": [ + -2.9299511909484863 + ], + "logprob_reward": -0.05495113879442215, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495113879442215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19573943316936493, + 0.08310158550739288, + 0.06930754333734512, + 0.07149672508239746, + 0.04993249848484993, + 0.08851834386587143, + 0.057517729699611664, + 0.0663987398147583, + 0.14457236230373383, + 0.17341503500938416 + ], + "credit_map": [ + 0.045550387352705, + 0.0320725291967392, + 0.01675562560558319, + 0.014565011486411095, + 0.02392849512398243, + 0.025260042399168015, + 0.01611378788948059, + 0.021493544802069664, + 0.016477402299642563, + 0.012971476651728153, + 0.005923931021243334, + 0.007654922548681498, + 0.01084987074136734, + 0.01252355519682169, + 0.008080574683845043, + 0.008619846776127815, + 0.016202827915549278, + 0.012030466459691525, + 0.007412828039377928, + 0.004211158957332373, + 0.009485025890171528, + 0.008469700813293457, + 0.005567245651036501, + 0.005928286351263523, + 0.010554122738540173, + 0.009790096431970596, + 0.008828784339129925, + 0.007268080487847328, + 0.01002540159970522, + 0.013771314173936844, + 0.005495618563145399, + 0.005763295106589794, + 0.018295779824256897, + 0.009599300101399422, + 0.003730736905708909, + 0.0037125246599316597, + 0.003291830886155367, + 0.003823102219030261, + 0.0034612773451954126, + 0.004017941188067198, + 0.02979479916393757, + 0.01547896210104227, + 0.00800422951579094, + 0.006475679110735655, + 0.006738210562616587, + 0.006293920334428549, + 0.008300703950226307, + 0.007431831210851669, + 0.011963735334575176, + 0.009072769433259964, + 0.0076148491352796555, + 0.004470297135412693, + 0.0053414045833051205, + 0.008484995923936367, + 0.005840883124619722, + 0.004728793632239103, + 0.007326080463826656, + 0.01024218276143074, + 0.005017411895096302, + 0.007306274957954884, + 0.01254489179700613, + 0.010764694772660732, + 0.0073464177548885345, + 0.0058507779613137245, + 0.040543388575315475, + 0.022531043738126755, + 0.012080846354365349, + 0.010088696144521236, + 0.014311998151242733, + 0.01659497804939747, + 0.01165526919066906, + 0.016766129061579704, + 0.05457499995827675, + 0.028271453455090523, + 0.010986397042870522, + 0.013096525333821774, + 0.011494863778352737, + 0.01839127205312252, + 0.01915338821709156, + 0.0174461267888546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.05495113879442215 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium-close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing professional black headphones and a simple black T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man rhythmically nodding his head with his eyes slightly closed and expression focused?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's right hand gripping a silver EQ knob between his thumb and index finger to adjust it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit only by multicolored LEDs from the mixing console and background screens, creating deep dark tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075095273554325 + ], + "opposite_logprob_tokens": [ + -4.511075019836426 + ], + "logprob_reward": -0.011075095273554325, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075095273554325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22038479149341583, + 0.24799294769763947, + 0.09139223396778107, + 0.05928957089781761, + 0.03629683330655098, + 0.08629553765058517, + 0.05428679287433624, + 0.08505452424287796, + 0.06973191350698471, + 0.04927487298846245 + ], + "credit_map": [ + 0.04227601736783981, + 0.03791576251387596, + 0.015467426739633083, + 0.01610039919614792, + 0.021316755563020706, + 0.026667485013604164, + 0.0327448733150959, + 0.027896059677004814, + 0.07615510374307632, + 0.04579060152173042, + 0.024025065824389458, + 0.014157891273498535, + 0.01680990494787693, + 0.027076223865151405, + 0.017730621621012688, + 0.02624753676354885, + 0.017899982631206512, + 0.01434317510575056, + 0.012749876827001572, + 0.008071497082710266, + 0.010458940640091896, + 0.010888151824474335, + 0.008180989883840084, + 0.008799616247415543, + 0.013627900741994381, + 0.010548023506999016, + 0.004918569233268499, + 0.003161683678627014, + 0.006995209958404303, + 0.008994461968541145, + 0.0043572913855314255, + 0.006686421576887369, + 0.005297010764479637, + 0.005729463882744312, + 0.003682423382997513, + 0.004164108075201511, + 0.004702586214989424, + 0.005071909632533789, + 0.0038470423314720392, + 0.003802285296842456, + 0.016039060428738594, + 0.01424553245306015, + 0.007093602791428566, + 0.011181987822055817, + 0.008376531302928925, + 0.010279993526637554, + 0.008873536251485348, + 0.010205289348959923, + 0.014315841719508171, + 0.009259606711566448, + 0.0037312456406652927, + 0.004447972401976585, + 0.007097623310983181, + 0.005945663433521986, + 0.004716070368885994, + 0.0047727650962769985, + 0.01539505086839199, + 0.014824156649410725, + 0.006969064474105835, + 0.008504527620971203, + 0.009266842156648636, + 0.012738484889268875, + 0.0072829402051866055, + 0.010073455050587654, + 0.012847374193370342, + 0.01013331301510334, + 0.0046048820950090885, + 0.0066420938819646835, + 0.00833098404109478, + 0.010299148969352245, + 0.006996949203312397, + 0.009877161122858524, + 0.011167194694280624, + 0.007412560284137726, + 0.005147058516740799, + 0.004058574791997671, + 0.004760420881211758, + 0.00607830286026001, + 0.004779663868248463, + 0.005871094763278961 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.011075095273554325 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium-close-up shot from a level angle focuses on a man immersed in musical composition. He has short, dark brown hair and wears a pair of professional black headphones. He rhythmically and immersively nods his head up and down to the rhythmic music, his eyes slightly closed, his expression intensely focused. He is in a dimly lit recording studio, facing a complex professional mixing console, its surface densely packed with faders and illuminated buttons. His right hand's fingers move delicately and precisely across the control panel, carefully gripping a silver EQ knob between his thumb and index finger, steadily and deliberately adjusting it to find the perfect tonal balance. The only light comes from the multicolored LEDs flashing on the mixing console and the screens of the various devices in the background. These dynamic dots of light cast shifting shadows on his face, emphasizing his simple black T-shirt and his focused profile. The whole picture is based on deep dark tones, dotted with blue, red and green light emitted by the equipment, creating an intimate, passionate and creative atmosphere.", + "seed": 17716, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_1/index_3928_seed_17716_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.029256658907979727, + "vlm_logprob": -0.029256658907979727, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.029256658907979727, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium-close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing professional black headphones and a simple black T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man rhythmically nodding his head with his eyes slightly closed and expression focused?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's right hand gripping a silver EQ knob between his thumb and index finger to adjust it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit only by multicolored LEDs from the mixing console and background screens, creating deep dark tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.047850534319877625 + ], + "opposite_logprob_tokens": [ + -4.297850608825684 + ], + "logprob_reward": -0.047850534319877625, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.047850534319877625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.035010337829589844, + 0.07342823594808578, + 0.037444960325956345, + 0.017569012939929962, + 0.05517234280705452, + 0.45330408215522766, + 0.033177174627780914, + 0.044083427637815475, + 0.2213822454214096, + 0.0294281467795372 + ], + "credit_map": [ + 0.0038803638890385628, + 0.0035528249572962523, + 0.005673957522958517, + 0.003713481593877077, + 0.005493019241839647, + 0.005339254159480333, + 0.002875017002224922, + 0.004482423420995474, + 0.0045101516880095005, + 0.002587180119007826, + 0.026098456233739853, + 0.00258849561214447, + 0.003942118491977453, + 0.026946816593408585, + 0.002195424400269985, + 0.0045595974661409855, + 0.0034571492578834295, + 0.005167434457689524, + 0.004721860401332378, + 0.0030841208063066006, + 0.00786413624882698, + 0.005104880314320326, + 0.0039189099334180355, + 0.004126472398638725, + 0.002037971280515194, + 0.001921566086821258, + 0.003405078547075391, + 0.0015050644287839532, + 0.0020559211261570454, + 0.002781124087050557, + 0.001889858627691865, + 0.0019724308513104916, + 0.006717059761285782, + 0.006171813700348139, + 0.011189415119588375, + 0.003674951381981373, + 0.005433439742773771, + 0.010526550002396107, + 0.006152286659926176, + 0.005306832492351532, + 0.01603849232196808, + 0.009108824655413628, + 0.19058260321617126, + 0.01515680830925703, + 0.011271989904344082, + 0.19365572929382324, + 0.007345024961978197, + 0.010144629515707493, + 0.0040046097710728645, + 0.003574765520170331, + 0.005019933916628361, + 0.005273646675050259, + 0.0037958889733999968, + 0.004256592132151127, + 0.004001155961304903, + 0.003250584937632084, + 0.004494784865528345, + 0.00695067597553134, + 0.008149325847625732, + 0.004069821909070015, + 0.005713069345802069, + 0.0068171159364283085, + 0.004054559860378504, + 0.003834074828773737, + 0.00959346815943718, + 0.00901852361857891, + 0.08283383399248123, + 0.006803299766033888, + 0.008837797679007053, + 0.0893145203590393, + 0.006002150010317564, + 0.008978673256933689, + 0.002941899234429002, + 0.0034785326570272446, + 0.004505874123424292, + 0.0036204997450113297, + 0.0024574894923716784, + 0.004665623418986797, + 0.003145788796246052, + 0.004612440709024668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 1 + ], + "selected_reward": -0.047850534319877625 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium-close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing professional black headphones and a simple black T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man rhythmically nodding his head with his eyes slightly closed and expression focused?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's right hand gripping a silver EQ knob between his thumb and index finger to adjust it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit only by multicolored LEDs from the mixing console and background screens, creating deep dark tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023304030299186707 + ], + "opposite_logprob_tokens": [ + -3.773303985595703 + ], + "logprob_reward": -0.023304030299186707, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023304030299186707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07130268961191177, + 0.230113685131073, + 0.08509577810764313, + 0.04497300833463669, + 0.037308987230062485, + 0.11047272384166718, + 0.036657463759183884, + 0.04666062071919441, + 0.2975686192512512, + 0.03984643146395683 + ], + "credit_map": [ + 0.006779173854738474, + 0.008223292417824268, + 0.011531428433954716, + 0.006228776648640633, + 0.011616874486207962, + 0.01013928186148405, + 0.005697687156498432, + 0.011086169630289078, + 0.010188787244260311, + 0.006538225803524256, + 0.08489876240491867, + 0.00722516281530261, + 0.010294673964381218, + 0.09730249643325806, + 0.005887126084417105, + 0.0077784438617527485, + 0.007966180332005024, + 0.013808616437017918, + 0.010856812819838524, + 0.0067437319085001945, + 0.012610072270035744, + 0.011708149686455727, + 0.01207021251320839, + 0.009332000277936459, + 0.00691770575940609, + 0.004745957907289267, + 0.0072393473237752914, + 0.003547822358086705, + 0.006252461113035679, + 0.007060585543513298, + 0.004556551575660706, + 0.004652578849345446, + 0.004769155289977789, + 0.003619362600147724, + 0.008261049166321754, + 0.0026086997240781784, + 0.0029973252676427364, + 0.008364136330783367, + 0.003226291621103883, + 0.003462968859821558, + 0.004904500208795071, + 0.006171941291540861, + 0.03377588465809822, + 0.005780460312962532, + 0.013627475127577782, + 0.03400382772088051, + 0.005504624452441931, + 0.0067040082067251205, + 0.0033741749357432127, + 0.004106946289539337, + 0.009993067011237144, + 0.0024894222151488066, + 0.0027064073365181684, + 0.008532922714948654, + 0.0033487523905932903, + 0.0021057697013020515, + 0.004957355558872223, + 0.006991044618189335, + 0.007068895734846592, + 0.006276553496718407, + 0.005754293408244848, + 0.006268539000302553, + 0.0038314738776534796, + 0.005512467119842768, + 0.02977333590388298, + 0.024729683995246887, + 0.08708449453115463, + 0.018104489892721176, + 0.018997395411133766, + 0.09142746031284332, + 0.012752903625369072, + 0.014698836021125317, + 0.004305179230868816, + 0.005603120196610689, + 0.006058605387806892, + 0.005305347498506308, + 0.00419008219614625, + 0.005085539538413286, + 0.004145825281739235, + 0.005152725614607334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 5 + ], + "selected_reward": -0.023304030299186707 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium-close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing professional black headphones and a simple black T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man rhythmically nodding his head with his eyes slightly closed and expression focused?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's right hand gripping a silver EQ knob between his thumb and index finger to adjust it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit only by multicolored LEDs from the mixing console and background screens, creating deep dark tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545594945549965 + ], + "opposite_logprob_tokens": [ + -4.387545585632324 + ], + "logprob_reward": -0.012545594945549965, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545594945549965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03814879432320595, + 0.07723177224397659, + 0.04201845824718475, + 0.021825239062309265, + 0.04048657417297363, + 0.44806110858917236, + 0.02140762284398079, + 0.08677122741937637, + 0.19380401074886322, + 0.030245183035731316 + ], + "credit_map": [ + 0.002593451412394643, + 0.0035005772951990366, + 0.006001729518175125, + 0.0035719822626560926, + 0.005674007348716259, + 0.007434376515448093, + 0.0034913038834929466, + 0.005881367716938257, + 0.0031591099686920643, + 0.00407427130267024, + 0.02529536746442318, + 0.0026208225172013044, + 0.0053582931868731976, + 0.029065443202853203, + 0.002532734302803874, + 0.005125734023749828, + 0.0038392189890146255, + 0.005221737548708916, + 0.007162205874919891, + 0.005005710292607546, + 0.006074805278331041, + 0.005605390295386314, + 0.0031178323552012444, + 0.005991562269628048, + 0.00196012482047081, + 0.0032771776895970106, + 0.0033026831224560738, + 0.001609943574294448, + 0.003780539846047759, + 0.003346833400428295, + 0.0014951289631426334, + 0.0030528095085173845, + 0.004019475542008877, + 0.0036504322197288275, + 0.010119738057255745, + 0.0022596099879592657, + 0.0033455498050898314, + 0.010155588388442993, + 0.003632847685366869, + 0.0033033385407179594, + 0.015949318185448647, + 0.009146253578364849, + 0.1885925829410553, + 0.014470437541604042, + 0.010339422151446342, + 0.19186510145664215, + 0.007299966644495726, + 0.01039808988571167, + 0.0026299222372472286, + 0.0020444223191589117, + 0.0044401888735592365, + 0.0020472947508096695, + 0.0025067448150366545, + 0.0030888488981872797, + 0.0014242129400372505, + 0.0032259898725897074, + 0.010628562420606613, + 0.013140561990439892, + 0.01459298562258482, + 0.0067266253754496574, + 0.012759404256939888, + 0.011223992332816124, + 0.009983946569263935, + 0.007715154904872179, + 0.009632010012865067, + 0.008916649967432022, + 0.07234092056751251, + 0.005826364737004042, + 0.006585683207958937, + 0.07698719203472137, + 0.005375843495130539, + 0.008139372803270817, + 0.003006917890161276, + 0.003780560102313757, + 0.004989916458725929, + 0.003510653506964445, + 0.002847121562808752, + 0.004561940673738718, + 0.0031310259364545345, + 0.004417050164192915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 7 + ], + "selected_reward": -0.012545594945549965 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium-close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing professional black headphones and a simple black T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man rhythmically nodding his head with his eyes slightly closed and expression focused?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's right hand gripping a silver EQ knob between his thumb and index finger to adjust it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit only by multicolored LEDs from the mixing console and background screens, creating deep dark tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054956551641225815 + ], + "opposite_logprob_tokens": [ + -2.9299564361572266 + ], + "logprob_reward": -0.054956551641225815, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054956551641225815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02070012129843235, + 0.04402437061071396, + 0.05032847449183464, + 0.01554557029157877, + 0.01784166693687439, + 0.6621657013893127, + 0.015197559259831905, + 0.021870063617825508, + 0.12104187160730362, + 0.03128461539745331 + ], + "credit_map": [ + 0.001607787678949535, + 0.0036186252254992723, + 0.0033121530432254076, + 0.0017843921668827534, + 0.0032528068404644728, + 0.0031024408526718616, + 0.0015997326700016856, + 0.0024221830535680056, + 0.0020694080740213394, + 0.002203729236498475, + 0.015011419542133808, + 0.0016968591371551156, + 0.003291582688689232, + 0.016437051817774773, + 0.0014402401866391301, + 0.001874082488939166, + 0.005326383747160435, + 0.007228884845972061, + 0.00652875704690814, + 0.003924649208784103, + 0.010828865692019463, + 0.007633200380951166, + 0.00443366589024663, + 0.0044240690767765045, + 0.0019080836791545153, + 0.0014685440110042691, + 0.0022758953273296356, + 0.0011802997905761003, + 0.002585233887657523, + 0.0025305987801402807, + 0.0016226496081799269, + 0.001974265556782484, + 0.0017513619968667626, + 0.0014948762254789472, + 0.004506303928792477, + 0.0010534032480791211, + 0.0014276717556640506, + 0.004694889299571514, + 0.0015058561693876982, + 0.0014073048951104283, + 0.024937627837061882, + 0.011868342757225037, + 0.2800041139125824, + 0.021452363580465317, + 0.015006455592811108, + 0.2836834490299225, + 0.01049033273011446, + 0.01472303457558155, + 0.002176719717681408, + 0.0011442231480032206, + 0.0032297244761139154, + 0.0009103367337957025, + 0.001875169575214386, + 0.002887145383283496, + 0.0011154728708788753, + 0.0018587673548609018, + 0.0022404540795832872, + 0.00336265005171299, + 0.003915531560778618, + 0.0017083004349842668, + 0.0035780766047537327, + 0.002747870050370693, + 0.0017938348464667797, + 0.0025233454070985317, + 0.0043520997278392315, + 0.003313856665045023, + 0.04558045044541359, + 0.0032442533411085606, + 0.006226010620594025, + 0.049031343311071396, + 0.0034938191529363394, + 0.005800035782158375, + 0.0018094496335834265, + 0.001403274480253458, + 0.008387953974306583, + 0.0020320990588515997, + 0.0022219158709049225, + 0.009440232999622822, + 0.0014445343986153603, + 0.004545156378298998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 2 + ], + "selected_reward": -0.054956551641225815 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium-close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing professional black headphones and a simple black T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man rhythmically nodding his head with his eyes slightly closed and expression focused?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's right hand gripping a silver EQ knob between his thumb and index finger to adjust it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit only by multicolored LEDs from the mixing console and background screens, creating deep dark tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007626583334058523 + ], + "opposite_logprob_tokens": [ + -4.882626533508301 + ], + "logprob_reward": -0.007626583334058523, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007626583334058523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06576711684465408, + 0.04861969128251076, + 0.07844194024801254, + 0.020612655207514763, + 0.024506913498044014, + 0.5457495450973511, + 0.03562666475772858, + 0.08602843433618546, + 0.062315307557582855, + 0.03233175724744797 + ], + "credit_map": [ + 0.009441163390874863, + 0.007350662723183632, + 0.009783787652850151, + 0.0066484990529716015, + 0.00866096280515194, + 0.009948214516043663, + 0.0060646007768809795, + 0.007869224064052105, + 0.0024482212029397488, + 0.0022673960775136948, + 0.013780968263745308, + 0.0019857008010149, + 0.00466505391523242, + 0.016597095876932144, + 0.0016345425974577665, + 0.005240715574473143, + 0.007360411342233419, + 0.010791025124490261, + 0.009318728931248188, + 0.0062512666918337345, + 0.01761718839406967, + 0.012035365216434002, + 0.008152367547154427, + 0.006915587466210127, + 0.001744946581311524, + 0.0016504040686413646, + 0.004649436101317406, + 0.00169720861595124, + 0.0027396914083510637, + 0.0037827580235898495, + 0.0019504926167428493, + 0.002397717209532857, + 0.0025682293344289064, + 0.0021632052958011627, + 0.005388174671679735, + 0.002007588744163513, + 0.002896360820159316, + 0.004910363350063562, + 0.002440819051116705, + 0.002132171532139182, + 0.019791653379797935, + 0.010468777269124985, + 0.22870968282222748, + 0.017764614894986153, + 0.016268931329250336, + 0.23111364245414734, + 0.009061780758202076, + 0.01257043145596981, + 0.0023427261039614677, + 0.0023461647797375917, + 0.007170733064413071, + 0.003608182305470109, + 0.004763963632285595, + 0.006843954790383577, + 0.003648273414000869, + 0.004902667831629515, + 0.006482166238129139, + 0.00772194005548954, + 0.01849394105374813, + 0.005009590182453394, + 0.017339617013931274, + 0.014890559017658234, + 0.007959771901369095, + 0.008130853995680809, + 0.002616102574393153, + 0.00312881194986403, + 0.022698432207107544, + 0.0022718438412994146, + 0.003365271259099245, + 0.02396533265709877, + 0.0018800058169290423, + 0.00238951132632792, + 0.0027732064481824636, + 0.0049476055428385735, + 0.004690902307629585, + 0.0034793647937476635, + 0.0021104440093040466, + 0.004360073246061802, + 0.003294898895546794, + 0.006675261072814465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 2 + ], + "selected_reward": -0.007626583334058523 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium-close-up shot from a level angle focuses on a man immersed in musical composition. He has short, dark brown hair and wears a pair of professional black headphones. He rhythmically and immersively nods his head up and down to the rhythmic music, his eyes slightly closed, his expression intensely focused. He is in a dimly lit recording studio, facing a complex professional mixing console, its surface densely packed with faders and illuminated buttons. His right hand's fingers move delicately and precisely across the control panel, carefully gripping a silver EQ knob between his thumb and index finger, steadily and deliberately adjusting it to find the perfect tonal balance. The only light comes from the multicolored LEDs flashing on the mixing console and the screens of the various devices in the background. These dynamic dots of light cast shifting shadows on his face, emphasizing his simple black T-shirt and his focused profile. The whole picture is based on deep dark tones, dotted with blue, red and green light emitted by the equipment, creating an intimate, passionate and creative atmosphere.", + "seed": 17717, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_1/index_3928_seed_17717_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03271994711831212, + "vlm_logprob": -0.03271994711831212, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.03271994711831212, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium-close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing professional black headphones and a simple black T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man rhythmically nodding his head with his eyes slightly closed and expression focused?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's right hand gripping a silver EQ knob between his thumb and index finger to adjust it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit only by multicolored LEDs from the mixing console and background screens, creating deep dark tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038491398096084595 + ], + "opposite_logprob_tokens": [ + -4.788491249084473 + ], + "logprob_reward": -0.038491398096084595, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.038491398096084595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16435924172401428, + 0.33634212613105774, + 0.07770413905382156, + 0.08428863435983658, + 0.04135556519031525, + 0.03210814669728279, + 0.0824858769774437, + 0.07525666058063507, + 0.053972382098436356, + 0.052127305418252945 + ], + "credit_map": [ + 0.025143947452306747, + 0.01574900560081005, + 0.018505634739995003, + 0.011201522313058376, + 0.02936590276658535, + 0.023857520893216133, + 0.014984183013439178, + 0.025551507249474525, + 0.05513753369450569, + 0.03247370570898056, + 0.030037185177206993, + 0.036005668342113495, + 0.03626062348484993, + 0.06909246742725372, + 0.03544028103351593, + 0.04189465939998627, + 0.00985034555196762, + 0.010719891637563705, + 0.009271073155105114, + 0.0072768256068229675, + 0.013083761557936668, + 0.010313715785741806, + 0.007380027789622545, + 0.009808494709432125, + 0.0088172797113657, + 0.0073409415781497955, + 0.004738342948257923, + 0.0056290035136044025, + 0.00973251648247242, + 0.020659271627664566, + 0.009397553279995918, + 0.017973722890019417, + 0.005712888203561306, + 0.004256079439073801, + 0.00429858174175024, + 0.003798913210630417, + 0.005230102222412825, + 0.007403065916150808, + 0.004607771057635546, + 0.006048162002116442, + 0.004264602903276682, + 0.002939020749181509, + 0.005733219441026449, + 0.00262322835624218, + 0.0034831948578357697, + 0.004682658240199089, + 0.0037537103053182364, + 0.004628506489098072, + 0.010577321983873844, + 0.009343392215669155, + 0.014535869471728802, + 0.005510896444320679, + 0.015865910798311234, + 0.010412163101136684, + 0.0072226570919156075, + 0.00901765190064907, + 0.008851537480950356, + 0.006785331293940544, + 0.006084820255637169, + 0.004963478539139032, + 0.007567852269858122, + 0.01706015318632126, + 0.005976666230708361, + 0.017966806888580322, + 0.007331974804401398, + 0.004270966164767742, + 0.005202552769333124, + 0.004003420937806368, + 0.008259925991296768, + 0.011001551523804665, + 0.00525909336283803, + 0.008642892353236675, + 0.005644708406180143, + 0.0037156487815082073, + 0.0046614957973361015, + 0.007968115620315075, + 0.009372423402965069, + 0.005713360849767923, + 0.005769846495240927, + 0.009281697683036327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.038491398096084595 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium-close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing professional black headphones and a simple black T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man rhythmically nodding his head with his eyes slightly closed and expression focused?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's right hand gripping a silver EQ knob between his thumb and index finger to adjust it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit only by multicolored LEDs from the mixing console and background screens, creating deep dark tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331637404859066 + ], + "opposite_logprob_tokens": [ + -3.7733163833618164 + ], + "logprob_reward": -0.02331637404859066, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331637404859066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11658012866973877, + 0.45558375120162964, + 0.0793469250202179, + 0.06323298066854477, + 0.054641082882881165, + 0.06873098760843277, + 0.020147103816270828, + 0.029872113838791847, + 0.10120642930269241, + 0.010658537968993187 + ], + "credit_map": [ + 0.015309213660657406, + 0.011087478138506413, + 0.010789969936013222, + 0.0162583589553833, + 0.020480260252952576, + 0.011279216967523098, + 0.013423163443803787, + 0.017952462658286095, + 0.07904978841543198, + 0.03378775343298912, + 0.03908824920654297, + 0.05305083096027374, + 0.04724931716918945, + 0.10425029695034027, + 0.037074264138936996, + 0.06203325092792511, + 0.013916420750319958, + 0.012575924396514893, + 0.005851303692907095, + 0.004366928245872259, + 0.008707963861525059, + 0.011491157114505768, + 0.007833155803382397, + 0.014604066498577595, + 0.004797180648893118, + 0.0035350483376532793, + 0.002673594281077385, + 0.003742491826415062, + 0.004573259502649307, + 0.02044486440718174, + 0.0038401067722588778, + 0.019626427441835403, + 0.007238280493766069, + 0.004714326001703739, + 0.003822032595053315, + 0.0036745413672178984, + 0.010953125543892384, + 0.008275036700069904, + 0.006250161211937666, + 0.00971357710659504, + 0.0058183385990560055, + 0.0038685393519699574, + 0.005354594439268112, + 0.011596575379371643, + 0.007831417955458164, + 0.010176782496273518, + 0.009686588309705257, + 0.014398157596588135, + 0.0033967187628149986, + 0.0027755501214414835, + 0.001844969461672008, + 0.0013472461141645908, + 0.0028065149672329426, + 0.0025737800169736147, + 0.0026016768533736467, + 0.0028006487991660833, + 0.002423793077468872, + 0.001683388720266521, + 0.0013708359329029918, + 0.000922702020034194, + 0.0017167029436677694, + 0.010594378225505352, + 0.0012537213042378426, + 0.009906590916216373, + 0.010520262643694878, + 0.015856387093663216, + 0.006263617426156998, + 0.012062648311257362, + 0.017761660739779472, + 0.009919248521327972, + 0.013117295689880848, + 0.01570531353354454, + 0.0017069770256057382, + 0.001238196506164968, + 0.0006247193668968976, + 0.001209039008244872, + 0.001310488791204989, + 0.001511304872110486, + 0.0009838321711868048, + 0.002073979936540127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.02331637404859066 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium-close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing professional black headphones and a simple black T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man rhythmically nodding his head with his eyes slightly closed and expression focused?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's right hand gripping a silver EQ knob between his thumb and index finger to adjust it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit only by multicolored LEDs from the mixing console and background screens, creating deep dark tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016074232757091522 + ], + "opposite_logprob_tokens": [ + -4.141074180603027 + ], + "logprob_reward": -0.016074232757091522, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016074232757091522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1664285808801651, + 0.2573378086090088, + 0.0628567785024643, + 0.10198108106851578, + 0.07603326439857483, + 0.06595627218484879, + 0.10133696347475052, + 0.09867841750383377, + 0.04397624358534813, + 0.02541463077068329 + ], + "credit_map": [ + 0.01947273313999176, + 0.015633760020136833, + 0.013486217707395554, + 0.025745313614606857, + 0.027515575289726257, + 0.020633947104215622, + 0.01939881406724453, + 0.02454221434891224, + 0.042992185801267624, + 0.0298428013920784, + 0.026244016364216805, + 0.03196655586361885, + 0.02276579476892948, + 0.05043693259358406, + 0.021005285903811455, + 0.03208421543240547, + 0.009422188624739647, + 0.007189523428678513, + 0.005097600165754557, + 0.007604067679494619, + 0.012177624739706516, + 0.008634095080196857, + 0.005770752672106028, + 0.006960922386497259, + 0.00893812719732523, + 0.006716277915984392, + 0.011289475485682487, + 0.006703890394419432, + 0.012238130904734135, + 0.025713399052619934, + 0.00887357909232378, + 0.021508188918232918, + 0.011433028616011143, + 0.009342662058770657, + 0.0045916251838207245, + 0.0056123328395187855, + 0.012860743328928947, + 0.01358379703015089, + 0.007778299506753683, + 0.010830767452716827, + 0.013082113116979599, + 0.008564019575715065, + 0.007602305617183447, + 0.00519745284691453, + 0.007014285307377577, + 0.010103005915880203, + 0.007130098529160023, + 0.007262987084686756, + 0.021498525515198708, + 0.02068943902850151, + 0.00852973572909832, + 0.007009970489889383, + 0.012081998400390148, + 0.014430687762796879, + 0.0066266790963709354, + 0.010469922795891762, + 0.007447586394846439, + 0.005601674783974886, + 0.0046414281241595745, + 0.0052449642680585384, + 0.007808774244040251, + 0.0303086806088686, + 0.0065948329865932465, + 0.03103047050535679, + 0.007680671755224466, + 0.0042471145279705524, + 0.003022537799552083, + 0.0027906238101422787, + 0.0048334053717553616, + 0.010721052065491676, + 0.0030728403944522142, + 0.007607993204146624, + 0.00423471862450242, + 0.00220950273796916, + 0.002699006348848343, + 0.0016562859527766705, + 0.0031298946123570204, + 0.003759772516787052, + 0.0028827604837715626, + 0.004842686001211405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.016074232757091522 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium-close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing professional black headphones and a simple black T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man rhythmically nodding his head with his eyes slightly closed and expression focused?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's right hand gripping a silver EQ knob between his thumb and index finger to adjust it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit only by multicolored LEDs from the mixing console and background screens, creating deep dark tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898272573947906 + ], + "opposite_logprob_tokens": [ + -2.5789828300476074 + ], + "logprob_reward": -0.07898272573947906, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898272573947906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30460384488105774, + 0.19876351952552795, + 0.10359945148229599, + 0.05615369975566864, + 0.03780185431241989, + 0.05768934264779091, + 0.03207835555076599, + 0.12692472338676453, + 0.04499289020895958, + 0.03739232197403908 + ], + "credit_map": [ + 0.05381157249212265, + 0.035447098314762115, + 0.02629757858812809, + 0.020769722759723663, + 0.04082619026303291, + 0.057887718081474304, + 0.022524505853652954, + 0.047039471566677094, + 0.03188939765095711, + 0.02147570438683033, + 0.020591873675584793, + 0.024726852774620056, + 0.024307377636432648, + 0.03262105584144592, + 0.017850469797849655, + 0.025300774723291397, + 0.02240706793963909, + 0.01868547312915325, + 0.00934165995568037, + 0.008423453196883202, + 0.013574606738984585, + 0.012203200720250607, + 0.008254621177911758, + 0.010709363967180252, + 0.0063368333503603935, + 0.005147443152964115, + 0.007756995502859354, + 0.00746879680082202, + 0.006537153385579586, + 0.0090516721829772, + 0.007208290509879589, + 0.006646513938903809, + 0.009413939900696278, + 0.004152955953031778, + 0.004195143468677998, + 0.0019945139065384865, + 0.003990181256085634, + 0.0059883687645196915, + 0.002446433063596487, + 0.005620317533612251, + 0.010492459870874882, + 0.004803113639354706, + 0.008037878200411797, + 0.007869144901633263, + 0.007167469244450331, + 0.007485335227102041, + 0.004528217017650604, + 0.0073057194240391254, + 0.0052322703413665295, + 0.005180192645639181, + 0.0033672323916107416, + 0.0026813128497451544, + 0.0036286257673054934, + 0.0046025472693145275, + 0.0025079892948269844, + 0.0048781828954815865, + 0.006546890828758478, + 0.005548721645027399, + 0.005098461173474789, + 0.003884725971147418, + 0.008954237215220928, + 0.045886073261499405, + 0.00613419571891427, + 0.04487140104174614, + 0.008190887048840523, + 0.007713811472058296, + 0.004610976204276085, + 0.0032774177379906178, + 0.00538856303319335, + 0.006538217421621084, + 0.0027345616836100817, + 0.006538453046232462, + 0.007097235415130854, + 0.003859123447909951, + 0.0060704671777784824, + 0.0024718723725527525, + 0.0035175823140889406, + 0.005134278442710638, + 0.0029676419217139482, + 0.006274121347814798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.07898272573947906 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium-close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing professional black headphones and a simple black T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man rhythmically nodding his head with his eyes slightly closed and expression focused?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's right hand gripping a silver EQ knob between his thumb and index finger to adjust it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit only by multicolored LEDs from the mixing console and background screens, creating deep dark tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00673500495031476 + ], + "opposite_logprob_tokens": [ + -5.006734848022461 + ], + "logprob_reward": -0.00673500495031476, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00673500495031476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2558216452598572, + 0.23131203651428223, + 0.08319630473852158, + 0.05189264193177223, + 0.13716676831245422, + 0.030577849596738815, + 0.04356219619512558, + 0.035095877945423126, + 0.1015356257557869, + 0.02983909845352173 + ], + "credit_map": [ + 0.038701944053173065, + 0.02484068274497986, + 0.025937935337424278, + 0.028216956183314323, + 0.04115258902311325, + 0.03346121311187744, + 0.02212914638221264, + 0.04138118773698807, + 0.032858360558748245, + 0.01984355039894581, + 0.0228135883808136, + 0.0278860405087471, + 0.025219794362783432, + 0.045349474996328354, + 0.026457102969288826, + 0.030884146690368652, + 0.010471614077687263, + 0.009852730669081211, + 0.0068205539137125015, + 0.008623041212558746, + 0.016052423045039177, + 0.011403226293623447, + 0.007651714142411947, + 0.012321005575358868, + 0.00850380677729845, + 0.004758512135595083, + 0.0040623098611831665, + 0.002766577061265707, + 0.005149718839675188, + 0.01079539768397808, + 0.005052412860095501, + 0.010803906247019768, + 0.028358522802591324, + 0.012436902150511742, + 0.011923507787287235, + 0.0076294285245239735, + 0.013594101183116436, + 0.02555391937494278, + 0.013236502185463905, + 0.024433888494968414, + 0.004960997495800257, + 0.0027857499662786722, + 0.002973152557387948, + 0.003256524447351694, + 0.004054426681250334, + 0.003837296273559332, + 0.0035061475355178118, + 0.00520355487242341, + 0.005959449801594019, + 0.005151437595486641, + 0.0030484022572636604, + 0.0036802846007049084, + 0.008696828037500381, + 0.005871415138244629, + 0.0033398461528122425, + 0.007814531214535236, + 0.0029070554301142693, + 0.002261785324662924, + 0.0024500112049281597, + 0.0012915469706058502, + 0.0027760411612689495, + 0.011311368085443974, + 0.0019828106742352247, + 0.010115258395671844, + 0.016967512667179108, + 0.011092285625636578, + 0.010102017782628536, + 0.010691543109714985, + 0.010170815512537956, + 0.018196873366832733, + 0.009525568224489689, + 0.014789002947509289, + 0.006414559204131365, + 0.004688603803515434, + 0.0032576879020780325, + 0.0018523698672652245, + 0.004088431131094694, + 0.004005585331469774, + 0.0023851145524531603, + 0.0031467447988688946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.00673500495031476 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_10/index_tensor([9666, 9666])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_10/index_tensor([9666, 9666])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d2ac37b96e23c436548aafc55b5802c743da7e7f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_10/index_tensor([9666, 9666])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a stunning close-up, the camera executes a slow, graceful 360-degree orbital pan around an exquisite globe, perched on an antique, dark wood base and supported by a slightly rusted brass theodolite stand. Its surface has the texture of yellowed parchment, revealing the delicate lines of a hand-drawn classical map. Place names are marked in elegant serif fonts, and the subtly raised mountain ranges are clearly visible in the light. A slender, bony hand with neatly manicured nails enters the frame. The tip of its index finger delicately and almost without pressure smoothly moves from the ancient Eurasian continent on the globe to the vast Pacific Ocean, as if tactilely tracing the threads of history and geographical change. The entire scene is illuminated by a warm, focused beam from an off-screen desk lamp, its soft light casting gentle highlights on the globe's curved surface and reflecting flecks of golden light on the brass stand. The study in the background is shrouded in deep shadow. The entire painting is dominated by warm brown, gold and cream tones, creating an atmosphere of tranquility, contemplation, wisdom and history.", + "seed": 14031, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_10/index_9666_seed_14031_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04162171250209212, + "vlm_logprob": -0.04162171250209212, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.04162171250209212, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, graceful 360-degree orbital pan around the central object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a globe with a yellowed parchment texture and hand-drawn classical map details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, bony hand with neatly manicured nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand's index finger move from the ancient Eurasian continent to the Pacific Ocean on the globe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, focused beam from an off-screen desk lamp casting highlights on the globe and brass stand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03165292739868164 + ], + "opposite_logprob_tokens": [ + -4.531652927398682 + ], + "logprob_reward": -0.03165292739868164, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.03165292739868164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07482808083295822, + 0.09387192130088806, + 0.09314768016338348, + 0.21350689232349396, + 0.06966958194971085, + 0.11814945936203003, + 0.055071618407964706, + 0.11647574603557587, + 0.108990877866745, + 0.056288108229637146 + ], + "credit_map": [ + 0.006369065493345261, + 0.010121652856469154, + 0.0068900687620043755, + 0.00800775270909071, + 0.011082259938120842, + 0.009261833503842354, + 0.0122698238119483, + 0.010825629346072674, + 0.01277187280356884, + 0.00693677831441164, + 0.009175479412078857, + 0.006677433382719755, + 0.004884487017989159, + 0.016649888828396797, + 0.008855785243213177, + 0.027920197695493698, + 0.007503956090658903, + 0.006830643862485886, + 0.009401473216712475, + 0.013210172764956951, + 0.007809335365891457, + 0.014640361070632935, + 0.015567239373922348, + 0.018184497952461243, + 0.01630926877260208, + 0.016351910308003426, + 0.021802812814712524, + 0.043516580015420914, + 0.018779022619128227, + 0.026287835091352463, + 0.04199058562517166, + 0.028468867763876915, + 0.008906126953661442, + 0.0062680356204509735, + 0.01083123404532671, + 0.009400198236107826, + 0.006435327231884003, + 0.006963412277400494, + 0.011343387886881828, + 0.00952185783535242, + 0.012760009616613388, + 0.02140996605157852, + 0.015778202563524246, + 0.014190071262419224, + 0.014108404517173767, + 0.008653325960040092, + 0.01631952077150345, + 0.014929957687854767, + 0.006829245947301388, + 0.008446583524346352, + 0.007547011598944664, + 0.009275468066334724, + 0.0054769739508628845, + 0.005276109557598829, + 0.006205727346241474, + 0.0060144951567053795, + 0.011785050854086876, + 0.007390995509922504, + 0.017895186319947243, + 0.01943095400929451, + 0.005816388409584761, + 0.009564254432916641, + 0.013773617334663868, + 0.030819298699498177, + 0.008532244712114334, + 0.01043388620018959, + 0.019238537177443504, + 0.02979048155248165, + 0.00626160716637969, + 0.0083682332187891, + 0.011643871665000916, + 0.014722014777362347, + 0.003998870961368084, + 0.00821469072252512, + 0.008348033763468266, + 0.007094471249729395, + 0.005811875686049461, + 0.005886520259082317, + 0.011610224843025208, + 0.005323418881744146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 7 + ], + "selected_reward": -0.03165292739868164 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, graceful 360-degree orbital pan around the central object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a globe with a yellowed parchment texture and hand-drawn classical map details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, bony hand with neatly manicured nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand's index finger move from the ancient Eurasian continent to the Pacific Ocean on the globe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, focused beam from an off-screen desk lamp casting highlights on the globe and brass stand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002509898506104946 + ], + "opposite_logprob_tokens": [ + -6.002510070800781 + ], + "logprob_reward": -0.002509898506104946, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002509898506104946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1320154368877411, + 0.10145734995603561, + 0.059469208121299744, + 0.23734869062900543, + 0.12739765644073486, + 0.04776083305478096, + 0.03525369241833687, + 0.14729823172092438, + 0.05765765905380249, + 0.054341282695531845 + ], + "credit_map": [ + 0.018860317766666412, + 0.005439602304250002, + 0.03482281416654587, + 0.01126772165298462, + 0.006302558816969395, + 0.020082080736756325, + 0.017427124083042145, + 0.01781321130692959, + 0.018357012420892715, + 0.006406466476619244, + 0.01006313506513834, + 0.009095892310142517, + 0.006350577808916569, + 0.0162233654409647, + 0.011339295655488968, + 0.02362159639596939, + 0.00552112003788352, + 0.004746230319142342, + 0.005715470295399427, + 0.007482362911105156, + 0.004275860730558634, + 0.010221373289823532, + 0.009597212076187134, + 0.011909574270248413, + 0.018602818250656128, + 0.018357450142502785, + 0.020566729828715324, + 0.052861254662275314, + 0.022402653470635414, + 0.018757320940494537, + 0.050964586436748505, + 0.03483587130904198, + 0.017766660079360008, + 0.008245233446359634, + 0.02606205642223358, + 0.01849224977195263, + 0.012929849326610565, + 0.011423242278397083, + 0.0192590169608593, + 0.013219340704381466, + 0.004883057437837124, + 0.00716398237273097, + 0.00585453724488616, + 0.005760862957686186, + 0.004918479360640049, + 0.005099899135529995, + 0.007759817410260439, + 0.00632020179182291, + 0.004129492677748203, + 0.003674808656796813, + 0.006615840829908848, + 0.004972703289240599, + 0.0032391257118433714, + 0.0033309240825474262, + 0.005289251916110516, + 0.004001548048108816, + 0.017067277804017067, + 0.012304622679948807, + 0.022128643468022346, + 0.02512325905263424, + 0.011981046758592129, + 0.0114230215549469, + 0.018995869904756546, + 0.028274493291974068, + 0.005514662247151136, + 0.006547793745994568, + 0.0066891866736114025, + 0.009030251763761044, + 0.005124503280967474, + 0.008124989457428455, + 0.009156147949397564, + 0.007470127195119858, + 0.003933884669095278, + 0.008256586268544197, + 0.008748560212552547, + 0.00963893998414278, + 0.004976022522896528, + 0.004660506267100573, + 0.007107630372047424, + 0.007019150536507368 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.002509898506104946 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, graceful 360-degree orbital pan around the central object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a globe with a yellowed parchment texture and hand-drawn classical map details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, bony hand with neatly manicured nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand's index finger move from the ancient Eurasian continent to the Pacific Ocean on the globe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, focused beam from an off-screen desk lamp casting highlights on the globe and brass stand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002092897891998 + ], + "opposite_logprob_tokens": [ + -2.695020914077759 + ], + "logprob_reward": -0.07002092897891998, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002092897891998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08175311982631683, + 0.04920472949743271, + 0.08699101209640503, + 0.48083269596099854, + 0.07615788280963898, + 0.07964440435171127, + 0.01812245324254036, + 0.037691861391067505, + 0.06881267577409744, + 0.020789194852113724 + ], + "credit_map": [ + 0.011087628081440926, + 0.006001492030918598, + 0.019337212666869164, + 0.007304089143872261, + 0.006168716121464968, + 0.011408699676394463, + 0.009840059094130993, + 0.010605226270854473, + 0.007756427861750126, + 0.003119960892945528, + 0.0058263312093913555, + 0.004396711476147175, + 0.004572222009301186, + 0.00726614985615015, + 0.006099311634898186, + 0.010167611762881279, + 0.005899391602724791, + 0.004218701273202896, + 0.0070322491228580475, + 0.007871772162616253, + 0.0060231052339077, + 0.019283318892121315, + 0.01099551934748888, + 0.02566695213317871, + 0.03639663755893707, + 0.03377506881952286, + 0.04263490065932274, + 0.106957346200943, + 0.042790647596120834, + 0.044888343662023544, + 0.1050153374671936, + 0.06837441772222519, + 0.009541521780192852, + 0.0037503542844206095, + 0.013618558645248413, + 0.011876106262207031, + 0.004694712348282337, + 0.007467708084732294, + 0.013341546058654785, + 0.011867373250424862, + 0.008161352016031742, + 0.014966612681746483, + 0.009943533688783646, + 0.009969203732907772, + 0.009278669022023678, + 0.0064444150775671005, + 0.010583139024674892, + 0.010297473520040512, + 0.0029557410161942244, + 0.0014459782978519797, + 0.0055338069796562195, + 0.0021210124250501394, + 0.001571343163959682, + 0.001262991689145565, + 0.001720626256428659, + 0.0015109536470845342, + 0.003804609412327409, + 0.0017747769597917795, + 0.00577991409227252, + 0.004853499587625265, + 0.003274235874414444, + 0.0030326510313898325, + 0.004729640204459429, + 0.01044253446161747, + 0.004948990885168314, + 0.006102696526795626, + 0.012567474506795406, + 0.019498556852340698, + 0.004156424198299646, + 0.005388977471739054, + 0.008295615203678608, + 0.007853934541344643, + 0.0016237576492130756, + 0.004132708068937063, + 0.002971910871565342, + 0.0036455641966313124, + 0.0021318187937140465, + 0.0015987853985279799, + 0.002777040470391512, + 0.0019076099852100015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.07002092897891998 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, graceful 360-degree orbital pan around the central object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a globe with a yellowed parchment texture and hand-drawn classical map details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, bony hand with neatly manicured nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand's index finger move from the ancient Eurasian continent to the Pacific Ocean on the globe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, focused beam from an off-screen desk lamp casting highlights on the globe and brass stand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031107813119888 + ], + "opposite_logprob_tokens": [ + -2.350311040878296 + ], + "logprob_reward": -0.10031107813119888, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031107813119888, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08658690005540848, + 0.056951895356178284, + 0.05082671344280243, + 0.3760831356048584, + 0.08673270046710968, + 0.12351108342409134, + 0.025352220982313156, + 0.07489752769470215, + 0.0827905610203743, + 0.03626728802919388 + ], + "credit_map": [ + 0.010534980334341526, + 0.007844257168471813, + 0.010499617084860802, + 0.010882370173931122, + 0.009585211053490639, + 0.01050695963203907, + 0.013826372101902962, + 0.01290713157504797, + 0.008514280430972576, + 0.003648781683295965, + 0.006709750276058912, + 0.005770855583250523, + 0.0056828162632882595, + 0.007821664214134216, + 0.008690203540027142, + 0.010113544762134552, + 0.0057378546334803104, + 0.0031689819879829884, + 0.006621843669563532, + 0.007167715113610029, + 0.003933822736144066, + 0.007082208059728146, + 0.009390424005687237, + 0.007723863236606121, + 0.0317363440990448, + 0.025149686262011528, + 0.033286791294813156, + 0.0812375470995903, + 0.03466863930225372, + 0.03938509523868561, + 0.07979986071586609, + 0.050819121301174164, + 0.014264223165810108, + 0.004411006346344948, + 0.019221879541873932, + 0.010050534270703793, + 0.0050836848095059395, + 0.008913139812648296, + 0.010860536247491837, + 0.013927701860666275, + 0.013544457964599133, + 0.023765120655298233, + 0.017554592341184616, + 0.014506754465401173, + 0.015219437889754772, + 0.00884896982461214, + 0.016233893111348152, + 0.013837860897183418, + 0.0043265013955533504, + 0.002300271764397621, + 0.007076319772750139, + 0.003162125125527382, + 0.0018054773099720478, + 0.0020659028086811304, + 0.002416083822026849, + 0.002199536422267556, + 0.007434292230755091, + 0.004280955996364355, + 0.011233288794755936, + 0.011289620772004128, + 0.006120924372226, + 0.006461977027356625, + 0.008792314678430557, + 0.01928415149450302, + 0.0059647224843502045, + 0.007983170449733734, + 0.015457934699952602, + 0.02356707490980625, + 0.005108407698571682, + 0.0061088246293365955, + 0.00878786202520132, + 0.009812567383050919, + 0.005262533202767372, + 0.007392836268991232, + 0.005074947606772184, + 0.004561515524983406, + 0.003737624501809478, + 0.0031684725545346737, + 0.0034944505896419287, + 0.003574908012524247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.10031107813119888 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, graceful 360-degree orbital pan around the central object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a globe with a yellowed parchment texture and hand-drawn classical map details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, bony hand with neatly manicured nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand's index finger move from the ancient Eurasian continent to the Pacific Ocean on the globe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, focused beam from an off-screen desk lamp casting highlights on the globe and brass stand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036137294955551624 + ], + "opposite_logprob_tokens": [ + -5.6286139488220215 + ], + "logprob_reward": -0.0036137294955551624, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036137294955551624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1495853215456009, + 0.15299643576145172, + 0.25213903188705444, + 0.06276653707027435, + 0.04380961135029793, + 0.05560148507356644, + 0.02167150191962719, + 0.08541605621576309, + 0.11155412346124649, + 0.06445988267660141 + ], + "credit_map": [ + 0.015549005940556526, + 0.01194009743630886, + 0.029876425862312317, + 0.01128995418548584, + 0.013439251109957695, + 0.02212020941078663, + 0.026402555406093597, + 0.018967820331454277, + 0.027881577610969543, + 0.008715140633285046, + 0.018981974571943283, + 0.013774317689239979, + 0.00723962252959609, + 0.0244913212954998, + 0.015679024159908295, + 0.03623345494270325, + 0.017915982753038406, + 0.014662948437035084, + 0.027650725096464157, + 0.02523268759250641, + 0.014221244491636753, + 0.05290834605693817, + 0.03273926302790642, + 0.0668078288435936, + 0.006076673977077007, + 0.0057640112936496735, + 0.007859991863369942, + 0.00765121029689908, + 0.007372539024800062, + 0.009160131216049194, + 0.007641252595931292, + 0.011240723542869091, + 0.006270460318773985, + 0.0031671642791479826, + 0.006986699998378754, + 0.005833637434989214, + 0.004630194045603275, + 0.004550036042928696, + 0.0051044016145169735, + 0.007267019711434841, + 0.006322027649730444, + 0.011528849601745605, + 0.008498850278556347, + 0.006234552711248398, + 0.006127483211457729, + 0.002960376674309373, + 0.008251133374869823, + 0.00567821366712451, + 0.003729665419086814, + 0.002663934137672186, + 0.00477689178660512, + 0.002547329757362604, + 0.0019039814360439777, + 0.0013535914476960897, + 0.0024136705324053764, + 0.0022824364714324474, + 0.010882635600864887, + 0.007717343978583813, + 0.012819383293390274, + 0.013032021000981331, + 0.009134279564023018, + 0.007155485916882753, + 0.012862827628850937, + 0.011812079697847366, + 0.008175553753972054, + 0.011165592819452286, + 0.01943870447576046, + 0.029192719608545303, + 0.008287321776151657, + 0.009483502246439457, + 0.013461611233651638, + 0.012349113821983337, + 0.00884187687188387, + 0.010881653986871243, + 0.010660444386303425, + 0.013498774729669094, + 0.005556010641157627, + 0.004041736014187336, + 0.0052418033592402935, + 0.005737584549933672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.0036137294955551624 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up, the camera executes a slow, graceful 360-degree orbital pan around an exquisite globe, perched on an antique, dark wood base and supported by a slightly rusted brass theodolite stand. Its surface has the texture of yellowed parchment, revealing the delicate lines of a hand-drawn classical map. Place names are marked in elegant serif fonts, and the subtly raised mountain ranges are clearly visible in the light. A slender, bony hand with neatly manicured nails enters the frame. The tip of its index finger delicately and almost without pressure smoothly moves from the ancient Eurasian continent on the globe to the vast Pacific Ocean, as if tactilely tracing the threads of history and geographical change. The entire scene is illuminated by a warm, focused beam from an off-screen desk lamp, its soft light casting gentle highlights on the globe's curved surface and reflecting flecks of golden light on the brass stand. The study in the background is shrouded in deep shadow. The entire painting is dominated by warm brown, gold and cream tones, creating an atmosphere of tranquility, contemplation, wisdom and history.", + "seed": 14032, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_10/index_9666_seed_14032_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04054766264744103, + "vlm_logprob": -0.04054766264744103, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.04054766264744103, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, graceful 360-degree orbital pan around the central object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a globe with a yellowed parchment texture and hand-drawn classical map details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, bony hand with neatly manicured nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand's index finger move from the ancient Eurasian continent to the Pacific Ocean on the globe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, focused beam from an off-screen desk lamp casting highlights on the globe and brass stand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033079542219638824 + ], + "opposite_logprob_tokens": [ + -5.033079624176025 + ], + "logprob_reward": -0.033079542219638824, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.033079542219638824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19470055401325226, + 0.05413300544023514, + 0.0742388442158699, + 0.08445340394973755, + 0.21130216121673584, + 0.07357658445835114, + 0.11898021399974823, + 0.058898139744997025, + 0.07464373111724854, + 0.05507328361272812 + ], + "credit_map": [ + 0.01936926878988743, + 0.03181413188576698, + 0.02401489019393921, + 0.026856673881411552, + 0.027562137693166733, + 0.021872134879231453, + 0.01713470183312893, + 0.026076627895236015, + 0.005428472999483347, + 0.009788447991013527, + 0.007128667086362839, + 0.00481990072876215, + 0.004201702773571014, + 0.0111076096072793, + 0.004421367775648832, + 0.007236838806420565, + 0.009461450390517712, + 0.008458133786916733, + 0.00872495211660862, + 0.006795215420424938, + 0.006124075036495924, + 0.01105837244540453, + 0.00556855509057641, + 0.01804809644818306, + 0.01025183405727148, + 0.009168545715510845, + 0.009961605072021484, + 0.009596328251063824, + 0.013250311836600304, + 0.012113701552152634, + 0.009715710766613483, + 0.010395379737019539, + 0.051844172179698944, + 0.014018288813531399, + 0.04035529866814613, + 0.013841601088643074, + 0.011984026059508324, + 0.02128729410469532, + 0.01058398000895977, + 0.04738751798868179, + 0.007473335601389408, + 0.007165438495576382, + 0.009058520197868347, + 0.01156102865934372, + 0.009381738491356373, + 0.008520121686160564, + 0.008235624991357327, + 0.01218077726662159, + 0.014361185021698475, + 0.009734030812978745, + 0.01621071621775627, + 0.017764950171113014, + 0.01577354036271572, + 0.010114986449480057, + 0.010621662251651287, + 0.024399150162935257, + 0.006294212304055691, + 0.00658596633002162, + 0.007224847562611103, + 0.00993188563734293, + 0.008593039587140083, + 0.005229652393609285, + 0.009829053655266762, + 0.0052094850689172745, + 0.016218459233641624, + 0.007862832397222519, + 0.01176913920789957, + 0.009378437884151936, + 0.0077605415135622025, + 0.005902384407818317, + 0.007371817249804735, + 0.008380129002034664, + 0.012496808543801308, + 0.0033278781920671463, + 0.007878272794187069, + 0.004564519505947828, + 0.005625903140753508, + 0.004507229197770357, + 0.009366900660097599, + 0.00730577576905489 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.033079542219638824 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, graceful 360-degree orbital pan around the central object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a globe with a yellowed parchment texture and hand-drawn classical map details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, bony hand with neatly manicured nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand's index finger move from the ancient Eurasian continent to the Pacific Ocean on the globe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, focused beam from an off-screen desk lamp casting highlights on the globe and brass stand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032144326251000166 + ], + "opposite_logprob_tokens": [ + -5.753214359283447 + ], + "logprob_reward": -0.0032144326251000166, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032144326251000166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06831379979848862, + 0.07213154435157776, + 0.18311813473701477, + 0.10206017643213272, + 0.07912333309650421, + 0.28386661410331726, + 0.11231136322021484, + 0.024764534085989, + 0.042494338005781174, + 0.03181616961956024 + ], + "credit_map": [ + 0.009497860446572304, + 0.009614725597202778, + 0.010631531476974487, + 0.005619285628199577, + 0.007218837738037109, + 0.008336266502737999, + 0.004441465251147747, + 0.012953825294971466, + 0.006438245996832848, + 0.010448492132127285, + 0.007657404523342848, + 0.009807019494473934, + 0.006712952628731728, + 0.013571965508162975, + 0.007520592305809259, + 0.009974868968129158, + 0.01891236938536167, + 0.023040546104311943, + 0.026436518877744675, + 0.01692311465740204, + 0.02164416015148163, + 0.03630037233233452, + 0.016043145209550858, + 0.023817909881472588, + 0.010502761229872704, + 0.008567905984818935, + 0.014906377531588078, + 0.00872765015810728, + 0.010963264852762222, + 0.025544559583067894, + 0.011819729581475258, + 0.011027919128537178, + 0.009072656743228436, + 0.006923009641468525, + 0.008978218771517277, + 0.011263178661465645, + 0.007519708946347237, + 0.0065941233187913895, + 0.014406846836209297, + 0.014365589246153831, + 0.023898478597402573, + 0.03912420570850372, + 0.03820671886205673, + 0.05886406451463699, + 0.029166748747229576, + 0.029978277161717415, + 0.029917916283011436, + 0.034710198640823364, + 0.013290314003825188, + 0.006277918349951506, + 0.015549900010228157, + 0.013551630079746246, + 0.014458385296165943, + 0.009232793003320694, + 0.01149809267371893, + 0.02845231629908085, + 0.0033328093122690916, + 0.002364509040489793, + 0.0035144658759236336, + 0.003935652319341898, + 0.0019609699957072735, + 0.0030114392284303904, + 0.0027150153182446957, + 0.003929671831429005, + 0.0069654714316129684, + 0.004529297351837158, + 0.005806049332022667, + 0.005326555576175451, + 0.004463555291295052, + 0.0029983411077409983, + 0.005107505712658167, + 0.007297561503946781, + 0.006332539953291416, + 0.0025897088926285505, + 0.003944212570786476, + 0.0032708949875086546, + 0.003741784719750285, + 0.0037277364172041416, + 0.004748511128127575, + 0.0034607776906341314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 6 + ], + "selected_reward": -0.0032144326251000166 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, graceful 360-degree orbital pan around the central object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a globe with a yellowed parchment texture and hand-drawn classical map details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, bony hand with neatly manicured nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand's index finger move from the ancient Eurasian continent to the Pacific Ocean on the globe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, focused beam from an off-screen desk lamp casting highlights on the globe and brass stand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204279512166977 + ], + "opposite_logprob_tokens": [ + -2.812042713165283 + ], + "logprob_reward": -0.06204279512166977, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204279512166977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03665895387530327, + 0.07025262713432312, + 0.10969525575637817, + 0.07152833044528961, + 0.07624763250350952, + 0.46292078495025635, + 0.03600873798131943, + 0.0321149080991745, + 0.06891237199306488, + 0.03566042333841324 + ], + "credit_map": [ + 0.0042772418819367886, + 0.006504965014755726, + 0.004902421031147242, + 0.004130473826080561, + 0.00438260892406106, + 0.0049255043268203735, + 0.0024601148907095194, + 0.005075622349977493, + 0.00690903514623642, + 0.011424676515161991, + 0.006380230188369751, + 0.007410040590912104, + 0.00993408914655447, + 0.010790382511913776, + 0.008093934506177902, + 0.009310231544077396, + 0.010936438106000423, + 0.011954205110669136, + 0.015313242562115192, + 0.0079732034355402, + 0.009954093024134636, + 0.02868575043976307, + 0.0070336926728487015, + 0.01784461736679077, + 0.008977366611361504, + 0.006629575043916702, + 0.006887713447213173, + 0.006135648116469383, + 0.008847341872751713, + 0.013181553222239017, + 0.009565899148583412, + 0.011303228326141834, + 0.01448275987058878, + 0.008815583772957325, + 0.012970678508281708, + 0.00963293481618166, + 0.008289576508104801, + 0.007574307266622782, + 0.007141692563891411, + 0.007340090349316597, + 0.04249115660786629, + 0.06692072749137878, + 0.06410673260688782, + 0.09203337132930756, + 0.04301686957478523, + 0.04819666966795921, + 0.045052360743284225, + 0.06110287457704544, + 0.004489200189709663, + 0.002873581601306796, + 0.0048439051024615765, + 0.00549398735165596, + 0.00485334312543273, + 0.0032415930181741714, + 0.004237369634211063, + 0.005975754465907812, + 0.0039396993815898895, + 0.0022462753113359213, + 0.005081998649984598, + 0.0046966965310275555, + 0.003555882256478071, + 0.004238126799464226, + 0.004007063340395689, + 0.004349164664745331, + 0.010663453489542007, + 0.009304073639214039, + 0.009422157891094685, + 0.00887000560760498, + 0.009993392042815685, + 0.004468667786568403, + 0.00822650920599699, + 0.007964110933244228, + 0.005089105572551489, + 0.0023276060819625854, + 0.004908463452011347, + 0.0029930712189525366, + 0.005870187189429998, + 0.0034230004530400038, + 0.007558076176792383, + 0.003490908769890666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 4 + ], + "selected_reward": -0.06204279512166977 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, graceful 360-degree orbital pan around the central object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a globe with a yellowed parchment texture and hand-drawn classical map details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, bony hand with neatly manicured nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand's index finger move from the ancient Eurasian continent to the Pacific Ocean on the globe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, focused beam from an off-screen desk lamp casting highlights on the globe and brass stand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030806064605713 + ], + "opposite_logprob_tokens": [ + -2.3503079414367676 + ], + "logprob_reward": -0.10030806064605713, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030806064605713, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1455678939819336, + 0.041252728551626205, + 0.4066970944404602, + 0.060763757675886154, + 0.028061602264642715, + 0.10922621190547943, + 0.09366212785243988, + 0.02759900875389576, + 0.026363590732216835, + 0.06080600246787071 + ], + "credit_map": [ + 0.01612711511552334, + 0.024997040629386902, + 0.02418811433017254, + 0.01471073180437088, + 0.018328523263335228, + 0.015813112258911133, + 0.008363064378499985, + 0.02304019220173359, + 0.0045700776390731335, + 0.0050638956017792225, + 0.004165404476225376, + 0.004551163874566555, + 0.004449492786079645, + 0.008117461577057838, + 0.0051843938417732716, + 0.005150838755071163, + 0.0634671002626419, + 0.02519526146352291, + 0.04728349670767784, + 0.021649328991770744, + 0.02208566665649414, + 0.0754694789648056, + 0.02346845716238022, + 0.12807831168174744, + 0.011352889239788055, + 0.005335533060133457, + 0.008106490597128868, + 0.005227024666965008, + 0.006193716078996658, + 0.008156943134963512, + 0.007660136558115482, + 0.008731024339795113, + 0.00438679801300168, + 0.003473063465207815, + 0.004570323973894119, + 0.0035702541936188936, + 0.003716845763847232, + 0.0027240717317909002, + 0.002936151111498475, + 0.0026840916834771633, + 0.010801789350807667, + 0.01282504666596651, + 0.015350704081356525, + 0.027939653024077415, + 0.010513206012547016, + 0.010633138008415699, + 0.011469999328255653, + 0.009692682884633541, + 0.007274509873241186, + 0.0099483048543334, + 0.01238289475440979, + 0.008991934359073639, + 0.018573200330138206, + 0.01024724543094635, + 0.010779894888401031, + 0.015464145690202713, + 0.0028396074194461107, + 0.0020254580304026604, + 0.0038590722251683474, + 0.004008420277386904, + 0.00461824145168066, + 0.0031949246767908335, + 0.003636913141235709, + 0.0034163722302764654, + 0.003573360852897167, + 0.002917621284723282, + 0.0035073321778327227, + 0.002901360858231783, + 0.0032753534615039825, + 0.002508537145331502, + 0.005015264730900526, + 0.002664759289473295, + 0.013343751430511475, + 0.003761167172342539, + 0.008801435120403767, + 0.005161702632904053, + 0.0065822056494653225, + 0.006366363726556301, + 0.009528306312859058, + 0.007261068094521761 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.10030806064605713 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, graceful 360-degree orbital pan around the central object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a globe with a yellowed parchment texture and hand-drawn classical map details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, bony hand with neatly manicured nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand's index finger move from the ancient Eurasian continent to the Pacific Ocean on the globe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, focused beam from an off-screen desk lamp casting highlights on the globe and brass stand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040934826247394085 + ], + "opposite_logprob_tokens": [ + -5.504093647003174 + ], + "logprob_reward": -0.0040934826247394085, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040934826247394085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15552973747253418, + 0.04435299336910248, + 0.20527935028076172, + 0.039371587336063385, + 0.04130716249346733, + 0.05041441693902016, + 0.08856452256441116, + 0.062252484261989594, + 0.284768283367157, + 0.02815949358046055 + ], + "credit_map": [ + 0.030220257118344307, + 0.01665961928665638, + 0.023174308240413666, + 0.013215459883213043, + 0.012661210261285305, + 0.016888273879885674, + 0.008855453692376614, + 0.03385515883564949, + 0.006300603970885277, + 0.005127367563545704, + 0.007266263477504253, + 0.005038167349994183, + 0.0035507106222212315, + 0.007128877565264702, + 0.0030116296838968992, + 0.006929374299943447, + 0.031215958297252655, + 0.01390556525439024, + 0.024052206426858902, + 0.010532562620937824, + 0.011499777436256409, + 0.03643603250384331, + 0.012997258454561234, + 0.06463999301195145, + 0.007441243622452021, + 0.003767159068956971, + 0.005881349556148052, + 0.0029117667581886053, + 0.003954427782446146, + 0.006855514366179705, + 0.0027218449395149946, + 0.005838281475007534, + 0.0049434262327849865, + 0.0030475121457129717, + 0.00578672019764781, + 0.005793879274278879, + 0.003859581658616662, + 0.00411687046289444, + 0.004231199622154236, + 0.00952797383069992, + 0.00656314380466938, + 0.006308785639703274, + 0.007596455514431, + 0.009765502996742725, + 0.00443101953715086, + 0.004372151102870703, + 0.0048354570753872395, + 0.00654189707711339, + 0.015267734415829182, + 0.006077880505472422, + 0.015755943953990936, + 0.01298435777425766, + 0.005326692946255207, + 0.008303314447402954, + 0.008230241015553474, + 0.01661835052073002, + 0.011947756633162498, + 0.0026679132133722305, + 0.010943762958049774, + 0.005775380879640579, + 0.003527448745444417, + 0.008553672581911087, + 0.004428845830261707, + 0.014407706446945667, + 0.0529356375336647, + 0.03056943230330944, + 0.04087884724140167, + 0.03564544767141342, + 0.021635904908180237, + 0.024736681953072548, + 0.02489262819290161, + 0.05347370356321335, + 0.007334752939641476, + 0.002204756485298276, + 0.003955120220780373, + 0.0024047887418419123, + 0.0030051132198423147, + 0.0025256711523979902, + 0.0031930822879076004, + 0.003536208998411894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.0040934826247394085 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up, the camera executes a slow, graceful 360-degree orbital pan around an exquisite globe, perched on an antique, dark wood base and supported by a slightly rusted brass theodolite stand. Its surface has the texture of yellowed parchment, revealing the delicate lines of a hand-drawn classical map. Place names are marked in elegant serif fonts, and the subtly raised mountain ranges are clearly visible in the light. A slender, bony hand with neatly manicured nails enters the frame. The tip of its index finger delicately and almost without pressure smoothly moves from the ancient Eurasian continent on the globe to the vast Pacific Ocean, as if tactilely tracing the threads of history and geographical change. The entire scene is illuminated by a warm, focused beam from an off-screen desk lamp, its soft light casting gentle highlights on the globe's curved surface and reflecting flecks of golden light on the brass stand. The study in the background is shrouded in deep shadow. The entire painting is dominated by warm brown, gold and cream tones, creating an atmosphere of tranquility, contemplation, wisdom and history.", + "seed": 14033, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_10/index_9666_seed_14033_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.104546883655712, + "vlm_logprob": -0.104546883655712, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.104546883655712, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, graceful 360-degree orbital pan around the central object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a globe with a yellowed parchment texture and hand-drawn classical map details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, bony hand with neatly manicured nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand's index finger move from the ancient Eurasian continent to the Pacific Ocean on the globe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, focused beam from an off-screen desk lamp casting highlights on the globe and brass stand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08806448429822922 + ], + "opposite_logprob_tokens": [ + -2.713064432144165 + ], + "logprob_reward": -0.08806448429822922, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.08806448429822922, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1248561292886734, + 0.05463813990354538, + 0.05466187000274658, + 0.22390034794807434, + 0.06139915809035301, + 0.05703011900186539, + 0.09263773262500763, + 0.05684017017483711, + 0.1942242681980133, + 0.07981204986572266 + ], + "credit_map": [ + 0.02879861742258072, + 0.013467058539390564, + 0.03262599557638168, + 0.014061029069125652, + 0.006753657944500446, + 0.011247779242694378, + 0.01004903856664896, + 0.007852944545447826, + 0.00905510876327753, + 0.007519785314798355, + 0.009086103178560734, + 0.005640074610710144, + 0.004757319577038288, + 0.005062321200966835, + 0.009128402918577194, + 0.004389022942632437, + 0.00886355061084032, + 0.007189132738858461, + 0.007933401502668858, + 0.005438861437141895, + 0.005240144673734903, + 0.005539161153137684, + 0.005261571146547794, + 0.009196046739816666, + 0.030677050352096558, + 0.03316422179341316, + 0.03574727475643158, + 0.02483438327908516, + 0.028082089498639107, + 0.01959950476884842, + 0.026930319145321846, + 0.024865487590432167, + 0.008311555720865726, + 0.011806009337306023, + 0.00780344707891345, + 0.005318745970726013, + 0.0065034739673137665, + 0.007172827608883381, + 0.004607165697962046, + 0.00987593550235033, + 0.006969744805246592, + 0.009138684719800949, + 0.007573497015982866, + 0.0068099722266197205, + 0.005548466928303242, + 0.008381134830415249, + 0.0065854997374117374, + 0.006023115944117308, + 0.016190700232982635, + 0.01186403539031744, + 0.01281353086233139, + 0.0076153031550347805, + 0.00906471349298954, + 0.013617685995995998, + 0.009547705762088299, + 0.01192405540496111, + 0.011049146763980389, + 0.008020428940653801, + 0.008892690762877464, + 0.006693585775792599, + 0.003877190174534917, + 0.00853482075035572, + 0.004568350035697222, + 0.005203958135098219, + 0.032510992139577866, + 0.028492320328950882, + 0.02443024516105652, + 0.03173376992344856, + 0.014657862484455109, + 0.018265020102262497, + 0.024035809561610222, + 0.020098242908716202, + 0.01147711742669344, + 0.0111189940944314, + 0.009204883128404617, + 0.008246555924415588, + 0.007205476518720388, + 0.00639101630076766, + 0.01624254882335663, + 0.009925459511578083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.08806448429822922 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, graceful 360-degree orbital pan around the central object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a globe with a yellowed parchment texture and hand-drawn classical map details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, bony hand with neatly manicured nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand's index finger move from the ancient Eurasian continent to the Pacific Ocean on the globe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, focused beam from an off-screen desk lamp casting highlights on the globe and brass stand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004660577047616243 + ], + "opposite_logprob_tokens": [ + -5.379660606384277 + ], + "logprob_reward": -0.004660577047616243, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004660577047616243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17302703857421875, + 0.08072984963655472, + 0.04317931458353996, + 0.22237291932106018, + 0.06260260939598083, + 0.07386831939220428, + 0.07116885483264923, + 0.05748138204216957, + 0.11926478892564774, + 0.09630497545003891 + ], + "credit_map": [ + 0.02293766848742962, + 0.020214201882481575, + 0.03436339274048805, + 0.024244170635938644, + 0.017712118104100227, + 0.01690153405070305, + 0.02301887609064579, + 0.013635092414915562, + 0.011409967206418514, + 0.007133603561669588, + 0.010606196708977222, + 0.009221119806170464, + 0.006935596466064453, + 0.007988073863089085, + 0.01371813379228115, + 0.013717149384319782, + 0.005567234009504318, + 0.003940274007618427, + 0.006884309928864241, + 0.0050362166948616505, + 0.005218603648245335, + 0.004330756142735481, + 0.006369255017489195, + 0.005832668859511614, + 0.030485065653920174, + 0.03085699863731861, + 0.033057693392038345, + 0.023856256157159805, + 0.029998555779457092, + 0.02212585136294365, + 0.027652140706777573, + 0.024340348318219185, + 0.009454915300011635, + 0.009951092302799225, + 0.008421618491411209, + 0.008387605659663677, + 0.0068372092209756374, + 0.005321058444678783, + 0.006993597839027643, + 0.007235512603074312, + 0.009885270148515701, + 0.01269320398569107, + 0.009677158668637276, + 0.007928949780762196, + 0.00785445049405098, + 0.00863466877490282, + 0.01051590871065855, + 0.0066787125542759895, + 0.012850036844611168, + 0.011136500164866447, + 0.01070144772529602, + 0.005811942741274834, + 0.005889624357223511, + 0.009039164520800114, + 0.006608402822166681, + 0.009131735190749168, + 0.010225638747215271, + 0.00736177247017622, + 0.007799915503710508, + 0.005824501626193523, + 0.004330589435994625, + 0.010031557641923428, + 0.005021445918828249, + 0.006885961163789034, + 0.020694199949502945, + 0.017762983217835426, + 0.01545681431889534, + 0.016433414071798325, + 0.008752770721912384, + 0.012105206027626991, + 0.015368443913757801, + 0.012690954841673374, + 0.014616536907851696, + 0.016135813668370247, + 0.011193021200597286, + 0.014406111091375351, + 0.007796213962137699, + 0.009191364049911499, + 0.014724108390510082, + 0.008241798728704453 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.004660577047616243 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, graceful 360-degree orbital pan around the central object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a globe with a yellowed parchment texture and hand-drawn classical map details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, bony hand with neatly manicured nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand's index finger move from the ancient Eurasian continent to the Pacific Ocean on the globe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, focused beam from an off-screen desk lamp casting highlights on the globe and brass stand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22557903826236725 + ], + "opposite_logprob_tokens": [ + -1.600579023361206 + ], + "logprob_reward": -0.22557903826236725, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22557903826236725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10271734744310379, + 0.0730210691690445, + 0.0777311623096466, + 0.2630506157875061, + 0.11429907381534576, + 0.19052886962890625, + 0.07114102691411972, + 0.03141331672668457, + 0.03768914192914963, + 0.038408391177654266 + ], + "credit_map": [ + 0.02143246866762638, + 0.010104699060320854, + 0.021687453612685204, + 0.01136974710971117, + 0.006252085790038109, + 0.009330269880592823, + 0.008267940953373909, + 0.014272687025368214, + 0.008866140618920326, + 0.01030468475073576, + 0.010434587486088276, + 0.0075461771339178085, + 0.009372490458190441, + 0.006985919084399939, + 0.013247048482298851, + 0.00626402348279953, + 0.010080530308187008, + 0.007687979377806187, + 0.011345628648996353, + 0.008907046169042587, + 0.009230549447238445, + 0.007564399857074022, + 0.011763794347643852, + 0.011151237413287163, + 0.03638549894094467, + 0.038859184831380844, + 0.04146338999271393, + 0.02795116789638996, + 0.031824465841054916, + 0.026871098205447197, + 0.030594097450375557, + 0.02910170704126358, + 0.013585369102656841, + 0.013353581540286541, + 0.013875847682356834, + 0.010860223323106766, + 0.012804890051484108, + 0.01623988151550293, + 0.015402457676827908, + 0.018176820129156113, + 0.024222664535045624, + 0.0325620137155056, + 0.025130243971943855, + 0.02225325256586075, + 0.019764577969908714, + 0.020453112199902534, + 0.027643347159028053, + 0.01849965937435627, + 0.011744744144380093, + 0.010846957564353943, + 0.00961181428283453, + 0.0066059366799890995, + 0.005909142550081015, + 0.008926862850785255, + 0.008824042975902557, + 0.008671523071825504, + 0.004942281171679497, + 0.0028148400597274303, + 0.003738553263247013, + 0.0028513032011687756, + 0.003274470567703247, + 0.004985117819160223, + 0.004257705062627792, + 0.0045490446500480175, + 0.005922732409089804, + 0.005095900036394596, + 0.004765396472066641, + 0.0049610016867518425, + 0.003321351483464241, + 0.005047603510320187, + 0.004210799001157284, + 0.004364356864243746, + 0.005580844823271036, + 0.004778080619871616, + 0.004609223920851946, + 0.004764508455991745, + 0.0027817271184176207, + 0.003076722379773855, + 0.008180621080100536, + 0.004636658355593681 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.22557903826236725 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, graceful 360-degree orbital pan around the central object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a globe with a yellowed parchment texture and hand-drawn classical map details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, bony hand with neatly manicured nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand's index finger move from the ancient Eurasian continent to the Pacific Ocean on the globe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, focused beam from an off-screen desk lamp casting highlights on the globe and brass stand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20161320269107819 + ], + "opposite_logprob_tokens": [ + -1.701613187789917 + ], + "logprob_reward": -0.20161320269107819, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20161320269107819, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12991750240325928, + 0.08335430175065994, + 0.06845396757125854, + 0.33617544174194336, + 0.09677625447511673, + 0.10852666199207306, + 0.05095941573381424, + 0.05890415236353874, + 0.03651798889040947, + 0.03041435033082962 + ], + "credit_map": [ + 0.02764059603214264, + 0.014378853142261505, + 0.021039042621850967, + 0.014350554905831814, + 0.010281735099852085, + 0.012127259746193886, + 0.014994744211435318, + 0.015104721300303936, + 0.009000901132822037, + 0.010969734750688076, + 0.010924295522272587, + 0.00893451925367117, + 0.010912968777120113, + 0.008133606053888798, + 0.015914281830191612, + 0.0085639962926507, + 0.008365402929484844, + 0.0073418887332081795, + 0.009357393719255924, + 0.0077523463405668736, + 0.00885829795151949, + 0.007256982382386923, + 0.00993824191391468, + 0.009583412669599056, + 0.04833073914051056, + 0.050137005746364594, + 0.05597713962197304, + 0.03402816504240036, + 0.040917351841926575, + 0.03227931261062622, + 0.03767858445644379, + 0.03682713955640793, + 0.013830315321683884, + 0.015099775977432728, + 0.012791980989277363, + 0.006326050963252783, + 0.00973203219473362, + 0.012500021606683731, + 0.009244514629244804, + 0.017251569777727127, + 0.013801192864775658, + 0.018798284232616425, + 0.014182827435433865, + 0.013073848560452461, + 0.011321758851408958, + 0.011463839560747147, + 0.015329254791140556, + 0.010555654764175415, + 0.010248623788356781, + 0.006840656511485577, + 0.007122003007680178, + 0.0037140834610909224, + 0.004942279774695635, + 0.007772962097078562, + 0.005214373581111431, + 0.005104432813823223, + 0.01477388758212328, + 0.007064253091812134, + 0.009631350636482239, + 0.0037097514141350985, + 0.004742786753922701, + 0.00701090507209301, + 0.005643205251544714, + 0.006328013725578785, + 0.0063131460919976234, + 0.004861058201640844, + 0.0046547697857022285, + 0.004544558469206095, + 0.003141488181427121, + 0.004028421826660633, + 0.0048008207231760025, + 0.004173724912106991, + 0.006729265209287405, + 0.0046183448284864426, + 0.0038589774630963802, + 0.004030503798276186, + 0.0022503468208014965, + 0.0030214020516723394, + 0.003784926375374198, + 0.002120584947988391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.20161320269107819 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, graceful 360-degree orbital pan around the central object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a globe with a yellowed parchment texture and hand-drawn classical map details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, bony hand with neatly manicured nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand's index finger move from the ancient Eurasian continent to the Pacific Ocean on the globe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, focused beam from an off-screen desk lamp casting highlights on the globe and brass stand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002817115979269147 + ], + "opposite_logprob_tokens": [ + -5.877817153930664 + ], + "logprob_reward": -0.002817115979269147, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002817115979269147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20604820549488068, + 0.11225399374961853, + 0.06898292154073715, + 0.12949465215206146, + 0.1090373769402504, + 0.08901651948690414, + 0.05358847603201866, + 0.11119554191827774, + 0.07050592452287674, + 0.04987644776701927 + ], + "credit_map": [ + 0.044183991849422455, + 0.021751850843429565, + 0.03140750154852867, + 0.025939330458641052, + 0.014798780903220177, + 0.019340550526976585, + 0.020968448370695114, + 0.027657726779580116, + 0.013794446364045143, + 0.013268432579934597, + 0.020882485434412956, + 0.010583695955574512, + 0.007399844937026501, + 0.011956912465393543, + 0.012418056838214397, + 0.02195010893046856, + 0.011032870970666409, + 0.013346645049750805, + 0.009634717367589474, + 0.009080267511308193, + 0.0049871476367115974, + 0.006041626911610365, + 0.006342222914099693, + 0.008517410606145859, + 0.026217205449938774, + 0.017577532678842545, + 0.02515653520822525, + 0.008334976620972157, + 0.013765858486294746, + 0.012799752876162529, + 0.01162196509540081, + 0.01402081549167633, + 0.018199831247329712, + 0.0200776606798172, + 0.014493358321487904, + 0.008215936832129955, + 0.009314842522144318, + 0.015463334508240223, + 0.007628222927451134, + 0.01564418338239193, + 0.01626722142100334, + 0.01355568878352642, + 0.013133483938872814, + 0.010785224847495556, + 0.007675035856664181, + 0.012706588953733444, + 0.007828404195606709, + 0.007064858451485634, + 0.008445356041193008, + 0.006233304738998413, + 0.0070058428682386875, + 0.006247593555599451, + 0.005136873573064804, + 0.0095452219247818, + 0.004330300260335207, + 0.006643973756581545, + 0.015056295320391655, + 0.016494054347276688, + 0.015258540399372578, + 0.011726056225597858, + 0.010111508890986443, + 0.009988315403461456, + 0.01443222351372242, + 0.018128542229533195, + 0.012889855541288853, + 0.007786920294165611, + 0.009522642008960247, + 0.007224861066788435, + 0.004403105471283197, + 0.008554402738809586, + 0.006899615284055471, + 0.013224508613348007, + 0.01069126557558775, + 0.008787599392235279, + 0.006880050525069237, + 0.005891864188015461, + 0.004080984741449356, + 0.004451898857951164, + 0.0047849612310528755, + 0.004307821858674288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.002817115979269147 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up, the camera executes a slow, graceful 360-degree orbital pan around an exquisite globe, perched on an antique, dark wood base and supported by a slightly rusted brass theodolite stand. Its surface has the texture of yellowed parchment, revealing the delicate lines of a hand-drawn classical map. Place names are marked in elegant serif fonts, and the subtly raised mountain ranges are clearly visible in the light. A slender, bony hand with neatly manicured nails enters the frame. The tip of its index finger delicately and almost without pressure smoothly moves from the ancient Eurasian continent on the globe to the vast Pacific Ocean, as if tactilely tracing the threads of history and geographical change. The entire scene is illuminated by a warm, focused beam from an off-screen desk lamp, its soft light casting gentle highlights on the globe's curved surface and reflecting flecks of golden light on the brass stand. The study in the background is shrouded in deep shadow. The entire painting is dominated by warm brown, gold and cream tones, creating an atmosphere of tranquility, contemplation, wisdom and history.", + "seed": 14034, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_10/index_9666_seed_14034_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04402077328413725, + "vlm_logprob": -0.04402077328413725, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.04402077328413725, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, graceful 360-degree orbital pan around the central object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a globe with a yellowed parchment texture and hand-drawn classical map details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, bony hand with neatly manicured nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand's index finger move from the ancient Eurasian continent to the Pacific Ocean on the globe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, focused beam from an off-screen desk lamp casting highlights on the globe and brass stand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0612945482134819 + ], + "opposite_logprob_tokens": [ + -3.8112945556640625 + ], + "logprob_reward": -0.0612945482134819, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0612945482134819, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2005258947610855, + 0.20966707170009613, + 0.05237141624093056, + 0.05297277122735977, + 0.04458169266581535, + 0.07856012135744095, + 0.011456482112407684, + 0.2212628275156021, + 0.11335775256156921, + 0.015243981964886189 + ], + "credit_map": [ + 0.019274655729532242, + 0.023673389106988907, + 0.03043449856340885, + 0.027667688205838203, + 0.04295124486088753, + 0.018577665090560913, + 0.03020387329161167, + 0.007742882706224918, + 0.018231691792607307, + 0.02879583090543747, + 0.027563054114580154, + 0.024023868143558502, + 0.029658975079655647, + 0.0342133603990078, + 0.02729036845266819, + 0.019889920949935913, + 0.004190423525869846, + 0.006632938981056213, + 0.011395547538995743, + 0.005543472245335579, + 0.004950759932398796, + 0.00819579791277647, + 0.0043767220340669155, + 0.007085748016834259, + 0.004545897711068392, + 0.004248440265655518, + 0.00585842365399003, + 0.007360219955444336, + 0.005372697487473488, + 0.00719944154843688, + 0.011191198602318764, + 0.007196453399956226, + 0.006528348196297884, + 0.0028568326961249113, + 0.006992712151259184, + 0.004491329658776522, + 0.00814275536686182, + 0.005037173628807068, + 0.005593812558799982, + 0.004938725847750902, + 0.011254743672907352, + 0.009454019367694855, + 0.00857260450720787, + 0.010425000451505184, + 0.0071294610388576984, + 0.012192036025226116, + 0.011659598909318447, + 0.007872659713029861, + 0.0016874598804861307, + 0.001258037518709898, + 0.0017952320631593466, + 0.0018209846457466483, + 0.0011154783423990011, + 0.0013276385143399239, + 0.001334169995971024, + 0.0011174812680110335, + 0.014104451052844524, + 0.0321933887898922, + 0.02798282913863659, + 0.045528072863817215, + 0.02674287185072899, + 0.028311042115092278, + 0.02488759718835354, + 0.02151256427168846, + 0.00979467574506998, + 0.009108222089707851, + 0.009405238553881645, + 0.011980886571109295, + 0.009542443789541721, + 0.031126772984862328, + 0.01602213829755783, + 0.016377372667193413, + 0.0018228053813800216, + 0.0012516190763562918, + 0.0023890286684036255, + 0.0023902247194200754, + 0.0016473684227094054, + 0.001380356028676033, + 0.0025143746752291918, + 0.0018482041778042912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.0612945482134819 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, graceful 360-degree orbital pan around the central object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a globe with a yellowed parchment texture and hand-drawn classical map details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, bony hand with neatly manicured nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand's index finger move from the ancient Eurasian continent to the Pacific Ocean on the globe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, focused beam from an off-screen desk lamp casting highlights on the globe and brass stand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003653638530522585 + ], + "opposite_logprob_tokens": [ + -5.628653526306152 + ], + "logprob_reward": -0.003653638530522585, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003653638530522585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20462989807128906, + 0.1703336387872696, + 0.03647753968834877, + 0.06487105786800385, + 0.08011670410633087, + 0.06736671924591064, + 0.03147934377193451, + 0.10530318319797516, + 0.09189315885305405, + 0.1475287526845932 + ], + "credit_map": [ + 0.015368511900305748, + 0.02111975848674774, + 0.03588083013892174, + 0.04619607329368591, + 0.02528652921319008, + 0.020576396957039833, + 0.027690505608916283, + 0.012511316686868668, + 0.011529731564223766, + 0.018634941428899765, + 0.022846443578600883, + 0.021244699135422707, + 0.0225260928273201, + 0.023290658369660378, + 0.03392913565039635, + 0.016331937164068222, + 0.0036536408588290215, + 0.0040552811697125435, + 0.0045057181268930435, + 0.00632417481392622, + 0.003975264262408018, + 0.005184996407479048, + 0.003912684973329306, + 0.004865776281803846, + 0.004890512675046921, + 0.0047204699367284775, + 0.0076590063981711864, + 0.007869143970310688, + 0.006272585596889257, + 0.0064711300656199455, + 0.01450735330581665, + 0.012480859644711018, + 0.00971665047109127, + 0.005475458689033985, + 0.01476882491260767, + 0.009125552140176296, + 0.015135039575397968, + 0.008323084563016891, + 0.01058497279882431, + 0.006987120024859905, + 0.010411270894110203, + 0.007320999167859554, + 0.008255884982645512, + 0.008530608378350735, + 0.006105520762503147, + 0.008369882591068745, + 0.01064552366733551, + 0.007727023214101791, + 0.0035815455485135317, + 0.0031530142296105623, + 0.003837338648736477, + 0.004596959333866835, + 0.0033656463492661715, + 0.005305366590619087, + 0.004014915321022272, + 0.003624559845775366, + 0.007840214297175407, + 0.015920177102088928, + 0.013155081309378147, + 0.01512374822050333, + 0.013508075848221779, + 0.014823667705059052, + 0.01212495006620884, + 0.012807266786694527, + 0.008402272127568722, + 0.007333744782954454, + 0.014219078235328197, + 0.010436490178108215, + 0.009386206977069378, + 0.008874702267348766, + 0.01811317913234234, + 0.015127480030059814, + 0.009711318649351597, + 0.015526287257671356, + 0.01863240823149681, + 0.020503252744674683, + 0.01608414761722088, + 0.018140094354748726, + 0.03136150911450386, + 0.017569728195667267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.003653638530522585 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, graceful 360-degree orbital pan around the central object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a globe with a yellowed parchment texture and hand-drawn classical map details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, bony hand with neatly manicured nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand's index finger move from the ancient Eurasian continent to the Pacific Ocean on the globe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, focused beam from an off-screen desk lamp casting highlights on the globe and brass stand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810303658246994 + ], + "opposite_logprob_tokens": [ + -3.2881031036376953 + ], + "logprob_reward": -0.03810303658246994, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810303658246994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1208162009716034, + 0.3420860171318054, + 0.05320695787668228, + 0.03647487610578537, + 0.029192861169576645, + 0.02872326597571373, + 0.023176265880465508, + 0.10691216588020325, + 0.2491026520729065, + 0.01030877884477377 + ], + "credit_map": [ + 0.008891222067177296, + 0.01461535319685936, + 0.018104203045368195, + 0.021863993257284164, + 0.01778368093073368, + 0.017052026465535164, + 0.016304906457662582, + 0.006200817413628101, + 0.024157186970114708, + 0.035202473402023315, + 0.049221303313970566, + 0.05323757976293564, + 0.02589588798582554, + 0.04485636577010155, + 0.07411488890647888, + 0.03540031239390373, + 0.0045392150059342384, + 0.00652278633788228, + 0.01265360414981842, + 0.006816766690462828, + 0.0045938934199512005, + 0.00654465239495039, + 0.006163547746837139, + 0.005372493527829647, + 0.003246917622163892, + 0.003161621978506446, + 0.00396738899871707, + 0.00411980040371418, + 0.005673619918525219, + 0.0054660397581756115, + 0.005911923013627529, + 0.004927567671984434, + 0.0035413678269833326, + 0.0024549742229282856, + 0.0033906938042491674, + 0.002687993925064802, + 0.0038493145257234573, + 0.005307138431817293, + 0.003938544075936079, + 0.004022834822535515, + 0.0035598266404122114, + 0.0029185356106609106, + 0.0037680319510400295, + 0.0048896633088588715, + 0.0026783980429172516, + 0.003754965728148818, + 0.003920488990843296, + 0.003233355237171054, + 0.0020248317159712315, + 0.0018796746153384447, + 0.0030182988848537207, + 0.0030638324096798897, + 0.0019966403488069773, + 0.005353303626179695, + 0.0023092746268957853, + 0.0035304096527397633, + 0.007597021758556366, + 0.014726435765624046, + 0.013886656612157822, + 0.025204911828041077, + 0.012470640242099762, + 0.01234497968107462, + 0.011170676909387112, + 0.009510844945907593, + 0.022494573146104813, + 0.021481694653630257, + 0.018871158361434937, + 0.02391340769827366, + 0.024767626076936722, + 0.07992682605981827, + 0.023045985028147697, + 0.03460138291120529, + 0.0008476261864416301, + 0.0011033400660380721, + 0.0011652602115646005, + 0.0013526275288313627, + 0.0010250586783513427, + 0.0013397630536928773, + 0.0020462621469050646, + 0.0014288423117250204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.03810303658246994 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, graceful 360-degree orbital pan around the central object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a globe with a yellowed parchment texture and hand-drawn classical map details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, bony hand with neatly manicured nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand's index finger move from the ancient Eurasian continent to the Pacific Ocean on the globe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, focused beam from an off-screen desk lamp casting highlights on the globe and brass stand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.112955242395401 + ], + "opposite_logprob_tokens": [ + -2.237955331802368 + ], + "logprob_reward": -0.112955242395401, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.112955242395401, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2183661311864853, + 0.24959692358970642, + 0.0933847650885582, + 0.03517318144440651, + 0.048945557326078415, + 0.05399427190423012, + 0.07081210613250732, + 0.0484035350382328, + 0.1239481195807457, + 0.05737540125846863 + ], + "credit_map": [ + 0.01738583669066429, + 0.029109589755535126, + 0.03131328895688057, + 0.02846318855881691, + 0.02872178517282009, + 0.02924400381743908, + 0.04027719423174858, + 0.013851242139935493, + 0.019763698801398277, + 0.029884785413742065, + 0.030429264530539513, + 0.031504642218351364, + 0.028534522280097008, + 0.03213079273700714, + 0.047973860055208206, + 0.029375378042459488, + 0.007973259314894676, + 0.011323719285428524, + 0.023126469925045967, + 0.011394643224775791, + 0.008412765339016914, + 0.011934444308280945, + 0.008073100820183754, + 0.01114636566489935, + 0.0038472702726721764, + 0.0026987663004547358, + 0.00461318576708436, + 0.003957844339311123, + 0.004301855340600014, + 0.00510709173977375, + 0.006469473708420992, + 0.004177697468549013, + 0.004453811328858137, + 0.003428244963288307, + 0.006317379418760538, + 0.0042000627145171165, + 0.00590521888807416, + 0.008898206986486912, + 0.010004434734582901, + 0.005738197825849056, + 0.004785081837326288, + 0.004361244384199381, + 0.006308810319751501, + 0.010203246027231216, + 0.00473325839266181, + 0.007632679771631956, + 0.009748511016368866, + 0.006221444346010685, + 0.0056679644621908665, + 0.006150421220809221, + 0.007160006556659937, + 0.011463155969977379, + 0.006782570853829384, + 0.014202555641531944, + 0.008491462096571922, + 0.010893970727920532, + 0.0032448628917336464, + 0.00678674504160881, + 0.0052498760633170605, + 0.009498009458184242, + 0.006337455008178949, + 0.0069054653868079185, + 0.00583229772746563, + 0.0045488192699849606, + 0.008498696610331535, + 0.01016725692898035, + 0.012825599871575832, + 0.013918710872530937, + 0.021393727511167526, + 0.027996715158224106, + 0.014297720044851303, + 0.014849696308374405, + 0.0038854116573929787, + 0.00589397456496954, + 0.007609950844198465, + 0.009496745653450489, + 0.0062988135032355785, + 0.004950992297381163, + 0.012268139980733395, + 0.006971373222768307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.112955242395401 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, graceful 360-degree orbital pan around the central object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a globe with a yellowed parchment texture and hand-drawn classical map details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, bony hand with neatly manicured nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand's index finger move from the ancient Eurasian continent to the Pacific Ocean on the globe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, focused beam from an off-screen desk lamp casting highlights on the globe and brass stand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097400698810816 + ], + "opposite_logprob_tokens": [ + -5.5040974617004395 + ], + "logprob_reward": -0.004097400698810816, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097400698810816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22677837312221527, + 0.09231490641832352, + 0.06953577697277069, + 0.03256598114967346, + 0.033367183059453964, + 0.06456345319747925, + 0.2608650028705597, + 0.1253829151391983, + 0.06957799196243286, + 0.025048496201634407 + ], + "credit_map": [ + 0.018649259582161903, + 0.025766799226403236, + 0.04329099506139755, + 0.044483501464128494, + 0.02583894319832325, + 0.02807912603020668, + 0.024476632475852966, + 0.01619313471019268, + 0.00739782489836216, + 0.009795459918677807, + 0.013948492705821991, + 0.015085830353200436, + 0.012880475260317326, + 0.011928610503673553, + 0.009786522015929222, + 0.011491687037050724, + 0.006498666480183601, + 0.009232373908162117, + 0.021397802978754044, + 0.006595179904252291, + 0.005009245593100786, + 0.005954355001449585, + 0.005716323386877775, + 0.009131827391684055, + 0.004877328407019377, + 0.0029680223669856787, + 0.004437933210283518, + 0.005714119877666235, + 0.0029079525265842676, + 0.003941280767321587, + 0.004557386972010136, + 0.003161957487463951, + 0.0036295915488153696, + 0.0024935484398156404, + 0.0035010690335184336, + 0.004019059706479311, + 0.004075019154697657, + 0.004468706902116537, + 0.0056094215251505375, + 0.0055707660503685474, + 0.009494015946984291, + 0.007359426002949476, + 0.00748076057061553, + 0.008456170558929443, + 0.006042297463864088, + 0.008497776463627815, + 0.009578165598213673, + 0.007654836401343346, + 0.029531460255384445, + 0.031977809965610504, + 0.05111055448651314, + 0.03478815406560898, + 0.016790097579360008, + 0.040507491677999496, + 0.0327717624604702, + 0.023387635126709938, + 0.0090880636125803, + 0.017089881002902985, + 0.01598549447953701, + 0.031461335718631744, + 0.014431335031986237, + 0.01382956001907587, + 0.012744178995490074, + 0.010753056034445763, + 0.0064111026003956795, + 0.00575011083856225, + 0.0066046337597072124, + 0.007216269616037607, + 0.007644264027476311, + 0.011759014800190926, + 0.010346216149628162, + 0.013846376910805702, + 0.0017767930403351784, + 0.002224392956122756, + 0.0032716356217861176, + 0.003634318709373474, + 0.0035031891893595457, + 0.0029312081169337034, + 0.004974862094968557, + 0.0027320983354002237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.004097400698810816 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_12/index_tensor([8978, 8978])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_12/index_tensor([8978, 8978])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0c2ef8a260b09e68bee183cd0d8cb41507efc3e3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_12/index_tensor([8978, 8978])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "shot shot from an extremely low angle, the camera slowly and steadily tilts upward, following the majestic and deliberate rise of a massive humpback whale from the deep, inky depths of the ocean. The beast's skin is a deep, almost black blue, covered in ancient scars and patches of barnacles. But most breathtakingly, its entire body is covered in millions of tiny, symbiotic organisms, a flowing river of stars, emitting a soft, mysterious blue-green bioluminescence. In this abyssal depths, devoid of visible natural light, these bioluminescent organisms serve as the sole source of light. Their glow pierces the inky blackness, not only clearly outlining the whale's massive and graceful silhouette but also illuminating the tiny particles suspended around it, making them appear like cosmic dust. With uncanny grace, the whale breaks through the current, its massive body barely causing any turbulence as it steadily and slowly moves toward the unknown surface. The entire scene is shrouded in a sacred and tranquil atmosphere, creating a breathtaking sense of awe and an otherworldly, dreamlike beauty.", + "seed": 21019, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_12/index_8978_seed_21019_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2229563444852829, + "vlm_logprob": -0.2229563444852829, + "vlm_yes_no_margin": 1.9, + "vlm_token_credit": -0.2229563444852829, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive humpback whale visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the whale's skin dark blue/black with visible barnacles and bioluminescent organisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the whale's bioluminescence in an otherwise dark environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5991312861442566 + ], + "opposite_logprob_tokens": [ + -0.8491312861442566 + ], + "logprob_reward": -0.5991312861442566, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5991312861442566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0680251196026802, + 0.06830694526433945, + 0.029353786259889603, + 0.04809362068772316, + 0.47365930676460266, + 0.04862527549266815, + 0.031161120161414146, + 0.03460671752691269, + 0.08917249739170074, + 0.10899558663368225 + ], + "credit_map": [ + 0.005463328678160906, + 0.009052014909684658, + 0.010851794853806496, + 0.008741342462599277, + 0.006371538154780865, + 0.009894023649394512, + 0.006747871171683073, + 0.010903207585215569, + 0.011234553530812263, + 0.008767389692366123, + 0.008978056721389294, + 0.006969011854380369, + 0.006444818340241909, + 0.010232564993202686, + 0.006200488191097975, + 0.009480062872171402, + 0.00391859607771039, + 0.0025546392425894737, + 0.00583465164527297, + 0.00338366674259305, + 0.004087857902050018, + 0.004744399338960648, + 0.0025940651539713144, + 0.002235909691080451, + 0.004709526896476746, + 0.005602682940661907, + 0.005401920527219772, + 0.010154378600418568, + 0.004506589844822884, + 0.007293348666280508, + 0.005821668077260256, + 0.004603501409292221, + 0.0253894105553627, + 0.019350653514266014, + 0.17003796994686127, + 0.019385099411010742, + 0.014886903576552868, + 0.19371236860752106, + 0.0129958251491189, + 0.017901064828038216, + 0.0060135978274047375, + 0.008447489701211452, + 0.006866324692964554, + 0.005567178130149841, + 0.004789059516042471, + 0.006514246575534344, + 0.00407823920249939, + 0.006349144037812948, + 0.003744880435988307, + 0.003527646418660879, + 0.0039452603086829185, + 0.003934598062187433, + 0.0022951613646000624, + 0.00400528172031045, + 0.004304398316890001, + 0.005403895862400532, + 0.004354942124336958, + 0.0029711537063121796, + 0.006283515132963657, + 0.002799061592668295, + 0.0031387624330818653, + 0.0064336745999753475, + 0.004103658255189657, + 0.0045219468884170055, + 0.012919805943965912, + 0.01059125829488039, + 0.01830449514091015, + 0.007815390825271606, + 0.0055175176821649075, + 0.007824106141924858, + 0.013026791624724865, + 0.013173134997487068, + 0.01027479488402605, + 0.008446067571640015, + 0.03262845426797867, + 0.009696630761027336, + 0.0051580918952822685, + 0.006392227485775948, + 0.022582771256566048, + 0.013816551305353642 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.5991312861442566 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive humpback whale visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the whale's skin dark blue/black with visible barnacles and bioluminescent organisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the whale's bioluminescence in an otherwise dark environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20160064101219177 + ], + "opposite_logprob_tokens": [ + -1.7016006708145142 + ], + "logprob_reward": -0.20160064101219177, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20160064101219177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05314464867115021, + 0.10116419196128845, + 0.025062359869480133, + 0.061883967369794846, + 0.6053234934806824, + 0.03240323066711426, + 0.028184402734041214, + 0.030721493065357208, + 0.021129440516233444, + 0.04098270460963249 + ], + "credit_map": [ + 0.004156277980655432, + 0.00704521220177412, + 0.007031241897493601, + 0.008768396452069283, + 0.004656508564949036, + 0.007542353589087725, + 0.005639730487018824, + 0.008304926566779613, + 0.01148146390914917, + 0.013948614709079266, + 0.01310688629746437, + 0.012471210211515427, + 0.013077215291559696, + 0.014754654839634895, + 0.011390679515898228, + 0.010933471843600273, + 0.0026549226604402065, + 0.0030178262386471033, + 0.004063248168677092, + 0.003671195823699236, + 0.0018854084191843867, + 0.004066512919962406, + 0.0030924915336072445, + 0.00261075422167778, + 0.006906002294272184, + 0.00883842259645462, + 0.009646047838032246, + 0.009066086262464523, + 0.006010089535266161, + 0.010275674983859062, + 0.005159602500498295, + 0.005982041824609041, + 0.02537824958562851, + 0.00918307900428772, + 0.2399515062570572, + 0.015269158408045769, + 0.0174520593136549, + 0.2693866491317749, + 0.012221330776810646, + 0.016481488943099976, + 0.004632031079381704, + 0.0025018786545842886, + 0.0070781102403998375, + 0.003092593513429165, + 0.002789278281852603, + 0.004821788053959608, + 0.0028986434917896986, + 0.004588905721902847, + 0.0032157001551240683, + 0.0033006579615175724, + 0.0049974555149674416, + 0.0027662639040499926, + 0.0026828462723642588, + 0.0031777103431522846, + 0.0036882346030324697, + 0.004355531185865402, + 0.004051493480801582, + 0.003623519791290164, + 0.005851371213793755, + 0.0030297355260699987, + 0.0022672382183372974, + 0.006077099125832319, + 0.0031400464940816164, + 0.002680990844964981, + 0.0024596683215349913, + 0.001647257013246417, + 0.004104147665202618, + 0.001395952538587153, + 0.0018685891991481185, + 0.00202031503431499, + 0.004803440999239683, + 0.0028300685808062553, + 0.0053849718533456326, + 0.003079820889979601, + 0.009874953888356686, + 0.0034728317987173796, + 0.001849704422056675, + 0.003187765134498477, + 0.006104599684476852, + 0.008028056472539902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.20160064101219177 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive humpback whale visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the whale's skin dark blue/black with visible barnacles and bioluminescent organisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the whale's bioluminescence in an otherwise dark environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026348866522312164 + ], + "opposite_logprob_tokens": [ + -3.651348829269409 + ], + "logprob_reward": -0.026348866522312164, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026348866522312164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07807325571775436, + 0.026547860354185104, + 0.03802235424518585, + 0.17960792779922485, + 0.2476656138896942, + 0.21006685495376587, + 0.05475737899541855, + 0.07183442264795303, + 0.00966473389416933, + 0.08375958353281021 + ], + "credit_map": [ + 0.006389346905052662, + 0.012342056259512901, + 0.010988974012434483, + 0.016363797709345818, + 0.005044078454375267, + 0.008735100738704205, + 0.008435605093836784, + 0.009774292819201946, + 0.002222768496721983, + 0.002905388129875064, + 0.0034058874007314444, + 0.0033971352968364954, + 0.004116572439670563, + 0.0037743407301604748, + 0.003763430053368211, + 0.0029623364098370075, + 0.0036674626171588898, + 0.00405493238940835, + 0.005804709158837795, + 0.006440264638513327, + 0.002775830216705799, + 0.00531859090551734, + 0.004828360863029957, + 0.005132200196385384, + 0.023783022537827492, + 0.02730502188205719, + 0.027446845546364784, + 0.02498898096382618, + 0.01609252206981182, + 0.025448104366660118, + 0.017968345433473587, + 0.01657509058713913, + 0.011673633009195328, + 0.009588894434273243, + 0.08701300621032715, + 0.009841199964284897, + 0.008680210448801517, + 0.09679783135652542, + 0.009853639639914036, + 0.014217209070920944, + 0.024963688105344772, + 0.035623949021101, + 0.03038480319082737, + 0.019251877442002296, + 0.022543026134371758, + 0.027368925511837006, + 0.02011508122086525, + 0.029815489426255226, + 0.005392142571508884, + 0.008645856752991676, + 0.007555739488452673, + 0.005975933745503426, + 0.00475962832570076, + 0.0061540198512375355, + 0.0069817169569432735, + 0.009292344562709332, + 0.008835329674184322, + 0.006574273109436035, + 0.012753954157233238, + 0.006905091926455498, + 0.005750606767833233, + 0.008418083190917969, + 0.012627624906599522, + 0.009969465434551239, + 0.001423839945346117, + 0.00142528279684484, + 0.001555567141622305, + 0.000857183535117656, + 0.0008595687686465681, + 0.0009164582588709891, + 0.0013302246807143092, + 0.0012966090580448508, + 0.008842670358717442, + 0.006227909587323666, + 0.02222239226102829, + 0.00745961582288146, + 0.004400443751364946, + 0.006417041644454002, + 0.015403740108013153, + 0.012785765342414379 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.026348866522312164 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive humpback whale visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the whale's skin dark blue/black with visible barnacles and bioluminescent organisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the whale's bioluminescence in an otherwise dark environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06208469718694687 + ], + "opposite_logprob_tokens": [ + -2.812084674835205 + ], + "logprob_reward": -0.06208469718694687, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06208469718694687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050747696310281754, + 0.062497057020664215, + 0.06959149986505508, + 0.07508191466331482, + 0.3480149209499359, + 0.06765410304069519, + 0.08893337845802307, + 0.07504469156265259, + 0.020839311182498932, + 0.14159537851810455 + ], + "credit_map": [ + 0.005118028726428747, + 0.0042844656854867935, + 0.009997781366109848, + 0.007157525047659874, + 0.003083372488617897, + 0.009842431172728539, + 0.004350609611719847, + 0.0069134850054979324, + 0.00870266743004322, + 0.00867860671132803, + 0.009269907139241695, + 0.010259997099637985, + 0.0075875259935855865, + 0.007420264184474945, + 0.0051580676808953285, + 0.005420031026005745, + 0.005974444095045328, + 0.0065974765457212925, + 0.0119924396276474, + 0.01272413320839405, + 0.005692940205335617, + 0.009763052687048912, + 0.009426156058907509, + 0.007420859765261412, + 0.008149177767336369, + 0.008813280612230301, + 0.01081039197742939, + 0.011123837903141975, + 0.00793085154145956, + 0.01191651914268732, + 0.00884697400033474, + 0.007490890566259623, + 0.022321125492453575, + 0.018803097307682037, + 0.11094153672456741, + 0.018549969419836998, + 0.01632140390574932, + 0.12444155663251877, + 0.01659327559173107, + 0.020042989403009415, + 0.007347112987190485, + 0.011511312797665596, + 0.008791917003691196, + 0.00663843983784318, + 0.006472758948802948, + 0.007539502345025539, + 0.00618837820366025, + 0.013164680451154709, + 0.00914468988776207, + 0.013203264214098454, + 0.012702020816504955, + 0.008398603647947311, + 0.008339808322489262, + 0.01057048887014389, + 0.011302525177598, + 0.015271982178092003, + 0.01215166412293911, + 0.010250684805214405, + 0.012044725008308887, + 0.006867060903459787, + 0.0056744907051324844, + 0.008339072577655315, + 0.009358788840472698, + 0.010358207859098911, + 0.002960270270705223, + 0.0027665679808706045, + 0.003603445366024971, + 0.002017872640863061, + 0.0021508566569536924, + 0.002127123763784766, + 0.002557607600465417, + 0.002655569463968277, + 0.016195382922887802, + 0.013267210684716702, + 0.03216670826077461, + 0.01257417444139719, + 0.011326122097671032, + 0.015131142921745777, + 0.025541795417666435, + 0.01539286132901907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 6 + ], + "selected_reward": -0.06208469718694687 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive humpback whale visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the whale's skin dark blue/black with visible barnacles and bioluminescent organisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the whale's bioluminescence in an otherwise dark environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2256162315607071 + ], + "opposite_logprob_tokens": [ + -1.600616216659546 + ], + "logprob_reward": -0.2256162315607071, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2256162315607071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029080232605338097, + 0.014372068457305431, + 0.027844678610563278, + 0.0633709579706192, + 0.6932898759841919, + 0.03696748614311218, + 0.06448312103748322, + 0.022923564538359642, + 0.02214028872549534, + 0.0255277082324028 + ], + "credit_map": [ + 0.003242356237024069, + 0.002623508684337139, + 0.004537471104413271, + 0.004599398002028465, + 0.0018062733579427004, + 0.005214579403400421, + 0.0031291027553379536, + 0.003927544225007296, + 0.0017788622062653303, + 0.0021857498213648796, + 0.0020027542486786842, + 0.0014984917361289263, + 0.001572311855852604, + 0.0019115518080070615, + 0.0018203584477305412, + 0.001601989264599979, + 0.003872271627187729, + 0.0027356427162885666, + 0.006084415130317211, + 0.0030810884200036526, + 0.002569283591583371, + 0.005355057772248983, + 0.0020236673299223185, + 0.002123252023011446, + 0.008884340524673462, + 0.01092346291989088, + 0.011247452348470688, + 0.005337948445230722, + 0.006383426021784544, + 0.010095812380313873, + 0.004811392165720463, + 0.005687125492841005, + 0.027488769963383675, + 0.012159666046500206, + 0.2728973925113678, + 0.019499875605106354, + 0.01914973370730877, + 0.30418989062309265, + 0.015965115278959274, + 0.021939437836408615, + 0.004619657527655363, + 0.003432282479479909, + 0.005003617610782385, + 0.0033704049419611692, + 0.003499189857393503, + 0.005980879068374634, + 0.004257309250533581, + 0.0068041435442864895, + 0.006624050438404083, + 0.010026487521827221, + 0.008714601397514343, + 0.007346963044255972, + 0.005533052608370781, + 0.007115398533642292, + 0.007934598252177238, + 0.011187976226210594, + 0.0030062159057706594, + 0.0022955473978072405, + 0.003912705462425947, + 0.0019471374107524753, + 0.0018743335967883468, + 0.004302722401916981, + 0.0027383374981582165, + 0.0028465662617236376, + 0.0027871313504874706, + 0.0015386417508125305, + 0.005991293117403984, + 0.0017039591912180185, + 0.0018434205558151007, + 0.0015368630411103368, + 0.0038480579387396574, + 0.0028909232933074236, + 0.0030520171858370304, + 0.001926614553667605, + 0.004909760784357786, + 0.002323396969586611, + 0.001494656433351338, + 0.001829660264775157, + 0.003960242960602045, + 0.006031361408531666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 3 + ], + "selected_reward": -0.2256162315607071 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot shot from an extremely low angle, the camera slowly and steadily tilts upward, following the majestic and deliberate rise of a massive humpback whale from the deep, inky depths of the ocean. The beast's skin is a deep, almost black blue, covered in ancient scars and patches of barnacles. But most breathtakingly, its entire body is covered in millions of tiny, symbiotic organisms, a flowing river of stars, emitting a soft, mysterious blue-green bioluminescence. In this abyssal depths, devoid of visible natural light, these bioluminescent organisms serve as the sole source of light. Their glow pierces the inky blackness, not only clearly outlining the whale's massive and graceful silhouette but also illuminating the tiny particles suspended around it, making them appear like cosmic dust. With uncanny grace, the whale breaks through the current, its massive body barely causing any turbulence as it steadily and slowly moves toward the unknown surface. The entire scene is shrouded in a sacred and tranquil atmosphere, creating a breathtaking sense of awe and an otherworldly, dreamlike beauty.", + "seed": 21020, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_12/index_8978_seed_21020_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1492491458542645, + "vlm_logprob": -0.1492491458542645, + "vlm_yes_no_margin": 2.874999988079071, + "vlm_token_credit": -0.1492491458542645, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive humpback whale visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the whale's skin dark blue/black with visible barnacles and bioluminescent organisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the whale's bioluminescence in an otherwise dark environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3077836334705353 + ], + "opposite_logprob_tokens": [ + -1.432783603668213 + ], + "logprob_reward": -0.3077836334705353, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.3077836334705353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09556640684604645, + 0.16655106842517853, + 0.2073529213666916, + 0.060691360384225845, + 0.22461117804050446, + 0.060260333120822906, + 0.03390767052769661, + 0.02046855539083481, + 0.07978242635726929, + 0.05080803856253624 + ], + "credit_map": [ + 0.009905476123094559, + 0.0097499443218112, + 0.005694231949746609, + 0.006515986751765013, + 0.008711610920727253, + 0.020613813772797585, + 0.010607492178678513, + 0.023767853155732155, + 0.013260326348245144, + 0.014476110227406025, + 0.013674866408109665, + 0.013519265688955784, + 0.01949404738843441, + 0.0425170361995697, + 0.01629222370684147, + 0.033317193388938904, + 0.014590597711503506, + 0.011634107679128647, + 0.07472312450408936, + 0.009422294795513153, + 0.01015715952962637, + 0.03802596405148506, + 0.012909148819744587, + 0.03589051961898804, + 0.005216892343014479, + 0.004172358196228743, + 0.009316629730165005, + 0.004041304811835289, + 0.009412428364157677, + 0.016560683026909828, + 0.005924640689045191, + 0.0060464246198534966, + 0.011773835867643356, + 0.006007686723023653, + 0.07521388679742813, + 0.019640937447547913, + 0.02343852072954178, + 0.06552039831876755, + 0.00560256140306592, + 0.01741335727274418, + 0.0049355775117874146, + 0.005683582276105881, + 0.008300749585032463, + 0.007548666559159756, + 0.009625345468521118, + 0.007898719049990177, + 0.0062414128333330154, + 0.010026276111602783, + 0.0029860876966267824, + 0.003136437153443694, + 0.0047394344583153725, + 0.005045374855399132, + 0.003878566436469555, + 0.005680513102561235, + 0.0036909417249262333, + 0.004750315565615892, + 0.0019378839060664177, + 0.0016713717486709356, + 0.003324835328385234, + 0.002333087380975485, + 0.0017408269923180342, + 0.0018767010187730193, + 0.0030635250732302666, + 0.004520322661846876, + 0.007208815775811672, + 0.0076656281016767025, + 0.014019707217812538, + 0.011202216148376465, + 0.006396150682121515, + 0.008344634436070919, + 0.010723451152443886, + 0.014221825636923313, + 0.0050615351647138596, + 0.007465691305696964, + 0.009511197917163372, + 0.008184930309653282, + 0.004731766413897276, + 0.004081604536622763, + 0.007313517387956381, + 0.00445779412984848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.3077836334705353 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive humpback whale visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the whale's skin dark blue/black with visible barnacles and bioluminescent organisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the whale's bioluminescence in an otherwise dark environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005960494279861 + ], + "opposite_logprob_tokens": [ + -2.6950595378875732 + ], + "logprob_reward": -0.07005960494279861, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005960494279861, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09657254815101624, + 0.0899420753121376, + 0.1788957715034485, + 0.07577920705080032, + 0.2640146017074585, + 0.11573465168476105, + 0.04169077053666115, + 0.0359543077647686, + 0.02232196368277073, + 0.07909407466650009 + ], + "credit_map": [ + 0.008669409900903702, + 0.009480452165007591, + 0.007260835263878107, + 0.0077745323069393635, + 0.00649626087397337, + 0.017459647729992867, + 0.011405477300286293, + 0.028025934472680092, + 0.006522151175886393, + 0.011487231589853764, + 0.01239475142210722, + 0.011474951170384884, + 0.010523323900997639, + 0.012699401006102562, + 0.011457288637757301, + 0.0133829889819026, + 0.015717625617980957, + 0.009819943457841873, + 0.030276848003268242, + 0.007611257489770651, + 0.012974904850125313, + 0.040708255022764206, + 0.013283012434840202, + 0.048503924161195755, + 0.008719293400645256, + 0.0077562290243804455, + 0.010040421970188618, + 0.010363739915192127, + 0.009981737472116947, + 0.010010302998125553, + 0.009355326183140278, + 0.009552162140607834, + 0.013583119958639145, + 0.005741586443036795, + 0.08810030668973923, + 0.022632330656051636, + 0.027890192344784737, + 0.07953670620918274, + 0.006616293452680111, + 0.01991410180926323, + 0.010361658409237862, + 0.00865996815264225, + 0.015236026607453823, + 0.013723373413085938, + 0.019572608172893524, + 0.016638418659567833, + 0.014086425304412842, + 0.0174561757594347, + 0.00389162078499794, + 0.003212164156138897, + 0.006633699405938387, + 0.005762591492384672, + 0.004071737639605999, + 0.00557073624804616, + 0.005999017506837845, + 0.006549205631017685, + 0.004527031909674406, + 0.006205912679433823, + 0.005658593960106373, + 0.0038554396014660597, + 0.002706024330109358, + 0.0032161481212824583, + 0.00566868158057332, + 0.004116479307413101, + 0.0019652913324534893, + 0.0019609590526670218, + 0.0033644824288785458, + 0.0027517133858054876, + 0.0020560063421726227, + 0.002242222661152482, + 0.0028592469170689583, + 0.005122040864080191, + 0.007352736312896013, + 0.011088772676885128, + 0.015513760969042778, + 0.011058812029659748, + 0.009158832021057606, + 0.007887305691838264, + 0.009880343452095985, + 0.007153508719056845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.07005960494279861 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive humpback whale visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the whale's skin dark blue/black with visible barnacles and bioluminescent organisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the whale's bioluminescence in an otherwise dark environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036199060268700123 + ], + "opposite_logprob_tokens": [ + -5.62861967086792 + ], + "logprob_reward": -0.0036199060268700123, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036199060268700123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1008305475115776, + 0.08405148983001709, + 0.1910921335220337, + 0.11215707659721375, + 0.0293183084577322, + 0.14857012033462524, + 0.1870613694190979, + 0.04514586180448532, + 0.068704754114151, + 0.03306836634874344 + ], + "credit_map": [ + 0.009662652388215065, + 0.012272843159735203, + 0.008320865221321583, + 0.01091906800866127, + 0.008602363988757133, + 0.01899952068924904, + 0.009365733712911606, + 0.022687500342726707, + 0.008690946735441685, + 0.008259057998657227, + 0.007602647878229618, + 0.010687973350286484, + 0.012263215146958828, + 0.015549900010228157, + 0.01065050344914198, + 0.010347241535782814, + 0.014174266718327999, + 0.006833174731582403, + 0.041379909962415695, + 0.00663747638463974, + 0.015679165720939636, + 0.07193129509687424, + 0.011802714318037033, + 0.022654136642813683, + 0.0106139425188303, + 0.011736652813851833, + 0.017097989097237587, + 0.011176100932061672, + 0.015288139693439007, + 0.025340789929032326, + 0.011952456086874008, + 0.008951004594564438, + 0.003132430138066411, + 0.002722683595493436, + 0.00583527609705925, + 0.0034871199168264866, + 0.0020923439878970385, + 0.004425365012139082, + 0.004481980111449957, + 0.0031411093659698963, + 0.013047311455011368, + 0.01791665516793728, + 0.024039894342422485, + 0.024223316460847855, + 0.017747411504387856, + 0.02020738460123539, + 0.016678836196660995, + 0.014709308743476868, + 0.014941605739295483, + 0.02146441861987114, + 0.02062230370938778, + 0.020396485924720764, + 0.019014403223991394, + 0.0294977854937315, + 0.022819194942712784, + 0.038305167108774185, + 0.0035095082130283117, + 0.004018878564238548, + 0.007699363864958286, + 0.005550776142627001, + 0.004259353969246149, + 0.003582723205909133, + 0.0077250502072274685, + 0.008800210431218147, + 0.0064921011216938496, + 0.008601506240665913, + 0.013004401698708534, + 0.00688264612108469, + 0.0064243534579873085, + 0.00964200310409069, + 0.008068726398050785, + 0.009589013643562794, + 0.005102082155644894, + 0.0028245067223906517, + 0.005133000668138266, + 0.003971699625253677, + 0.004376077558845282, + 0.0030562717001885176, + 0.003287221072241664, + 0.005317509174346924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 5 + ], + "selected_reward": -0.0036199060268700123 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive humpback whale visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the whale's skin dark blue/black with visible barnacles and bioluminescent organisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the whale's bioluminescence in an otherwise dark environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016087723895907402 + ], + "opposite_logprob_tokens": [ + -4.141087532043457 + ], + "logprob_reward": -0.016087723895907402, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016087723895907402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16902635991573334, + 0.07692807912826538, + 0.18648532032966614, + 0.02901884727180004, + 0.16684049367904663, + 0.1417318433523178, + 0.026059454306960106, + 0.11901308596134186, + 0.028580859303474426, + 0.05631563439965248 + ], + "credit_map": [ + 0.00912181381136179, + 0.012230059131979942, + 0.007495279889553785, + 0.009812269359827042, + 0.009757639840245247, + 0.0404987558722496, + 0.02120758593082428, + 0.05890294536948204, + 0.006826843600720167, + 0.008542017079889774, + 0.00653487304225564, + 0.006410595029592514, + 0.008702791295945644, + 0.016489990055561066, + 0.009503637440502644, + 0.013917334377765656, + 0.019601136445999146, + 0.02000025473535061, + 0.05021928250789642, + 0.009922763332724571, + 0.011182785034179688, + 0.01742803119122982, + 0.016175735741853714, + 0.04195532947778702, + 0.0027285953983664513, + 0.0031674306374043226, + 0.0033594558481127024, + 0.002535139210522175, + 0.005276203621178865, + 0.00489759398624301, + 0.004416829906404018, + 0.0026375979650765657, + 0.011741172522306442, + 0.007451524026691914, + 0.04511298984289169, + 0.016590233892202377, + 0.01826995238661766, + 0.043821144849061966, + 0.011220221407711506, + 0.012633253820240498, + 0.012197484262287617, + 0.01144202146679163, + 0.022482505068182945, + 0.01849345490336418, + 0.017880424857139587, + 0.024129392579197884, + 0.011889670975506306, + 0.023216884583234787, + 0.00259521440602839, + 0.0028319861739873886, + 0.0029674391262233257, + 0.00274648517370224, + 0.0033644880168139935, + 0.003043506061658263, + 0.003800254547968507, + 0.004710081499069929, + 0.008882720023393631, + 0.010554144158959389, + 0.020723624154925346, + 0.01425629761070013, + 0.012784240767359734, + 0.00917404517531395, + 0.01825696974992752, + 0.02438104897737503, + 0.0029506715945899487, + 0.0020110628101974726, + 0.006843501701951027, + 0.003126106457784772, + 0.0021092358510941267, + 0.002185514895245433, + 0.0028214193880558014, + 0.006533343810588121, + 0.006994518917053938, + 0.009117201901972294, + 0.006203892640769482, + 0.006273254286497831, + 0.010186518542468548, + 0.0063733854331076145, + 0.005718190688639879, + 0.005448672920465469 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.016087723895907402 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive humpback whale visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the whale's skin dark blue/black with visible barnacles and bioluminescent organisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the whale's bioluminescence in an otherwise dark environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486948609352112 + ], + "opposite_logprob_tokens": [ + -1.2236948013305664 + ], + "logprob_reward": -0.3486948609352112, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.3486948609352112, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21264031529426575, + 0.07603080570697784, + 0.2836230397224426, + 0.02722005359828472, + 0.06565450131893158, + 0.13512884080410004, + 0.032638903707265854, + 0.04251610115170479, + 0.07732921838760376, + 0.04721824452280998 + ], + "credit_map": [ + 0.011023432947695255, + 0.01285523734986782, + 0.007520281709730625, + 0.008143862709403038, + 0.012320474721491337, + 0.053563717752695084, + 0.02875618450343609, + 0.07845713198184967, + 0.007393995299935341, + 0.007119426038116217, + 0.007061162032186985, + 0.006257287226617336, + 0.008082335814833641, + 0.015871964395046234, + 0.008582798764109612, + 0.01566183567047119, + 0.03020365536212921, + 0.023702139034867287, + 0.04651450738310814, + 0.011575527489185333, + 0.01801803894340992, + 0.06621789932250977, + 0.014828233048319817, + 0.0725630521774292, + 0.0027012124191969633, + 0.0022929443512111902, + 0.004366457927972078, + 0.002936112228780985, + 0.003383822739124298, + 0.00627734512090683, + 0.002638760255649686, + 0.002623398322612047, + 0.008327039889991283, + 0.0062761614099144936, + 0.010484423488378525, + 0.005682061426341534, + 0.007783604320138693, + 0.012599427253007889, + 0.00645009009167552, + 0.008051689714193344, + 0.01226795744150877, + 0.011476614512503147, + 0.02252521552145481, + 0.018841488286852837, + 0.016842195764183998, + 0.02477419003844261, + 0.011870547197759151, + 0.01653062365949154, + 0.00276183127425611, + 0.0033575347624719143, + 0.0047119976952672005, + 0.005059550516307354, + 0.003689501201733947, + 0.004757040645927191, + 0.003097956068813801, + 0.0052034915424883366, + 0.0035832813009619713, + 0.005889387335628271, + 0.007029595319181681, + 0.0054482934065163136, + 0.003552683861926198, + 0.004563948139548302, + 0.006462580990046263, + 0.005986328702419996, + 0.0090609360486269, + 0.006206268910318613, + 0.01544172316789627, + 0.009157931432127953, + 0.005854774732142687, + 0.006748244632035494, + 0.00933653861284256, + 0.01552280131727457, + 0.005900437943637371, + 0.008804630488157272, + 0.005178281106054783, + 0.005519872531294823, + 0.008361198008060455, + 0.00462486082687974, + 0.004894418176263571, + 0.003934547770768404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.3486948609352112 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot shot from an extremely low angle, the camera slowly and steadily tilts upward, following the majestic and deliberate rise of a massive humpback whale from the deep, inky depths of the ocean. The beast's skin is a deep, almost black blue, covered in ancient scars and patches of barnacles. But most breathtakingly, its entire body is covered in millions of tiny, symbiotic organisms, a flowing river of stars, emitting a soft, mysterious blue-green bioluminescence. In this abyssal depths, devoid of visible natural light, these bioluminescent organisms serve as the sole source of light. Their glow pierces the inky blackness, not only clearly outlining the whale's massive and graceful silhouette but also illuminating the tiny particles suspended around it, making them appear like cosmic dust. With uncanny grace, the whale breaks through the current, its massive body barely causing any turbulence as it steadily and slowly moves toward the unknown surface. The entire scene is shrouded in a sacred and tranquil atmosphere, creating a breathtaking sense of awe and an otherworldly, dreamlike beauty.", + "seed": 21021, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_12/index_8978_seed_21021_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.28382545528002084, + "vlm_logprob": -0.28382545528002084, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.28382545528002084, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive humpback whale visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the whale's skin dark blue/black with visible barnacles and bioluminescent organisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the whale's bioluminescence in an otherwise dark environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31125232577323914 + ], + "opposite_logprob_tokens": [ + -1.4362523555755615 + ], + "logprob_reward": -0.31125232577323914, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.31125232577323914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05753248557448387, + 0.03604942560195923, + 0.028331568464636803, + 0.05680253729224205, + 0.049457646906375885, + 0.08458314090967178, + 0.34706705808639526, + 0.2523440420627594, + 0.03922252729535103, + 0.048609618097543716 + ], + "credit_map": [ + 0.00629316596314311, + 0.007316627074033022, + 0.006906129419803619, + 0.007157676853239536, + 0.006216022651642561, + 0.0066000670194625854, + 0.008267536759376526, + 0.008775254711508751, + 0.003454074962064624, + 0.004913972690701485, + 0.0037592558655887842, + 0.0051027811132371426, + 0.004728961735963821, + 0.0029950698371976614, + 0.0061014131642878056, + 0.004993892274796963, + 0.002954823663458228, + 0.0025113457813858986, + 0.00375988963060081, + 0.0019543112721294165, + 0.004778623580932617, + 0.002202017931267619, + 0.0036542904563248158, + 0.0065162633545696735, + 0.005584260914474726, + 0.014060376212000847, + 0.006079781334847212, + 0.006522560957819223, + 0.006161443889141083, + 0.005691886879503727, + 0.005504895932972431, + 0.007197322323918343, + 0.005100048612803221, + 0.007692696526646614, + 0.00745739508420229, + 0.00818505510687828, + 0.004757014103233814, + 0.004558338783681393, + 0.005381613504141569, + 0.0063254875130951405, + 0.008014839142560959, + 0.011017552576959133, + 0.009859338402748108, + 0.009559967555105686, + 0.012483264319598675, + 0.009800933301448822, + 0.011845460161566734, + 0.012001775205135345, + 0.02537897229194641, + 0.0750647783279419, + 0.047119785100221634, + 0.035063713788986206, + 0.048285990953445435, + 0.035348307341337204, + 0.048639506101608276, + 0.03216596692800522, + 0.020628469064831734, + 0.04255649819970131, + 0.031659629195928574, + 0.039017338305711746, + 0.024435371160507202, + 0.02548699826002121, + 0.039069242775440216, + 0.02949046716094017, + 0.0029536501970142126, + 0.004451568238437176, + 0.00450780289247632, + 0.006744706071913242, + 0.00452338345348835, + 0.004219729453325272, + 0.007510730996727943, + 0.0043109506368637085, + 0.003079248359426856, + 0.007869753986597061, + 0.007287762593477964, + 0.00714501878246665, + 0.005630649160593748, + 0.004096238408237696, + 0.0073430719785392284, + 0.0061578694730997086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.31125232577323914 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive humpback whale visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the whale's skin dark blue/black with visible barnacles and bioluminescent organisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the whale's bioluminescence in an otherwise dark environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17992182075977325 + ], + "opposite_logprob_tokens": [ + -1.8049218654632568 + ], + "logprob_reward": -0.17992182075977325, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17992182075977325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07122062891721725, + 0.05096020549535751, + 0.07449639588594437, + 0.06873997300863266, + 0.04668182507157326, + 0.06509959697723389, + 0.1947154700756073, + 0.1433747559785843, + 0.15745876729488373, + 0.12725237011909485 + ], + "credit_map": [ + 0.008343013934791088, + 0.012231175787746906, + 0.010845843702554703, + 0.01043607760220766, + 0.0077249822206795216, + 0.007195982616394758, + 0.007575535215437412, + 0.006868019234389067, + 0.005913112312555313, + 0.006667394656687975, + 0.005433639511466026, + 0.00507762236520648, + 0.0046986984089016914, + 0.005386236123740673, + 0.009639400988817215, + 0.008144104853272438, + 0.007510353345423937, + 0.008347644470632076, + 0.0090012913569808, + 0.0049444278702139854, + 0.015458360314369202, + 0.005005472805351019, + 0.010505126789212227, + 0.013723728246986866, + 0.006733037065714598, + 0.01576862670481205, + 0.008097977377474308, + 0.006781199015676975, + 0.0075815292075276375, + 0.006673878058791161, + 0.009604220278561115, + 0.0074995094910264015, + 0.003454962745308876, + 0.005747949704527855, + 0.006523660384118557, + 0.007531248033046722, + 0.0072772265411913395, + 0.005337503273040056, + 0.005211521405726671, + 0.005597760435193777, + 0.009328775107860565, + 0.00802717823535204, + 0.008418160490691662, + 0.008410566486418247, + 0.009987201541662216, + 0.006191619671881199, + 0.009022867307066917, + 0.005713231395930052, + 0.017807012423872948, + 0.0313071683049202, + 0.025253616273403168, + 0.014992483891546726, + 0.027865329757332802, + 0.03141417354345322, + 0.021264977753162384, + 0.02481074258685112, + 0.013241911306977272, + 0.02327205240726471, + 0.016845256090164185, + 0.017903851345181465, + 0.017744049429893494, + 0.014891490340232849, + 0.02024713158607483, + 0.01922902837395668, + 0.012962406501173973, + 0.020567454397678375, + 0.01849324442446232, + 0.022799862548708916, + 0.01939988136291504, + 0.015679772943258286, + 0.02918749302625656, + 0.018368659541010857, + 0.009015575051307678, + 0.020187266170978546, + 0.020274806767702103, + 0.018460018560290337, + 0.014231930486857891, + 0.014524790458381176, + 0.01683400571346283, + 0.013723981566727161 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.17992182075977325 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive humpback whale visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the whale's skin dark blue/black with visible barnacles and bioluminescent organisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the whale's bioluminescence in an otherwise dark environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031966085080057383 + ], + "opposite_logprob_tokens": [ + -5.753196716308594 + ], + "logprob_reward": -0.0031966085080057383, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031966085080057383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13240407407283783, + 0.04603435844182968, + 0.04815196245908737, + 0.055457744747400284, + 0.17193111777305603, + 0.10888341069221497, + 0.2571946978569031, + 0.07397259026765823, + 0.0528283417224884, + 0.053141701966524124 + ], + "credit_map": [ + 0.012799341231584549, + 0.016097696498036385, + 0.014316432178020477, + 0.0184165146201849, + 0.014298936352133751, + 0.0186859630048275, + 0.019215036183595657, + 0.01857415959239006, + 0.0041519091464579105, + 0.004634452052414417, + 0.006675807759165764, + 0.004958314821124077, + 0.005909040104597807, + 0.004824589006602764, + 0.008247695863246918, + 0.00663255387917161, + 0.004274312872439623, + 0.011395363137125969, + 0.006385675631463528, + 0.004110828973352909, + 0.007104725576937199, + 0.0036244881339371204, + 0.005781087558716536, + 0.0054754819720983505, + 0.005125689320266247, + 0.00965788308531046, + 0.006679045967757702, + 0.005049842409789562, + 0.00716736912727356, + 0.005599908530712128, + 0.00884833000600338, + 0.007329675368964672, + 0.013818361796438694, + 0.03366474807262421, + 0.025315476581454277, + 0.022419249638915062, + 0.028718145564198494, + 0.02143935114145279, + 0.013243278488516808, + 0.01331249624490738, + 0.0072887735441327095, + 0.01544103492051363, + 0.017226524651050568, + 0.014016546308994293, + 0.015008201822638512, + 0.010732877068221569, + 0.013499663211405277, + 0.01566978543996811, + 0.018265558406710625, + 0.047130659222602844, + 0.03475230932235718, + 0.030826421454548836, + 0.035670891404151917, + 0.024176141247153282, + 0.028168389573693275, + 0.03820435330271721, + 0.00817179400473833, + 0.01203926745802164, + 0.009068424813449383, + 0.010525698773562908, + 0.008063659071922302, + 0.0067764404229819775, + 0.009845510125160217, + 0.009481790475547314, + 0.004342024680227041, + 0.006277798675000668, + 0.0063123139552772045, + 0.007430544123053551, + 0.006177169270813465, + 0.005646765697747469, + 0.009821632876992226, + 0.006820092909038067, + 0.004842768423259258, + 0.007202443201094866, + 0.008047749288380146, + 0.008468510583043098, + 0.0058914292603731155, + 0.0071024466305971146, + 0.005849991925060749, + 0.005736359860748053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.0031966085080057383 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive humpback whale visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the whale's skin dark blue/black with visible barnacles and bioluminescent organisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the whale's bioluminescence in an otherwise dark environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02637278474867344 + ], + "opposite_logprob_tokens": [ + -3.6513726711273193 + ], + "logprob_reward": -0.02637278474867344, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02637278474867344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16670633852481842, + 0.11399606615304947, + 0.05243150144815445, + 0.055433448404073715, + 0.07574524730443954, + 0.04237229377031326, + 0.1266183853149414, + 0.08356018364429474, + 0.044122498482465744, + 0.23901399970054626 + ], + "credit_map": [ + 0.010150526650249958, + 0.0252253208309412, + 0.018360279500484467, + 0.02398381382226944, + 0.019143449142575264, + 0.023710263893008232, + 0.020017579197883606, + 0.026115115731954575, + 0.007993925362825394, + 0.01808115281164646, + 0.010635231621563435, + 0.015904445201158524, + 0.019397126510739326, + 0.008917047642171383, + 0.020787499845027924, + 0.012279640883207321, + 0.005507887341082096, + 0.007062429562211037, + 0.006922062020748854, + 0.003951191436499357, + 0.009487200528383255, + 0.0036650870461016893, + 0.0060566021129488945, + 0.009779039770364761, + 0.007645582780241966, + 0.011927084997296333, + 0.007368926424533129, + 0.007084397133439779, + 0.006318245083093643, + 0.005486762151122093, + 0.0056978692300617695, + 0.003904579905793071, + 0.007905750535428524, + 0.014410917647182941, + 0.011535950936377048, + 0.009382758289575577, + 0.010322969406843185, + 0.009855215437710285, + 0.005418830085545778, + 0.006912857759743929, + 0.004834209568798542, + 0.004685656167566776, + 0.007362062577158213, + 0.005744631867855787, + 0.006510596722364426, + 0.00502042006701231, + 0.004122809506952763, + 0.004091908223927021, + 0.009880882687866688, + 0.023818131536245346, + 0.01530268881469965, + 0.01046944223344326, + 0.0198095440864563, + 0.014308201149106026, + 0.013250529766082764, + 0.01977897249162197, + 0.009030519053339958, + 0.011212167330086231, + 0.01013468112796545, + 0.011533461511135101, + 0.010427898727357388, + 0.008950488641858101, + 0.01176489982753992, + 0.010506060905754566, + 0.004088529851287603, + 0.006613757461309433, + 0.005719151813536882, + 0.007520594168454409, + 0.004088699817657471, + 0.004038556013256311, + 0.006966371554881334, + 0.005086835939437151, + 0.01702406257390976, + 0.017608005553483963, + 0.038808442652225494, + 0.042751237750053406, + 0.023363081738352776, + 0.05124228447675705, + 0.028128087520599365, + 0.020088814198970795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.02637278474867344 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive humpback whale visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the whale's skin dark blue/black with visible barnacles and bioluminescent organisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the whale's bioluminescence in an otherwise dark environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983837366104126 + ], + "opposite_logprob_tokens": [ + -0.5233837366104126 + ], + "logprob_reward": -0.8983837366104126, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983837366104126, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12630926072597504, + 0.05724024400115013, + 0.035439230501651764, + 0.0251481793820858, + 0.06485659629106522, + 0.07738596200942993, + 0.11235623806715012, + 0.09809320420026779, + 0.14331947267055511, + 0.2598516643047333 + ], + "credit_map": [ + 0.017344195395708084, + 0.014348558150231838, + 0.01566282846033573, + 0.01480892300605774, + 0.012314190156757832, + 0.01484635192900896, + 0.02327522449195385, + 0.013708991929888725, + 0.005634211003780365, + 0.007300744764506817, + 0.0070669022388756275, + 0.008312088437378407, + 0.007850603200495243, + 0.005595376715064049, + 0.008498906157910824, + 0.006981412880122662, + 0.0041124033741652966, + 0.005040193907916546, + 0.00464665750041604, + 0.003258339362218976, + 0.005858832970261574, + 0.0035377617459744215, + 0.00479623768478632, + 0.004188804887235165, + 0.002289092866703868, + 0.00396727304905653, + 0.003335474757477641, + 0.0025306683965027332, + 0.0028543025255203247, + 0.0025751173961907625, + 0.0035775485448539257, + 0.004018702544271946, + 0.00710639264434576, + 0.007140622474253178, + 0.00732463551685214, + 0.007823937572538853, + 0.009912666864693165, + 0.005883576348423958, + 0.00696111936122179, + 0.012703647837042809, + 0.005774531979113817, + 0.010434544645249844, + 0.010762802325189114, + 0.009691672399640083, + 0.009673953987658024, + 0.008582182228565216, + 0.010479338467121124, + 0.011986934579908848, + 0.011530493386089802, + 0.015618267469108105, + 0.015462742187082767, + 0.009545817039906979, + 0.014345629140734673, + 0.019032156094908714, + 0.01328310277312994, + 0.01353803277015686, + 0.009477769024670124, + 0.014904934912919998, + 0.013187473639845848, + 0.011878135614097118, + 0.013000210747122765, + 0.014942185021936893, + 0.01087480690330267, + 0.009827680885791779, + 0.009447700344026089, + 0.017119383439421654, + 0.018862977623939514, + 0.019661545753479004, + 0.020932205021381378, + 0.01812417060136795, + 0.025420352816581726, + 0.013751129619777203, + 0.021934034302830696, + 0.029206782579421997, + 0.03503923863172531, + 0.04384106025099754, + 0.03532424569129944, + 0.031209923326969147, + 0.03382603079080582, + 0.029470320791006088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.8983837366104126 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "shot shot from an extremely low angle, the camera slowly and steadily tilts upward, following the majestic and deliberate rise of a massive humpback whale from the deep, inky depths of the ocean. The beast's skin is a deep, almost black blue, covered in ancient scars and patches of barnacles. But most breathtakingly, its entire body is covered in millions of tiny, symbiotic organisms, a flowing river of stars, emitting a soft, mysterious blue-green bioluminescence. In this abyssal depths, devoid of visible natural light, these bioluminescent organisms serve as the sole source of light. Their glow pierces the inky blackness, not only clearly outlining the whale's massive and graceful silhouette but also illuminating the tiny particles suspended around it, making them appear like cosmic dust. With uncanny grace, the whale breaks through the current, its massive body barely causing any turbulence as it steadily and slowly moves toward the unknown surface. The entire scene is shrouded in a sacred and tranquil atmosphere, creating a breathtaking sense of awe and an otherworldly, dreamlike beauty.", + "seed": 21022, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_12/index_8978_seed_21022_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2938844673335552, + "vlm_logprob": -0.2938844673335552, + "vlm_yes_no_margin": 2.125, + "vlm_token_credit": -0.2938844673335552, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive humpback whale visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the whale's skin dark blue/black with visible barnacles and bioluminescent organisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the whale's bioluminescence in an otherwise dark environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2869255244731903 + ], + "opposite_logprob_tokens": [ + -1.5369255542755127 + ], + "logprob_reward": -0.2869255244731903, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2869255244731903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06360358744859695, + 0.11860602349042892, + 0.07636505365371704, + 0.5285332798957825, + 0.019227655604481697, + 0.021769287064671516, + 0.01684672385454178, + 0.08305902034044266, + 0.03611791133880615, + 0.03587139770388603 + ], + "credit_map": [ + 0.0068535576574504375, + 0.006936416029930115, + 0.007200976833701134, + 0.005603007972240448, + 0.007921789772808552, + 0.013901851139962673, + 0.00561801390722394, + 0.009567974135279655, + 0.009193232282996178, + 0.00589363370090723, + 0.006694583222270012, + 0.006068774964660406, + 0.008643798530101776, + 0.04308653250336647, + 0.00349099631421268, + 0.03553446754813194, + 0.008056110702455044, + 0.016554255038499832, + 0.009149634279310703, + 0.007133721839636564, + 0.009879117831587791, + 0.01074790395796299, + 0.006535152904689312, + 0.008309158496558666, + 0.05195075273513794, + 0.042365048080682755, + 0.02784590981900692, + 0.02308955229818821, + 0.04601915925741196, + 0.17203277349472046, + 0.03636110574007034, + 0.12886899709701538, + 0.0013791874516755342, + 0.0019093373557552695, + 0.0013355257688090205, + 0.0017275616992264986, + 0.0017784963129088283, + 0.004008826334029436, + 0.0030494588427245617, + 0.00403926195576787, + 0.0012787353480234742, + 0.002597293583676219, + 0.0017881459789350629, + 0.001872932887636125, + 0.0022122482769191265, + 0.0034377779811620712, + 0.00515418965369463, + 0.0034279630053788424, + 0.0018608449026942253, + 0.0014850795269012451, + 0.001904487842693925, + 0.002694498747587204, + 0.0026307045482099056, + 0.0022758301347494125, + 0.0017304072389379144, + 0.002264870097860694, + 0.007591533940285444, + 0.00809730589389801, + 0.006874068174511194, + 0.009320421144366264, + 0.014875160530209541, + 0.011848952621221542, + 0.01277906447649002, + 0.01167251355946064, + 0.002989997388795018, + 0.003891268977895379, + 0.005995620042085648, + 0.004347960464656353, + 0.006725704297423363, + 0.005043468438088894, + 0.0036652740091085434, + 0.0034586184192448854, + 0.0028545644599944353, + 0.004989166744053364, + 0.004503902047872543, + 0.005023973062634468, + 0.003790658200159669, + 0.007052746135741472, + 0.003576202318072319, + 0.004080182407051325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 7 + ], + "selected_reward": -0.2869255244731903 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive humpback whale visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the whale's skin dark blue/black with visible barnacles and bioluminescent organisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the whale's bioluminescence in an otherwise dark environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521466910839081 + ], + "opposite_logprob_tokens": [ + -1.5021467208862305 + ], + "logprob_reward": -0.2521466910839081, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521466910839081, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10411816090345383, + 0.10696728527545929, + 0.06158587336540222, + 0.05150476098060608, + 0.09072072058916092, + 0.13238625228405, + 0.07434765249490738, + 0.15620994567871094, + 0.15162162482738495, + 0.0705377459526062 + ], + "credit_map": [ + 0.012207300402224064, + 0.009616591967642307, + 0.01031466107815504, + 0.011601084843277931, + 0.014431838877499104, + 0.022792955860495567, + 0.006482827011495829, + 0.01667090319097042, + 0.008856356143951416, + 0.010042790323495865, + 0.01188149955123663, + 0.013120753690600395, + 0.01922539621591568, + 0.018264813348650932, + 0.012663708999752998, + 0.012911972589790821, + 0.006146132946014404, + 0.008206824772059917, + 0.007734565529972315, + 0.007872129790484905, + 0.006118051242083311, + 0.010320593602955341, + 0.005345797631889582, + 0.009841779246926308, + 0.004805128555744886, + 0.007272328715771437, + 0.007522439118474722, + 0.005415832623839378, + 0.006231551058590412, + 0.006402303464710712, + 0.004401698242872953, + 0.009453478269279003, + 0.005977426189929247, + 0.011720151640474796, + 0.010463004931807518, + 0.009387888945639133, + 0.007071215659379959, + 0.01053058635443449, + 0.01373465545475483, + 0.021835792809724808, + 0.004853059072047472, + 0.011254229582846165, + 0.008171570487320423, + 0.00791353639215231, + 0.009072705172002316, + 0.01273181289434433, + 0.04095516726374626, + 0.037434156984090805, + 0.006916843354701996, + 0.00983910821378231, + 0.007179524749517441, + 0.011152678169310093, + 0.008271894417703152, + 0.010338717140257359, + 0.009102804586291313, + 0.01154608279466629, + 0.016582246869802475, + 0.012844893150031567, + 0.01953764818608761, + 0.020953522995114326, + 0.015798354521393776, + 0.01989731378853321, + 0.029067886993288994, + 0.021528083831071854, + 0.012188625521957874, + 0.01637508161365986, + 0.01808825321495533, + 0.0164958406239748, + 0.029291942715644836, + 0.03180433809757233, + 0.0124745462089777, + 0.014902990311384201, + 0.005869613029062748, + 0.005153256934136152, + 0.01079825684428215, + 0.012120471335947514, + 0.007460484281182289, + 0.010805916041135788, + 0.011096660047769547, + 0.007233079057186842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.2521466910839081 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive humpback whale visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the whale's skin dark blue/black with visible barnacles and bioluminescent organisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the whale's bioluminescence in an otherwise dark environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640285581350327 + ], + "opposite_logprob_tokens": [ + -4.758640289306641 + ], + "logprob_reward": -0.008640285581350327, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640285581350327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31276071071624756, + 0.152112677693367, + 0.04054876044392586, + 0.17482979595661163, + 0.014788175001740456, + 0.057415589690208435, + 0.027638981118798256, + 0.09999633580446243, + 0.10435838252305984, + 0.015550562180578709 + ], + "credit_map": [ + 0.03365904465317726, + 0.02049889974296093, + 0.026498304679989815, + 0.017446788027882576, + 0.03668878600001335, + 0.09502072632312775, + 0.018904494121670723, + 0.06404366344213486, + 0.01163590420037508, + 0.007385779172182083, + 0.008859495632350445, + 0.005456339567899704, + 0.010121405124664307, + 0.05737612023949623, + 0.004377815872430801, + 0.04689982905983925, + 0.0034134795423597097, + 0.005583937279880047, + 0.004415951203554869, + 0.005558103788644075, + 0.003705107606947422, + 0.007322041783481836, + 0.004666585940867662, + 0.005883555393666029, + 0.016180766746401787, + 0.016744958236813545, + 0.011024056933820248, + 0.010366957634687424, + 0.015553416684269905, + 0.05193740129470825, + 0.012205048464238644, + 0.040817197412252426, + 0.0010359593434259295, + 0.0027607078664004803, + 0.0019177438225597143, + 0.002065944718196988, + 0.0016037027817219496, + 0.0019034924916923046, + 0.0015918646240606904, + 0.0019087591208517551, + 0.0021114745177328587, + 0.00452381232753396, + 0.0038551483303308487, + 0.0037523761857300997, + 0.0031967987306416035, + 0.004795898217707872, + 0.018287474289536476, + 0.016892606392502785, + 0.0030805913265794516, + 0.004170314874500036, + 0.003074716543778777, + 0.0032360025215893984, + 0.003568035550415516, + 0.0036712419241666794, + 0.003303466597571969, + 0.003534612711519003, + 0.008823171257972717, + 0.010670186020433903, + 0.008531952276825905, + 0.013310045935213566, + 0.014200404286384583, + 0.015317687764763832, + 0.016586989164352417, + 0.012555901892483234, + 0.009427833370864391, + 0.010608065873384476, + 0.01526082493364811, + 0.013881436549127102, + 0.015300877392292023, + 0.018008777871727943, + 0.010972078889608383, + 0.010898484848439693, + 0.0014885741984471679, + 0.0014554469380527735, + 0.0020732043776661158, + 0.0018282587407156825, + 0.0021637901663780212, + 0.003395628184080124, + 0.0014633305836468935, + 0.0016823287587612867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.008640285581350327 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive humpback whale visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the whale's skin dark blue/black with visible barnacles and bioluminescent organisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the whale's bioluminescence in an otherwise dark environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023316025733947754 + ], + "opposite_logprob_tokens": [ + -3.773315906524658 + ], + "logprob_reward": -0.023316025733947754, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023316025733947754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08925291895866394, + 0.06297993659973145, + 0.05572415515780449, + 0.4194188714027405, + 0.03880472481250763, + 0.05334752798080444, + 0.03540503606200218, + 0.09918863326311111, + 0.08168214559555054, + 0.06419598311185837 + ], + "credit_map": [ + 0.009847821667790413, + 0.009187777526676655, + 0.013496513478457928, + 0.010151795111596584, + 0.00931420736014843, + 0.01924201101064682, + 0.006097304169088602, + 0.011915495619177818, + 0.006272764876484871, + 0.004505285061895847, + 0.007948381826281548, + 0.005546536762267351, + 0.005716246087104082, + 0.015820356085896492, + 0.005897418595850468, + 0.011272944509983063, + 0.005338456481695175, + 0.006703447084873915, + 0.00732621643692255, + 0.006115449592471123, + 0.00632497388869524, + 0.009136056527495384, + 0.0073251319117844105, + 0.007454423699527979, + 0.045340798795223236, + 0.030525928363204002, + 0.02079257182776928, + 0.018756462261080742, + 0.0416128970682621, + 0.13545307517051697, + 0.03285830840468407, + 0.09407883882522583, + 0.0028786694165319204, + 0.008000281639397144, + 0.004293088801205158, + 0.0054507083259522915, + 0.005556229967623949, + 0.0037449246738106012, + 0.004266600124537945, + 0.00461422186344862, + 0.003897403832525015, + 0.00358993629924953, + 0.0035947079304605722, + 0.0033637462183833122, + 0.002562257694080472, + 0.004305495414882898, + 0.016067422926425934, + 0.01596655882894993, + 0.004000689834356308, + 0.004182328935712576, + 0.0031795573886483908, + 0.004136195406317711, + 0.004043580032885075, + 0.0047164843417704105, + 0.006276896223425865, + 0.004869303200393915, + 0.011142891831696033, + 0.010335661470890045, + 0.008891960605978966, + 0.010304545983672142, + 0.017051400616765022, + 0.013589225709438324, + 0.012924094684422016, + 0.014948857016861439, + 0.0062621100805699825, + 0.008366003632545471, + 0.012818384915590286, + 0.00886430311948061, + 0.012218127027153969, + 0.015277695842087269, + 0.007424426730722189, + 0.010451101697981358, + 0.005542164668440819, + 0.006259702146053314, + 0.007517318241298199, + 0.00794822908937931, + 0.011360461823642254, + 0.010771836154162884, + 0.008450611494481564, + 0.006345661822706461 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.023316025733947754 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive humpback whale visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the whale's skin dark blue/black with visible barnacles and bioluminescent organisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the whale's bioluminescence in an otherwise dark environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983938097953796 + ], + "opposite_logprob_tokens": [ + -0.5233938097953796 + ], + "logprob_reward": -0.8983938097953796, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983938097953796, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07915526628494263, + 0.018642090260982513, + 0.06372447311878204, + 0.5111016035079956, + 0.029789132997393608, + 0.02115989662706852, + 0.012873265892267227, + 0.10525017976760864, + 0.04388948529958725, + 0.11441462486982346 + ], + "credit_map": [ + 0.00923396646976471, + 0.005804181564599276, + 0.004914699587970972, + 0.005026521626859903, + 0.00731703033670783, + 0.02403445541858673, + 0.004959616344422102, + 0.017864791676402092, + 0.0017585932509973645, + 0.0020861055236309767, + 0.001856456627137959, + 0.0015452896477654576, + 0.004521453287452459, + 0.002993566682562232, + 0.002355879172682762, + 0.0015247464179992676, + 0.005553496070206165, + 0.006333623081445694, + 0.006796479225158691, + 0.006710036192089319, + 0.012377733364701271, + 0.007807367015630007, + 0.01045907661318779, + 0.007686658296734095, + 0.048754576593637466, + 0.03876632824540138, + 0.028112810105085373, + 0.0214205514639616, + 0.045537568628787994, + 0.16899466514587402, + 0.0346473790705204, + 0.12486770004034042, + 0.002221854403614998, + 0.005058214999735355, + 0.003169147064909339, + 0.0028227269649505615, + 0.0029095576610416174, + 0.004905987065285444, + 0.00361924828030169, + 0.005082397256046534, + 0.0016767772613093257, + 0.0030714301392436028, + 0.0028256913647055626, + 0.0027583602350205183, + 0.002121260156854987, + 0.001811557449400425, + 0.003852884052321315, + 0.003041936317458749, + 0.0014707734808325768, + 0.001895265537314117, + 0.001202746876515448, + 0.0015560807660222054, + 0.0017996605020016432, + 0.0016096378676593304, + 0.0016870960826054215, + 0.001652005361393094, + 0.010442633181810379, + 0.007727930322289467, + 0.012120017781853676, + 0.01289437897503376, + 0.02016054280102253, + 0.015698986127972603, + 0.015131764113903046, + 0.011073930189013481, + 0.003436559345573187, + 0.0048411390744149685, + 0.005320229567587376, + 0.005176043603569269, + 0.006768633611500263, + 0.007509747985750437, + 0.005209928844124079, + 0.005627203732728958, + 0.00988837517797947, + 0.010723579674959183, + 0.014355842024087906, + 0.014495272189378738, + 0.017195813357830048, + 0.01890536956489086, + 0.016840824857354164, + 0.012009548023343086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 7 + ], + "selected_reward": -0.8983938097953796 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_13/index_tensor([1338, 1338])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_13/index_tensor([1338, 1338])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..47c53e56ba82cc879b3c7ec89e4582109137a0ee --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_13/index_tensor([1338, 1338])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot in a flowing watercolor animation style, a low-angle shot with a subtle, organic drift of the handheld camera gazes reverently upward at a majestic, ancient tree. Its massive, gnarled trunk, rendered in washes of deep brown and mossy green, twists powerfully toward the sky, its vast canopy energized by a gentle mountain breeze. Broad, hand-shaped leaves rustle and dance in an enchanting rhythm, their forms defined by a vibrant tapestry of soft, emerald and ochre pigments delicately intermingling. Rhythmically swaying branches are hung with spiky, yellow-green shells, their sharp outlines softened by a painterly beauty. Golden afternoon light majestically filters through the leaves, casting speckled, shifting patterns of warm glow that make them appear translucent. Through gaps in the canopy, glimpses of the deep blue sky peek through, its color seamlessly flowing into the greens and yellows of the foliage, creating a moving, serene, painterly tableau, each element swaying in a calm, dreamlike visual poetry.", + "seed": 6928, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_13/index_1338_seed_6928_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.059167839307338, + "vlm_logprob": -0.059167839307338, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.059167839307338, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a flowing watercolor animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle looking upward at the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree trunk massive, gnarled, and colored with deep brown and mossy green washes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are spiky yellow-green shells hanging from the swaying branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does golden afternoon light filter through the leaves creating speckled, shifting patterns of warm glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0043362402357161045 + ], + "opposite_logprob_tokens": [ + -5.879336357116699 + ], + "logprob_reward": -0.0043362402357161045, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0043362402357161045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07006935775279999, + 0.04755710810422897, + 0.2527868151664734, + 0.0820625051856041, + 0.037255171686410904, + 0.043633174151182175, + 0.038735296577215195, + 0.22204291820526123, + 0.07941154390573502, + 0.12644603848457336 + ], + "credit_map": [ + 0.012320222333073616, + 0.01063550915569067, + 0.011336550116539001, + 0.009937400929629803, + 0.008695254102349281, + 0.003816072829067707, + 0.007137271575629711, + 0.006191079039126635, + 0.005265283863991499, + 0.012483984231948853, + 0.004873983096331358, + 0.005187745671719313, + 0.005064216908067465, + 0.004639300052076578, + 0.00705675920471549, + 0.002985834376886487, + 0.030997231602668762, + 0.09468306601047516, + 0.00988322775810957, + 0.009393041022121906, + 0.01413232646882534, + 0.011164262890815735, + 0.06687091290950775, + 0.01566276326775551, + 0.0118667958304286, + 0.010502524673938751, + 0.031808190047740936, + 0.006760852411389351, + 0.006015075370669365, + 0.003557733492925763, + 0.0062975091859698296, + 0.00525383185595274, + 0.005421211943030357, + 0.004023957531899214, + 0.004485880024731159, + 0.006318708416074514, + 0.0053025828674435616, + 0.002392915543168783, + 0.006649637594819069, + 0.0026602777652442455, + 0.006615892518311739, + 0.00510051753371954, + 0.004795638378709555, + 0.005985176656395197, + 0.004728824831545353, + 0.0036150768864899874, + 0.0071645379066467285, + 0.005627509672194719, + 0.006681618746370077, + 0.004736024886369705, + 0.005706008058041334, + 0.008339953608810902, + 0.0029248918872326612, + 0.004220546223223209, + 0.0036222224589437246, + 0.0025040307082235813, + 0.053248222917318344, + 0.011286371387541294, + 0.013109950348734856, + 0.017088335007429123, + 0.02144380286335945, + 0.07917051762342453, + 0.008597243577241898, + 0.018098492175340652, + 0.008540600538253784, + 0.024750426411628723, + 0.004893757868558168, + 0.009382438845932484, + 0.005932492669671774, + 0.01618299074470997, + 0.005616304464638233, + 0.0041125379502773285, + 0.019277555868029594, + 0.01881011389195919, + 0.020404431968927383, + 0.019018065184354782, + 0.017655014991760254, + 0.01037577260285616, + 0.010775732807815075, + 0.010129349306225777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 9 + ], + "selected_reward": -0.0043362402357161045 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a flowing watercolor animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle looking upward at the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree trunk massive, gnarled, and colored with deep brown and mossy green washes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are spiky yellow-green shells hanging from the swaying branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does golden afternoon light filter through the leaves creating speckled, shifting patterns of warm glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25215455889701843 + ], + "opposite_logprob_tokens": [ + -1.5021545886993408 + ], + "logprob_reward": -0.25215455889701843, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25215455889701843, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09837470203638077, + 0.047297291457653046, + 0.08768017590045929, + 0.16499686241149902, + 0.02373037114739418, + 0.035029973834753036, + 0.04686814174056053, + 0.06421709060668945, + 0.07433971762657166, + 0.3574656844139099 + ], + "credit_map": [ + 0.011332031339406967, + 0.01903088390827179, + 0.014500509947538376, + 0.013635851442813873, + 0.012014812789857388, + 0.005878044757992029, + 0.013282477855682373, + 0.008700091391801834, + 0.003561631543561816, + 0.006955810356885195, + 0.007325548212975264, + 0.009137785993516445, + 0.00523879611864686, + 0.0027704585809260607, + 0.007951688952744007, + 0.004355564247816801, + 0.008295644074678421, + 0.024924540892243385, + 0.004890678450465202, + 0.0047303750179708, + 0.008650346659123898, + 0.0052500078454613686, + 0.022173868492245674, + 0.008764703758060932, + 0.03005489520728588, + 0.021108528599143028, + 0.06560243666172028, + 0.011883148923516273, + 0.011647302657365799, + 0.005971989128738642, + 0.011482113040983677, + 0.007246446330100298, + 0.004477691370993853, + 0.003821720601990819, + 0.003090338548645377, + 0.0035891402512788773, + 0.002407374791800976, + 0.001618163543753326, + 0.0024634047877043486, + 0.0022625348065048456, + 0.005250014364719391, + 0.003796815872192383, + 0.004705803003162146, + 0.00433716643601656, + 0.004087742418050766, + 0.0027630385011434555, + 0.006536229979246855, + 0.003553160931915045, + 0.008115681819617748, + 0.004485047422349453, + 0.004102467559278011, + 0.00934441015124321, + 0.005820881109684706, + 0.008561226539313793, + 0.003395552048459649, + 0.0030428734607994556, + 0.014608899131417274, + 0.003933038096874952, + 0.0032007875852286816, + 0.004436418414115906, + 0.006806863006204367, + 0.02218521572649479, + 0.003517627716064453, + 0.005528238136321306, + 0.006404259242117405, + 0.008355123922228813, + 0.007597906515002251, + 0.011475978419184685, + 0.006677184719592333, + 0.013523395173251629, + 0.015692908316850662, + 0.00461295573040843, + 0.07057277113199234, + 0.0392131544649601, + 0.06383046507835388, + 0.060282908380031586, + 0.033603567630052567, + 0.025326916947960854, + 0.030642157420516014, + 0.033993713557720184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.25215455889701843 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a flowing watercolor animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle looking upward at the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree trunk massive, gnarled, and colored with deep brown and mossy green washes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are spiky yellow-green shells hanging from the swaying branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does golden afternoon light filter through the leaves creating speckled, shifting patterns of warm glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639931678771973 + ], + "opposite_logprob_tokens": [ + -4.758639812469482 + ], + "logprob_reward": -0.008639931678771973, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639931678771973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07239586859941483, + 0.02710818126797676, + 0.13818049430847168, + 0.036394257098436356, + 0.0255501177161932, + 0.1959797590970993, + 0.036905333399772644, + 0.31914326548576355, + 0.04189509525895119, + 0.10644759982824326 + ], + "credit_map": [ + 0.007597026880830526, + 0.01025979220867157, + 0.00844865757972002, + 0.01201354805380106, + 0.00985962338745594, + 0.005213114432990551, + 0.010736546479165554, + 0.00826756376773119, + 0.002607606118544936, + 0.0066335625015199184, + 0.002321922918781638, + 0.0031746800523251295, + 0.0026839044876396656, + 0.0024103038012981415, + 0.004693044815212488, + 0.002583156805485487, + 0.01779445819556713, + 0.05010484531521797, + 0.004409885499626398, + 0.00524042546749115, + 0.008451717905700207, + 0.006492637563496828, + 0.03575963154435158, + 0.009926906786859035, + 0.004203306045383215, + 0.004463902674615383, + 0.004845974966883659, + 0.003675964428111911, + 0.0054467166773974895, + 0.004233269952237606, + 0.004254511557519436, + 0.005270610563457012, + 0.003305431455373764, + 0.003764885012060404, + 0.0029497689101845026, + 0.004087858833372593, + 0.0028164489194750786, + 0.0024553651455789804, + 0.0037250712048262358, + 0.0024452898651361465, + 0.030200151726603508, + 0.03314683586359024, + 0.019746830686926842, + 0.027523398399353027, + 0.026065191254019737, + 0.014606231823563576, + 0.026177698746323586, + 0.018513431772589684, + 0.004964825231581926, + 0.0032701794989407063, + 0.004570451099425554, + 0.009266606532037258, + 0.003506773617118597, + 0.005272727459669113, + 0.0031357314437627792, + 0.002918040147051215, + 0.07790438085794449, + 0.016865478828549385, + 0.017603695392608643, + 0.022850461304187775, + 0.028260525315999985, + 0.11705204099416733, + 0.011165683157742023, + 0.02744101546704769, + 0.004544077441096306, + 0.005123162642121315, + 0.0049265543930232525, + 0.006695491727441549, + 0.003211350878700614, + 0.008785043843090534, + 0.005990417208522558, + 0.002618999220430851, + 0.0174674391746521, + 0.015224218368530273, + 0.01772310957312584, + 0.0163677167147398, + 0.01252820249646902, + 0.009031695313751698, + 0.010020202957093716, + 0.008085020817816257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 2 + ], + "selected_reward": -0.008639931678771973 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a flowing watercolor animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle looking upward at the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree trunk massive, gnarled, and colored with deep brown and mossy green washes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are spiky yellow-green shells hanging from the swaying branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does golden afternoon light filter through the leaves creating speckled, shifting patterns of warm glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012530408799648285 + ], + "opposite_logprob_tokens": [ + -4.387530326843262 + ], + "logprob_reward": -0.012530408799648285, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012530408799648285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.032821815460920334, + 0.036239635199308395, + 0.5005272626876831, + 0.05586700513958931, + 0.023187173530459404, + 0.031215805560350418, + 0.0481901690363884, + 0.09992929548025131, + 0.0979977399110794, + 0.07402411103248596 + ], + "credit_map": [ + 0.0034668061416596174, + 0.004339038394391537, + 0.004839763510972261, + 0.005529647693037987, + 0.004842459224164486, + 0.0019786593038588762, + 0.0046259197406470776, + 0.003199517959728837, + 0.0034459952730685472, + 0.008680733852088451, + 0.004705617669969797, + 0.0059334877878427505, + 0.002821797039359808, + 0.0013126382837072015, + 0.007275955751538277, + 0.0020634091924875975, + 0.06624176353216171, + 0.18685060739517212, + 0.02423934079706669, + 0.020671505481004715, + 0.030894635245203972, + 0.02064391039311886, + 0.12343081831932068, + 0.027554692700505257, + 0.006610665936022997, + 0.007202087435871363, + 0.008318489417433739, + 0.003918162547051907, + 0.009765563532710075, + 0.005866864696145058, + 0.005817462224513292, + 0.008367707952857018, + 0.00392162473872304, + 0.0023229538928717375, + 0.002692754380404949, + 0.0020825525280088186, + 0.0035708316136151552, + 0.002021843334659934, + 0.0023738164454698563, + 0.0042007965967059135, + 0.004690267611294985, + 0.004462249111384153, + 0.003390359925106168, + 0.004809003323316574, + 0.003780927276238799, + 0.003435496473684907, + 0.003718498395755887, + 0.002929004142060876, + 0.009555916301906109, + 0.004989910870790482, + 0.004867799114435911, + 0.006028051488101482, + 0.004488273989409208, + 0.011779178865253925, + 0.002865070244297385, + 0.003615968395024538, + 0.02353968657553196, + 0.0051039098761975765, + 0.00464273989200592, + 0.006723719649016857, + 0.009166013449430466, + 0.037229668349027634, + 0.004296289756894112, + 0.009227261878550053, + 0.010681808926165104, + 0.030653376132249832, + 0.005852059926837683, + 0.014804198406636715, + 0.010219558142125607, + 0.012358125299215317, + 0.010443123057484627, + 0.002985486527904868, + 0.011654547415673733, + 0.010721122846007347, + 0.012676646932959557, + 0.011365567333996296, + 0.009189137257635593, + 0.006180766969919205, + 0.006788047030568123, + 0.0054482705891132355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 8 + ], + "selected_reward": -0.012530408799648285 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a flowing watercolor animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle looking upward at the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree trunk massive, gnarled, and colored with deep brown and mossy green washes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are spiky yellow-green shells hanging from the swaying branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does golden afternoon light filter through the leaves creating speckled, shifting patterns of warm glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018178056925535202 + ], + "opposite_logprob_tokens": [ + -4.0181779861450195 + ], + "logprob_reward": -0.018178056925535202, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018178056925535202, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16404053568840027, + 0.05409326031804085, + 0.14675630629062653, + 0.03664860874414444, + 0.023142991587519646, + 0.0855557769536972, + 0.050602659583091736, + 0.1677885204553604, + 0.13291706144809723, + 0.13845428824424744 + ], + "credit_map": [ + 0.019550658762454987, + 0.030909638851881027, + 0.021377725526690483, + 0.02735123038291931, + 0.017938198521733284, + 0.009053884074091911, + 0.023946646600961685, + 0.013912541791796684, + 0.004086862318217754, + 0.009749436751008034, + 0.006921781692653894, + 0.008179253898561, + 0.006371601019054651, + 0.005595867056399584, + 0.008127903565764427, + 0.005060557276010513, + 0.018518999218940735, + 0.05184612423181534, + 0.006036717910319567, + 0.005612450651824474, + 0.008440681733191013, + 0.009211474098265171, + 0.03553669527173042, + 0.011553148739039898, + 0.00462542288005352, + 0.005495384801179171, + 0.008062244392931461, + 0.0038319360464811325, + 0.0032160100527107716, + 0.00391659839078784, + 0.003905913792550564, + 0.0035950983874499798, + 0.002338741673156619, + 0.002680859062820673, + 0.002551795681938529, + 0.004165175370872021, + 0.0032115443609654903, + 0.0021786114666610956, + 0.003370093647390604, + 0.002646170323714614, + 0.011006588116288185, + 0.009624899365007877, + 0.00885591097176075, + 0.009279368445277214, + 0.007666619960218668, + 0.007840792648494244, + 0.014988870359957218, + 0.016292722895741463, + 0.007788714487105608, + 0.0051155840046703815, + 0.006036573555320501, + 0.006987429689615965, + 0.006840733345597982, + 0.008969157934188843, + 0.004050159361213446, + 0.004814307205379009, + 0.04023546725511551, + 0.008600889705121517, + 0.010871521197259426, + 0.013409086503088474, + 0.01606777124106884, + 0.05840011686086655, + 0.00695403665304184, + 0.013249623589217663, + 0.013398963958024979, + 0.03584878891706467, + 0.010807197540998459, + 0.019377944990992546, + 0.013177755288779736, + 0.014349017292261124, + 0.02099955640733242, + 0.0049578468315303326, + 0.021444397047162056, + 0.020046476274728775, + 0.026801487430930138, + 0.021173207089304924, + 0.018072884529829025, + 0.009971769526600838, + 0.010988832451403141, + 0.009955235756933689 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.018178056925535202 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in a flowing watercolor animation style, a low-angle shot with a subtle, organic drift of the handheld camera gazes reverently upward at a majestic, ancient tree. Its massive, gnarled trunk, rendered in washes of deep brown and mossy green, twists powerfully toward the sky, its vast canopy energized by a gentle mountain breeze. Broad, hand-shaped leaves rustle and dance in an enchanting rhythm, their forms defined by a vibrant tapestry of soft, emerald and ochre pigments delicately intermingling. Rhythmically swaying branches are hung with spiky, yellow-green shells, their sharp outlines softened by a painterly beauty. Golden afternoon light majestically filters through the leaves, casting speckled, shifting patterns of warm glow that make them appear translucent. Through gaps in the canopy, glimpses of the deep blue sky peek through, its color seamlessly flowing into the greens and yellows of the foliage, creating a moving, serene, painterly tableau, each element swaying in a calm, dreamlike visual poetry.", + "seed": 6929, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_13/index_1338_seed_6929_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.029089044965803623, + "vlm_logprob": -0.029089044965803623, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.029089044965803623, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a flowing watercolor animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle looking upward at the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree trunk massive, gnarled, and colored with deep brown and mossy green washes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are spiky yellow-green shells hanging from the swaying branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does golden afternoon light filter through the leaves creating speckled, shifting patterns of warm glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0035403212532401085 + ], + "opposite_logprob_tokens": [ + -5.878540515899658 + ], + "logprob_reward": -0.0035403212532401085, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0035403212532401085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2724998891353607, + 0.10032776743173599, + 0.09228359907865524, + 0.1299266666173935, + 0.05569193512201309, + 0.06609953939914703, + 0.052247900515794754, + 0.07260081171989441, + 0.05213875323534012, + 0.10618315637111664 + ], + "credit_map": [ + 0.024722542613744736, + 0.011716341599822044, + 0.025065988302230835, + 0.01270492933690548, + 0.029657263308763504, + 0.08742428570985794, + 0.015982555225491524, + 0.06522595882415771, + 0.011964050121605396, + 0.011490797623991966, + 0.019919442012906075, + 0.008441505953669548, + 0.013497333973646164, + 0.012772210873663425, + 0.010322474874556065, + 0.011919941753149033, + 0.010019412264227867, + 0.01057465560734272, + 0.014619630761444569, + 0.008183053694665432, + 0.008731914684176445, + 0.014712151139974594, + 0.009257909841835499, + 0.01618487760424614, + 0.016287600621581078, + 0.012361414730548859, + 0.014303064905107021, + 0.009083638899028301, + 0.010081695392727852, + 0.030837509781122208, + 0.009156223386526108, + 0.027815520763397217, + 0.004300808534026146, + 0.005073046777397394, + 0.004968344699591398, + 0.00368345994502306, + 0.005260155536234379, + 0.01426351722329855, + 0.004722143989056349, + 0.01342045795172453, + 0.00851438008248806, + 0.006718919146806002, + 0.009525769390165806, + 0.005951499100774527, + 0.006487696897238493, + 0.014329147525131702, + 0.006309046875685453, + 0.008263079449534416, + 0.005810047499835491, + 0.005326550453901291, + 0.006124393083155155, + 0.005547828506678343, + 0.004323381464928389, + 0.007389903999865055, + 0.009771800599992275, + 0.007953991182148457, + 0.0065840971656143665, + 0.01089427713304758, + 0.008139791898429394, + 0.005951186642050743, + 0.014084076508879662, + 0.01104767993092537, + 0.006429907400161028, + 0.009469792246818542, + 0.005621280521154404, + 0.0057185981422662735, + 0.008991165086627007, + 0.004844799637794495, + 0.006754034198820591, + 0.008256210945546627, + 0.0037501780316233635, + 0.008202488534152508, + 0.0144257927313447, + 0.010970711708068848, + 0.0105489706620574, + 0.008601674810051918, + 0.010898319073021412, + 0.025251809507608414, + 0.016022121533751488, + 0.009463753551244736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.0035403212532401085 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a flowing watercolor animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle looking upward at the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree trunk massive, gnarled, and colored with deep brown and mossy green washes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are spiky yellow-green shells hanging from the swaying branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does golden afternoon light filter through the leaves creating speckled, shifting patterns of warm glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003103941679001 + ], + "opposite_logprob_tokens": [ + -2.695030927658081 + ], + "logprob_reward": -0.07003103941679001, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003103941679001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.202447772026062, + 0.06866149604320526, + 0.0678168386220932, + 0.0512920506298542, + 0.1270681917667389, + 0.129009410738945, + 0.06689416617155075, + 0.1262989640235901, + 0.0846986249089241, + 0.07581251114606857 + ], + "credit_map": [ + 0.014888599514961243, + 0.011583102867007256, + 0.014993944205343723, + 0.008852381259202957, + 0.0283849835395813, + 0.06353844702243805, + 0.013009041547775269, + 0.047197263687849045, + 0.008193442597985268, + 0.009731804952025414, + 0.01118888147175312, + 0.006508275866508484, + 0.009154774248600006, + 0.00877491943538189, + 0.0055372342467308044, + 0.009572161361575127, + 0.005919127259403467, + 0.004748423118144274, + 0.005961468908935785, + 0.004218782298266888, + 0.005733435042202473, + 0.01867063157260418, + 0.004557515028864145, + 0.018007446080446243, + 0.006492992863059044, + 0.006490122061222792, + 0.007415236439555883, + 0.004540859255939722, + 0.00692722899839282, + 0.0066241305321455, + 0.005315836519002914, + 0.00748564163222909, + 0.007948195561766624, + 0.008234655484557152, + 0.006941209081560373, + 0.004950153175741434, + 0.007857698015868664, + 0.043660182505846024, + 0.007481844164431095, + 0.0399942509829998, + 0.024625366553664207, + 0.01428239420056343, + 0.015531235374510288, + 0.00894329696893692, + 0.010688571259379387, + 0.018817758187651634, + 0.01449423935264349, + 0.021626535803079605, + 0.008041766472160816, + 0.007533762603998184, + 0.007395211141556501, + 0.0053986492566764355, + 0.007495391648262739, + 0.009833838790655136, + 0.007743044290691614, + 0.013452496379613876, + 0.017031265422701836, + 0.014547201804816723, + 0.013790998607873917, + 0.008385938592255116, + 0.01969337835907936, + 0.02048609033226967, + 0.012174329720437527, + 0.02018975093960762, + 0.008760778233408928, + 0.008415392599999905, + 0.015880463644862175, + 0.008807222358882427, + 0.008455759845674038, + 0.015937093645334244, + 0.008342506363987923, + 0.010099398903548717, + 0.008747327141463757, + 0.008242172189056873, + 0.012660265900194645, + 0.006314230617135763, + 0.00823934469372034, + 0.0187512394040823, + 0.007139470893889666, + 0.005718459375202656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.07003103941679001 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a flowing watercolor animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle looking upward at the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree trunk massive, gnarled, and colored with deep brown and mossy green washes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are spiky yellow-green shells hanging from the swaying branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does golden afternoon light filter through the leaves creating speckled, shifting patterns of warm glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024894457310438156 + ], + "opposite_logprob_tokens": [ + -6.0024895668029785 + ], + "logprob_reward": -0.0024894457310438156, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024894457310438156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23712490499019623, + 0.09730076789855957, + 0.05609234422445297, + 0.0590696707367897, + 0.11353510618209839, + 0.08620457351207733, + 0.05819141864776611, + 0.12480136007070541, + 0.05872534587979317, + 0.10895439237356186 + ], + "credit_map": [ + 0.01907060295343399, + 0.008724751882255077, + 0.017719075083732605, + 0.010219764895737171, + 0.027481725439429283, + 0.07721709460020065, + 0.01339050941169262, + 0.06330142170190811, + 0.01263450738042593, + 0.009295138530433178, + 0.01723804883658886, + 0.007210463285446167, + 0.012961143627762794, + 0.014254012145102024, + 0.00934100802987814, + 0.014366460964083672, + 0.005272495094686747, + 0.0033464631997048855, + 0.005194747354835272, + 0.0031941812485456467, + 0.004460976924747229, + 0.014496798627078533, + 0.004028684925287962, + 0.016098009422421455, + 0.005384696181863546, + 0.007318702060729265, + 0.006680082995444536, + 0.004545310046523809, + 0.005879478063434362, + 0.011299961246550083, + 0.0062784780748188496, + 0.011682967655360699, + 0.009670022875070572, + 0.008943278342485428, + 0.012075116857886314, + 0.00787246786057949, + 0.0102676497772336, + 0.029069216921925545, + 0.010106679983437061, + 0.02553069032728672, + 0.01517399214208126, + 0.009472995065152645, + 0.011077294126152992, + 0.007911387830972672, + 0.008155623450875282, + 0.011507917195558548, + 0.009258970618247986, + 0.013646409846842289, + 0.007142942864447832, + 0.006464175879955292, + 0.005596289876848459, + 0.0041481005027890205, + 0.006202473305165768, + 0.011721598915755749, + 0.008300087414681911, + 0.008615761995315552, + 0.016442734748125076, + 0.012802557088434696, + 0.015569424256682396, + 0.008819852955639362, + 0.01390667911618948, + 0.019608119502663612, + 0.011603794060647488, + 0.026048224419355392, + 0.006253137718886137, + 0.006589023862034082, + 0.009619314223527908, + 0.0047166007570922375, + 0.009334944188594818, + 0.01060872245579958, + 0.004172875080257654, + 0.007430739235132933, + 0.021805115044116974, + 0.008031724952161312, + 0.00804874300956726, + 0.005248465575277805, + 0.010120503604412079, + 0.033032190054655075, + 0.014185612089931965, + 0.008482062257826328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.0024894457310438156 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a flowing watercolor animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle looking upward at the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree trunk massive, gnarled, and colored with deep brown and mossy green washes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are spiky yellow-green shells hanging from the swaying branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does golden afternoon light filter through the leaves creating speckled, shifting patterns of warm glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430484414100647 + ], + "opposite_logprob_tokens": [ + -3.16804838180542 + ], + "logprob_reward": -0.0430484414100647, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430484414100647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07222623378038406, + 0.02641422487795353, + 0.05670394003391266, + 0.043962232768535614, + 0.14301656186580658, + 0.12716561555862427, + 0.07799381017684937, + 0.2047930806875229, + 0.03261224180459976, + 0.2151120901107788 + ], + "credit_map": [ + 0.00581449456512928, + 0.0035803020000457764, + 0.006443304941058159, + 0.0029409800190478563, + 0.008873137645423412, + 0.022652558982372284, + 0.004247036296874285, + 0.017674414440989494, + 0.00309367710724473, + 0.003392961574718356, + 0.0035929251462221146, + 0.0019920547492802143, + 0.0028541297651827335, + 0.003814804833382368, + 0.0027738180942833424, + 0.004899852443486452, + 0.00588660454377532, + 0.0055610439740121365, + 0.005218998529016972, + 0.0037012104876339436, + 0.007185301743447781, + 0.011743685230612755, + 0.005533298943191767, + 0.011873794719576836, + 0.005329497158527374, + 0.006924531422555447, + 0.00634392537176609, + 0.003800589358434081, + 0.005547372158616781, + 0.005067599005997181, + 0.006154622882604599, + 0.004794097039848566, + 0.010093259625136852, + 0.00840074010193348, + 0.01045583002269268, + 0.005897688679397106, + 0.010225852020084858, + 0.04773973673582077, + 0.01083432137966156, + 0.0393691286444664, + 0.02196037583053112, + 0.012581266462802887, + 0.012740821577608585, + 0.010127628222107887, + 0.013663257472217083, + 0.020708978176116943, + 0.016663724556565285, + 0.01871955394744873, + 0.008070065639913082, + 0.007665735669434071, + 0.008366509340703487, + 0.005410825368016958, + 0.008402685634791851, + 0.013788902200758457, + 0.008154808543622494, + 0.01813427545130253, + 0.023376988247036934, + 0.019679713994264603, + 0.018864789977669716, + 0.013392258435487747, + 0.024421846494078636, + 0.03931352496147156, + 0.028034279122948647, + 0.03770967572927475, + 0.003918668255209923, + 0.003220414975658059, + 0.005405012983828783, + 0.002692658454179764, + 0.004273659083992243, + 0.005779422353953123, + 0.0036883580032736063, + 0.0036340474616736174, + 0.038939669728279114, + 0.01924993470311165, + 0.019637204706668854, + 0.011343042366206646, + 0.02346321940422058, + 0.075648233294487, + 0.01166072953492403, + 0.015170054510235786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 4 + ], + "selected_reward": -0.0430484414100647 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a flowing watercolor animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle looking upward at the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree trunk massive, gnarled, and colored with deep brown and mossy green washes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are spiky yellow-green shells hanging from the swaying branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does golden afternoon light filter through the leaves creating speckled, shifting patterns of warm glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026335977017879486 + ], + "opposite_logprob_tokens": [ + -3.6513359546661377 + ], + "logprob_reward": -0.026335977017879486, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026335977017879486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26023998856544495, + 0.059708938002586365, + 0.06436681747436523, + 0.08169414848089218, + 0.06590747833251953, + 0.043692346662282944, + 0.05947541445493698, + 0.07049147039651871, + 0.06246841698884964, + 0.23195508122444153 + ], + "credit_map": [ + 0.02129802666604519, + 0.009277017787098885, + 0.019577208906412125, + 0.012331632897257805, + 0.0260182935744524, + 0.08640570193529129, + 0.014571481384336948, + 0.0707605853676796, + 0.007432158570736647, + 0.006254665087908506, + 0.005052455700933933, + 0.0055152652785182, + 0.007281667552888393, + 0.009975825436413288, + 0.007428871002048254, + 0.010768021456897259, + 0.007988848723471165, + 0.007145571988075972, + 0.008044942282140255, + 0.007476180791854858, + 0.00862622819840908, + 0.008458410389721394, + 0.008498767390847206, + 0.008127858862280846, + 0.007082975469529629, + 0.007781381253153086, + 0.010684906505048275, + 0.005678623449057341, + 0.008733777329325676, + 0.018647771328687668, + 0.0070814769715070724, + 0.016003232449293137, + 0.008104592561721802, + 0.005899517796933651, + 0.008060754276812077, + 0.004450110252946615, + 0.005467629060149193, + 0.016308683902025223, + 0.004926273599267006, + 0.012689914554357529, + 0.004927079658955336, + 0.004977541975677013, + 0.005497045814990997, + 0.0035970036406069994, + 0.005553657654672861, + 0.007114794105291367, + 0.004255938343703747, + 0.007769281510263681, + 0.008469512686133385, + 0.006937774363905191, + 0.008345097303390503, + 0.005348049569875002, + 0.006308204028755426, + 0.009378546848893166, + 0.005609290674328804, + 0.009078931994736195, + 0.011017571203410625, + 0.007072194013744593, + 0.008401613682508469, + 0.005817941855639219, + 0.00900219939649105, + 0.012603084556758404, + 0.0061194575391709805, + 0.010457400232553482, + 0.007684035692363977, + 0.008623550646007061, + 0.010487237945199013, + 0.004699677228927612, + 0.009344291873276234, + 0.008089946582913399, + 0.007095452398061752, + 0.006444219499826431, + 0.04638572037220001, + 0.014669446274638176, + 0.016146576032042503, + 0.0088827358558774, + 0.023603133857250214, + 0.0831756517291069, + 0.021148988977074623, + 0.017942799255251884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.026335977017879486 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in a flowing watercolor animation style, a low-angle shot with a subtle, organic drift of the handheld camera gazes reverently upward at a majestic, ancient tree. Its massive, gnarled trunk, rendered in washes of deep brown and mossy green, twists powerfully toward the sky, its vast canopy energized by a gentle mountain breeze. Broad, hand-shaped leaves rustle and dance in an enchanting rhythm, their forms defined by a vibrant tapestry of soft, emerald and ochre pigments delicately intermingling. Rhythmically swaying branches are hung with spiky, yellow-green shells, their sharp outlines softened by a painterly beauty. Golden afternoon light majestically filters through the leaves, casting speckled, shifting patterns of warm glow that make them appear translucent. Through gaps in the canopy, glimpses of the deep blue sky peek through, its color seamlessly flowing into the greens and yellows of the foliage, creating a moving, serene, painterly tableau, each element swaying in a calm, dreamlike visual poetry.", + "seed": 6930, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_13/index_1338_seed_6930_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.040120324259623884, + "vlm_logprob": -0.040120324259623884, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.040120324259623884, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a flowing watercolor animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle looking upward at the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree trunk massive, gnarled, and colored with deep brown and mossy green washes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are spiky yellow-green shells hanging from the swaying branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does golden afternoon light filter through the leaves creating speckled, shifting patterns of warm glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004660695791244507 + ], + "opposite_logprob_tokens": [ + -5.629660606384277 + ], + "logprob_reward": -0.004660695791244507, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.004660695791244507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08425474166870117, + 0.03647841885685921, + 0.03594527766108513, + 0.041643477976322174, + 0.025709185749292374, + 0.21589265763759613, + 0.42552104592323303, + 0.06641416996717453, + 0.024969249963760376, + 0.0431717149913311 + ], + "credit_map": [ + 0.009030221961438656, + 0.0129891661927104, + 0.01053079217672348, + 0.00897742435336113, + 0.01106405071914196, + 0.00904258992522955, + 0.013488676398992538, + 0.009131827391684055, + 0.0054887342266738415, + 0.005657930858433247, + 0.007420824375003576, + 0.004439233802258968, + 0.003664797404780984, + 0.0023644757457077503, + 0.004199263174086809, + 0.0032431602012366056, + 0.005827596411108971, + 0.005166224204003811, + 0.006237856578081846, + 0.003516213968396187, + 0.005674184300005436, + 0.0026996158994734287, + 0.002828205469995737, + 0.0039953868836164474, + 0.00737036345526576, + 0.004407825414091349, + 0.00611704308539629, + 0.00801890715956688, + 0.0056356699205935, + 0.002918539335951209, + 0.003631661180406809, + 0.0035434740129858255, + 0.003310549771413207, + 0.002151001710444689, + 0.0030015241354703903, + 0.005062953568994999, + 0.0033832143526524305, + 0.00210002763196826, + 0.0034695949871093035, + 0.0032303209882229567, + 0.0184247475117445, + 0.015290196985006332, + 0.09030579775571823, + 0.01931227371096611, + 0.018450003117322922, + 0.012367046438157558, + 0.02142486907541752, + 0.020317744463682175, + 0.03060295805335045, + 0.031236128881573677, + 0.024731222540140152, + 0.13541479408740997, + 0.1326378881931305, + 0.022188574075698853, + 0.02959573268890381, + 0.019113769754767418, + 0.007615650072693825, + 0.008926487527787685, + 0.01023096963763237, + 0.00914536602795124, + 0.009782078675925732, + 0.0065034725703299046, + 0.008366478607058525, + 0.005843674298375845, + 0.0033223580103367567, + 0.0033042049035429955, + 0.0035846384707838297, + 0.0036990956868976355, + 0.003771435935050249, + 0.002632939489558339, + 0.0025780219584703445, + 0.002076559467241168, + 0.006976461037993431, + 0.006360240280628204, + 0.0038696981500834227, + 0.005335030145943165, + 0.0041047255508601665, + 0.005017605610191822, + 0.004279992077499628, + 0.00722796656191349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.004660695791244507 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a flowing watercolor animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle looking upward at the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree trunk massive, gnarled, and colored with deep brown and mossy green washes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are spiky yellow-green shells hanging from the swaying branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does golden afternoon light filter through the leaves creating speckled, shifting patterns of warm glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16039907932281494 + ], + "opposite_logprob_tokens": [ + -1.910399079322815 + ], + "logprob_reward": -0.16039907932281494, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16039907932281494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26176923513412476, + 0.06935836374759674, + 0.053097181022167206, + 0.06770408153533936, + 0.06088699400424957, + 0.0764821395277977, + 0.022770294919610023, + 0.1916256546974182, + 0.14542168378829956, + 0.05088437348604202 + ], + "credit_map": [ + 0.026248767971992493, + 0.03985312953591347, + 0.038654327392578125, + 0.03529592975974083, + 0.030481966212391853, + 0.022836847230792046, + 0.04148067533969879, + 0.02691761963069439, + 0.008965705521404743, + 0.008737154304981232, + 0.009014072827994823, + 0.012044033966958523, + 0.006331687327474356, + 0.0045527126640081406, + 0.005748112685978413, + 0.01396489329636097, + 0.008075310848653316, + 0.005648862570524216, + 0.010767990723252296, + 0.0050877914763987064, + 0.007637071423232555, + 0.004696415271610022, + 0.004211577586829662, + 0.006972166243940592, + 0.011579377576708794, + 0.007497201208025217, + 0.010584385134279728, + 0.008637149818241596, + 0.005800831597298384, + 0.007606173399835825, + 0.008350168354809284, + 0.007648801896721125, + 0.008168524131178856, + 0.009281275793910027, + 0.0058289747685194016, + 0.010518302209675312, + 0.006104986183345318, + 0.004732778295874596, + 0.006811995059251785, + 0.009440160356462002, + 0.009375860914587975, + 0.0074060228653252125, + 0.010314073413610458, + 0.01654752716422081, + 0.010088852606713772, + 0.008295138366520405, + 0.008520697243511677, + 0.005933978594839573, + 0.003405071562156081, + 0.0027752849273383617, + 0.002831406192854047, + 0.00335674942471087, + 0.004213358275592327, + 0.0022849002853035927, + 0.002086372347548604, + 0.001817156095057726, + 0.043869901448488235, + 0.025138940662145615, + 0.01558597106486559, + 0.03178882226347923, + 0.026071224361658096, + 0.01459911186248064, + 0.019762525334954262, + 0.014809188432991505, + 0.024196147918701172, + 0.017406897619366646, + 0.016716934740543365, + 0.01765548065304756, + 0.019485417753458023, + 0.020100118592381477, + 0.016634689643979073, + 0.013226000592112541, + 0.004867044743150473, + 0.008445639163255692, + 0.0048967394977808, + 0.006840325426310301, + 0.006308478303253651, + 0.007391414139419794, + 0.007021530531346798, + 0.005113204475492239 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.16039907932281494 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a flowing watercolor animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle looking upward at the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree trunk massive, gnarled, and colored with deep brown and mossy green washes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are spiky yellow-green shells hanging from the swaying branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does golden afternoon light filter through the leaves creating speckled, shifting patterns of warm glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024895647075027227 + ], + "opposite_logprob_tokens": [ + -6.0024895668029785 + ], + "logprob_reward": -0.0024895647075027227, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024895647075027227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15307125449180603, + 0.17800934612751007, + 0.04057179018855095, + 0.02574021741747856, + 0.07250002771615982, + 0.11051035672426224, + 0.24347424507141113, + 0.09303167462348938, + 0.05586615577340126, + 0.02722487971186638 + ], + "credit_map": [ + 0.019239045679569244, + 0.025678317993879318, + 0.015310186892747879, + 0.019445914775133133, + 0.01952921226620674, + 0.017220579087734222, + 0.022342294454574585, + 0.014305726625025272, + 0.01795913837850094, + 0.03105371817946434, + 0.036055389791727066, + 0.024199288338422775, + 0.017262082546949387, + 0.0115400031208992, + 0.01975937746465206, + 0.020180361345410347, + 0.0043310848996043205, + 0.004834008403122425, + 0.009564301930367947, + 0.0036661471240222454, + 0.004067323170602322, + 0.0030339837539941072, + 0.003255022456869483, + 0.007819924503564835, + 0.004382513929158449, + 0.0027991740498691797, + 0.003651530249044299, + 0.0034045905340462923, + 0.0030990662053227425, + 0.0027847266755998135, + 0.0027738551143556833, + 0.0028447641525417566, + 0.010659154504537582, + 0.00866704247891903, + 0.008236585184931755, + 0.011412412859499454, + 0.009793956764042377, + 0.004312800709158182, + 0.00958588719367981, + 0.009832199662923813, + 0.010985822416841984, + 0.008969927206635475, + 0.03927670419216156, + 0.011977835558354855, + 0.012724630534648895, + 0.007368709892034531, + 0.0119469054043293, + 0.007259829435497522, + 0.01343655027449131, + 0.01644815504550934, + 0.009787656366825104, + 0.07636801898479462, + 0.0924587994813919, + 0.010421846061944962, + 0.015527987852692604, + 0.00902525894343853, + 0.013944610953330994, + 0.010875319130718708, + 0.010464857332408428, + 0.014127333648502827, + 0.01507115364074707, + 0.00872718170285225, + 0.011799728497862816, + 0.008021501824259758, + 0.008380517363548279, + 0.0070148915983736515, + 0.007389182690531015, + 0.007802294101566076, + 0.00769186532124877, + 0.006597806233912706, + 0.006126704625785351, + 0.004862900823354721, + 0.003004732308909297, + 0.0041225082240998745, + 0.0031682001426815987, + 0.0038031120784580708, + 0.0026493447367101908, + 0.003017804119735956, + 0.004665780812501907, + 0.0027933993842452765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.0024895647075027227 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a flowing watercolor animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle looking upward at the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree trunk massive, gnarled, and colored with deep brown and mossy green washes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are spiky yellow-green shells hanging from the swaying branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does golden afternoon light filter through the leaves creating speckled, shifting patterns of warm glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327677607536316 + ], + "opposite_logprob_tokens": [ + -3.7732768058776855 + ], + "logprob_reward": -0.02327677607536316, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327677607536316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.132553830742836, + 0.13643944263458252, + 0.06807944923639297, + 0.09337136149406433, + 0.029545653611421585, + 0.07435289025306702, + 0.2145594209432602, + 0.12229691445827484, + 0.05010597035288811, + 0.07869502156972885 + ], + "credit_map": [ + 0.020965712144970894, + 0.021240467205643654, + 0.013737675733864307, + 0.016559263691306114, + 0.015724025666713715, + 0.01302587054669857, + 0.02010638825595379, + 0.011194436810910702, + 0.015525813214480877, + 0.013491129502654076, + 0.014032372273504734, + 0.03361568972468376, + 0.009231585077941418, + 0.007171586621552706, + 0.011927175335586071, + 0.03144410625100136, + 0.010225201956927776, + 0.0073477728292346, + 0.007852323353290558, + 0.006592888850718737, + 0.007824476808309555, + 0.010782989673316479, + 0.008422239683568478, + 0.009031565859913826, + 0.0097260857000947, + 0.012586846947669983, + 0.010895653627812862, + 0.011066542007029057, + 0.018617795780301094, + 0.01185884140431881, + 0.010219570249319077, + 0.008400039747357368, + 0.004607037175446749, + 0.0028429480735212564, + 0.0042494237422943115, + 0.0037979206535965204, + 0.003724301466718316, + 0.002835616236552596, + 0.004689525347203016, + 0.0027988869696855545, + 0.007478997576981783, + 0.006734869442880154, + 0.0183823611587286, + 0.01242552138864994, + 0.010154727846384048, + 0.006734167691320181, + 0.008027518168091774, + 0.0044147358275949955, + 0.012558262795209885, + 0.014943835325539112, + 0.009325530380010605, + 0.06153900921344757, + 0.08694194257259369, + 0.009295151568949223, + 0.013811574317514896, + 0.006144146900624037, + 0.02788415551185608, + 0.01782282069325447, + 0.015444044023752213, + 0.011960366740822792, + 0.016507519409060478, + 0.012304081581532955, + 0.012408659793436527, + 0.007965276017785072, + 0.009621625766158104, + 0.0073524038307368755, + 0.00605163024738431, + 0.005968478042632341, + 0.0064787971787154675, + 0.005433099344372749, + 0.004957389086484909, + 0.004242554306983948, + 0.00979252252727747, + 0.009112556464970112, + 0.00901241134852171, + 0.00997057743370533, + 0.008145825006067753, + 0.012118567712605, + 0.009375019930303097, + 0.011167549528181553 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.02327677607536316 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a flowing watercolor animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle looking upward at the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree trunk massive, gnarled, and colored with deep brown and mossy green washes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are spiky yellow-green shells hanging from the swaying branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does golden afternoon light filter through the leaves creating speckled, shifting patterns of warm glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009775505401194096 + ], + "opposite_logprob_tokens": [ + -4.634775638580322 + ], + "logprob_reward": -0.009775505401194096, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009775505401194096, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11072533577680588, + 0.15478035807609558, + 0.08820941299200058, + 0.0241233017295599, + 0.08447448909282684, + 0.1369001865386963, + 0.24361233413219452, + 0.06767192482948303, + 0.05246549844741821, + 0.03703717514872551 + ], + "credit_map": [ + 0.012084247544407845, + 0.018108325079083443, + 0.011728879064321518, + 0.015274371020495892, + 0.014689686708152294, + 0.012044984847307205, + 0.016270894557237625, + 0.010523941367864609, + 0.015790268778800964, + 0.017950139939785004, + 0.03544085472822189, + 0.029431520029902458, + 0.011675700545310974, + 0.008681295439600945, + 0.01316078007221222, + 0.022649794816970825, + 0.012973652221262455, + 0.012808920815587044, + 0.015928473323583603, + 0.009347054176032543, + 0.00969183724373579, + 0.009041148237884045, + 0.008957087062299252, + 0.009461228735744953, + 0.0037870421074330807, + 0.0026500369422137737, + 0.004440493881702423, + 0.00334970373660326, + 0.001804068684577942, + 0.0028699066024273634, + 0.0028967023827135563, + 0.002325347624719143, + 0.011325610801577568, + 0.012037538923323154, + 0.00976554024964571, + 0.011037635616958141, + 0.009097767062485218, + 0.005668458063155413, + 0.009424878284335136, + 0.01611705869436264, + 0.011773851700127125, + 0.010979133658111095, + 0.03739483281970024, + 0.012485473416745663, + 0.016018586233258247, + 0.010622882284224033, + 0.017096947878599167, + 0.020528482273221016, + 0.013515596278011799, + 0.016904212534427643, + 0.010136730037629604, + 0.07487965375185013, + 0.0913022831082344, + 0.011339724995195866, + 0.015774188563227654, + 0.009759952314198017, + 0.01052737608551979, + 0.009165974333882332, + 0.007071679923683405, + 0.007727614603936672, + 0.009885153733193874, + 0.0076429592445492744, + 0.008576015010476112, + 0.007075154222548008, + 0.007795278914272785, + 0.0074967509135603905, + 0.006429631263017654, + 0.007104275748133659, + 0.008116224780678749, + 0.004335444420576096, + 0.006417052820324898, + 0.004770836792886257, + 0.0036044565495103598, + 0.00768553651869297, + 0.005399388261139393, + 0.00450120260939002, + 0.003411160781979561, + 0.004371951799839735, + 0.0051499661058187485, + 0.002913513919338584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 5 + ], + "selected_reward": -0.009775505401194096 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in a flowing watercolor animation style, a low-angle shot with a subtle, organic drift of the handheld camera gazes reverently upward at a majestic, ancient tree. Its massive, gnarled trunk, rendered in washes of deep brown and mossy green, twists powerfully toward the sky, its vast canopy energized by a gentle mountain breeze. Broad, hand-shaped leaves rustle and dance in an enchanting rhythm, their forms defined by a vibrant tapestry of soft, emerald and ochre pigments delicately intermingling. Rhythmically swaying branches are hung with spiky, yellow-green shells, their sharp outlines softened by a painterly beauty. Golden afternoon light majestically filters through the leaves, casting speckled, shifting patterns of warm glow that make them appear translucent. Through gaps in the canopy, glimpses of the deep blue sky peek through, its color seamlessly flowing into the greens and yellows of the foliage, creating a moving, serene, painterly tableau, each element swaying in a calm, dreamlike visual poetry.", + "seed": 6931, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_13/index_1338_seed_6931_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.011202219617553055, + "vlm_logprob": -0.011202219617553055, + "vlm_yes_no_margin": 5.125, + "vlm_token_credit": -0.011202219617553055, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a flowing watercolor animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle looking upward at the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree trunk massive, gnarled, and colored with deep brown and mossy green washes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are spiky yellow-green shells hanging from the swaying branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does golden afternoon light filter through the leaves creating speckled, shifting patterns of warm glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022574197500944138 + ], + "opposite_logprob_tokens": [ + -6.627257347106934 + ], + "logprob_reward": -0.0022574197500944138, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0022574197500944138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03152179345488548, + 0.6526802182197571, + 0.014586292207241058, + 0.16435670852661133, + 0.030317410826683044, + 0.02048509381711483, + 0.01621422730386257, + 0.017934782430529594, + 0.03267106041312218, + 0.01923242211341858 + ], + "credit_map": [ + 0.006127622909843922, + 0.003025870071724057, + 0.003632392967119813, + 0.0023820982314646244, + 0.0038495089393109083, + 0.005363773554563522, + 0.003027700586244464, + 0.004112825728952885, + 0.2416369467973709, + 0.03138116002082825, + 0.07372726500034332, + 0.02104877308011055, + 0.028558244928717613, + 0.12802520394325256, + 0.03809673339128494, + 0.09020590782165527, + 0.002052766503766179, + 0.001399292959831655, + 0.00164429668802768, + 0.0016015510773286223, + 0.0022045630030333996, + 0.0025084891822189093, + 0.001774105941876769, + 0.0014012266183272004, + 0.02146950736641884, + 0.024203674867749214, + 0.014694432727992535, + 0.018289124593138695, + 0.019895024597644806, + 0.016906507313251495, + 0.024137206375598907, + 0.02476123906672001, + 0.004343670792877674, + 0.0029011822771281004, + 0.00688906479626894, + 0.004600448068231344, + 0.003167049027979374, + 0.0024944632314145565, + 0.003175342222675681, + 0.0027461915742605925, + 0.002224231604486704, + 0.0025392738170921803, + 0.0028996483888477087, + 0.004367312416434288, + 0.0036715385504066944, + 0.0013480440247803926, + 0.0021117946598678827, + 0.001323249889537692, + 0.002418181858956814, + 0.0015399634139612317, + 0.0015884324675425887, + 0.002047757850959897, + 0.0022311280481517315, + 0.0030371765606105328, + 0.001948982710018754, + 0.0014026050921529531, + 0.0014799052150920033, + 0.001898854156024754, + 0.0013097284827381372, + 0.0017721803160384297, + 0.006321448367089033, + 0.0015260209329426289, + 0.0021212678402662277, + 0.0015053777024149895, + 0.004339323844760656, + 0.0038264133036136627, + 0.004412979818880558, + 0.004823253024369478, + 0.005041339434683323, + 0.0035907342098653316, + 0.0035090025048702955, + 0.0031280156690627337, + 0.0030801102984696627, + 0.002485818462446332, + 0.002136786002665758, + 0.003023295197635889, + 0.002139126881957054, + 0.0011850367300212383, + 0.0022044565994292498, + 0.002977792639285326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 8 + ], + "selected_reward": -0.0022574197500944138 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a flowing watercolor animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle looking upward at the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree trunk massive, gnarled, and colored with deep brown and mossy green washes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are spiky yellow-green shells hanging from the swaying branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does golden afternoon light filter through the leaves creating speckled, shifting patterns of warm glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029806073755025864 + ], + "opposite_logprob_tokens": [ + -3.529806137084961 + ], + "logprob_reward": -0.029806073755025864, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029806073755025864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09851524978876114, + 0.43588948249816895, + 0.026180293411016464, + 0.044465500861406326, + 0.03590802475810051, + 0.02366037666797638, + 0.1583467721939087, + 0.023492658510804176, + 0.0698879063129425, + 0.08365367352962494 + ], + "credit_map": [ + 0.02269885502755642, + 0.007583566941320896, + 0.014401703141629696, + 0.007328395266085863, + 0.01466588769108057, + 0.01052864734083414, + 0.010265040211379528, + 0.011043155565857887, + 0.16271759569644928, + 0.02160039357841015, + 0.05016603320837021, + 0.01377207599580288, + 0.01771957240998745, + 0.0866628885269165, + 0.02579190582036972, + 0.05745905265212059, + 0.0032581340055912733, + 0.0041335579007864, + 0.0028465609066188335, + 0.00266016717068851, + 0.004431391134858131, + 0.003468187293037772, + 0.0030823154374957085, + 0.002299980726093054, + 0.0076452502980828285, + 0.004653295502066612, + 0.007941099815070629, + 0.004135517869144678, + 0.005575675051659346, + 0.004247728735208511, + 0.004519285634160042, + 0.005747647024691105, + 0.005997190717607737, + 0.003677153727039695, + 0.004176137503236532, + 0.00396703602746129, + 0.0031146323308348656, + 0.0097477613016963, + 0.0031853155232965946, + 0.0020427971612662077, + 0.003079337300732732, + 0.0025947510730475187, + 0.002338131656870246, + 0.003216525772586465, + 0.0027557690627872944, + 0.005182419903576374, + 0.002247435972094536, + 0.0022460059262812138, + 0.030755622312426567, + 0.018680265173316002, + 0.01833059825003147, + 0.024577168747782707, + 0.01804955117404461, + 0.023933904245495796, + 0.01236681267619133, + 0.01165285985916853, + 0.0023967381566762924, + 0.002240953501313925, + 0.0023044520057737827, + 0.0024207395035773516, + 0.007473284844309092, + 0.002271124627441168, + 0.0022474571596831083, + 0.0021379098761826754, + 0.010687106288969517, + 0.006445417180657387, + 0.010337070561945438, + 0.011248256079852581, + 0.010408355854451656, + 0.008165000937879086, + 0.006938002537935972, + 0.005658692214637995, + 0.008722330443561077, + 0.007838567718863487, + 0.013180851005017757, + 0.01441918034106493, + 0.011417049914598465, + 0.008753358386456966, + 0.012088106945157051, + 0.007234233431518078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.029806073755025864 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a flowing watercolor animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle looking upward at the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree trunk massive, gnarled, and colored with deep brown and mossy green washes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are spiky yellow-green shells hanging from the swaying branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does golden afternoon light filter through the leaves creating speckled, shifting patterns of warm glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019426537910476327 + ], + "opposite_logprob_tokens": [ + -6.2519426345825195 + ], + "logprob_reward": -0.0019426537910476327, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019426537910476327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19876158237457275, + 0.48216667771339417, + 0.03014395572245121, + 0.02384399250149727, + 0.04035450518131256, + 0.017499692738056183, + 0.02517588436603546, + 0.04057616740465164, + 0.06666162610054016, + 0.07481586188077927 + ], + "credit_map": [ + 0.027072489261627197, + 0.018204661086201668, + 0.022604433819651604, + 0.01714196428656578, + 0.03168058022856712, + 0.027800967916846275, + 0.036219522356987, + 0.018036987632513046, + 0.17950062453746796, + 0.02310108207166195, + 0.05528417229652405, + 0.015348219312727451, + 0.020277515053749084, + 0.09638209640979767, + 0.02780945785343647, + 0.06446351855993271, + 0.003317685332149267, + 0.003894917666912079, + 0.0035238806158304214, + 0.0036382670514285564, + 0.0043602604418993, + 0.003634778782725334, + 0.00369494897313416, + 0.004079221747815609, + 0.0032262629829347134, + 0.004012837540358305, + 0.0036048507317900658, + 0.0030915753450244665, + 0.002899952931329608, + 0.0020128898322582245, + 0.0029977732338011265, + 0.0019978524651378393, + 0.006295957136899233, + 0.003743804292753339, + 0.0066079031676054, + 0.003977418411523104, + 0.003913805820047855, + 0.008698660880327225, + 0.004034749232232571, + 0.003082210198044777, + 0.0018618148751556873, + 0.0026072589680552483, + 0.001707659917883575, + 0.003571031615138054, + 0.002827528864145279, + 0.0013006188673898578, + 0.0019739519339054823, + 0.001649830024689436, + 0.003228860441595316, + 0.0028053107671439648, + 0.002204503398388624, + 0.003464905545115471, + 0.005507500842213631, + 0.003668349701911211, + 0.0025565873365849257, + 0.0017398686613887548, + 0.004874145612120628, + 0.00525237899273634, + 0.00497411610558629, + 0.00798097439110279, + 0.00651986850425601, + 0.003337134374305606, + 0.004533577710390091, + 0.0031039791647344828, + 0.008213463239371777, + 0.006839248817414045, + 0.009015185758471489, + 0.012312527745962143, + 0.010255028493702412, + 0.0070187668316066265, + 0.008669077418744564, + 0.0043383329175412655, + 0.008636405691504478, + 0.010021123103797436, + 0.0068888673558831215, + 0.01593838259577751, + 0.009961308911442757, + 0.007849039509892464, + 0.009836492128670216, + 0.005684252362698317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.0019426537910476327 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a flowing watercolor animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle looking upward at the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree trunk massive, gnarled, and colored with deep brown and mossy green washes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are spiky yellow-green shells hanging from the swaying branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does golden afternoon light filter through the leaves creating speckled, shifting patterns of warm glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016059335321187973 + ], + "opposite_logprob_tokens": [ + -4.141059398651123 + ], + "logprob_reward": -0.016059335321187973, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016059335321187973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04885275661945343, + 0.8227586150169373, + 0.027053209021687508, + 0.016860704869031906, + 0.020500708371400833, + 0.006821773014962673, + 0.011518548242747784, + 0.008003412745893002, + 0.014633706770837307, + 0.02299659512937069 + ], + "credit_map": [ + 0.0064159659668803215, + 0.004798186477273703, + 0.005390448961406946, + 0.006109834183007479, + 0.007390947081148624, + 0.00537863839417696, + 0.008858873508870602, + 0.004509856924414635, + 0.28323599696159363, + 0.04836432635784149, + 0.08716951310634613, + 0.030496865510940552, + 0.03325706720352173, + 0.08885039389133453, + 0.053106989711523056, + 0.19827745854854584, + 0.004623001907020807, + 0.004425575491040945, + 0.0032028888817876577, + 0.002136698691174388, + 0.002821007277816534, + 0.003958511166274548, + 0.002239430556073785, + 0.003646095050498843, + 0.002488710219040513, + 0.002087730448693037, + 0.0013618855737149715, + 0.002032878343015909, + 0.0017947704764083028, + 0.0016065090894699097, + 0.002808982040733099, + 0.0026792390272021294, + 0.0037091870326548815, + 0.0017128211911767721, + 0.0022261496633291245, + 0.0020422651432454586, + 0.0014865892007946968, + 0.005752162076532841, + 0.0017851744778454304, + 0.0017863594694063067, + 0.0011316332966089249, + 0.0006382038118317723, + 0.0006626634276472032, + 0.0008246447541750968, + 0.0006650708382949233, + 0.0012365123257040977, + 0.0009170880075544119, + 0.0007459562621079385, + 0.0015228328993543983, + 0.0011978748952969909, + 0.001450299401767552, + 0.0016076563624665141, + 0.00213295454159379, + 0.0013663156423717737, + 0.0011292442213743925, + 0.001111370394937694, + 0.0007331597735174, + 0.0007328083156608045, + 0.0007911375723779202, + 0.0008558391127735376, + 0.0020612638909369707, + 0.0007207448943518102, + 0.0008260234608314931, + 0.0012824359582737088, + 0.002092768205329776, + 0.0019948547706007957, + 0.0018031933577731252, + 0.0018706695409491658, + 0.002086313208565116, + 0.001171064330264926, + 0.0016116617480292916, + 0.0020031810272485018, + 0.0029918847139924765, + 0.0022097723558545113, + 0.004188149701803923, + 0.005223978776484728, + 0.002462034346535802, + 0.0018278388306498528, + 0.002368100918829441, + 0.0017248339718207717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.016059335321187973 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a flowing watercolor animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle looking upward at the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree trunk massive, gnarled, and colored with deep brown and mossy green washes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are spiky yellow-green shells hanging from the swaying branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does golden afternoon light filter through the leaves creating speckled, shifting patterns of warm glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005945615470409393 + ], + "opposite_logprob_tokens": [ + -5.130945682525635 + ], + "logprob_reward": -0.005945615470409393, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005945615470409393, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05547920987010002, + 0.6279283761978149, + 0.027627507224678993, + 0.06328673660755157, + 0.054598912596702576, + 0.026678966358304024, + 0.02955952100455761, + 0.04273517429828644, + 0.014680638909339905, + 0.057424988597631454 + ], + "credit_map": [ + 0.011381532996892929, + 0.004278790671378374, + 0.007220962084829807, + 0.005811937619000673, + 0.006323456764221191, + 0.006695183459669352, + 0.007755854167044163, + 0.006011481396853924, + 0.23313267529010773, + 0.0318458192050457, + 0.07112869620323181, + 0.019187746569514275, + 0.027473023161292076, + 0.14090557396411896, + 0.035843390971422195, + 0.06841140240430832, + 0.0045585911720991135, + 0.004179995507001877, + 0.0026804818771779537, + 0.002887962618842721, + 0.0033120980951935053, + 0.0043280525133013725, + 0.002504761330783367, + 0.003175562946125865, + 0.007837051525712013, + 0.009664420038461685, + 0.006329046096652746, + 0.00849633477628231, + 0.009679392911493778, + 0.005723833572119474, + 0.008257227018475533, + 0.007299421820789576, + 0.006460643373429775, + 0.004349095281213522, + 0.006533902138471603, + 0.005874364636838436, + 0.006720507983118296, + 0.007547491230070591, + 0.007710959296673536, + 0.009401940740644932, + 0.003145757829770446, + 0.0019565047696232796, + 0.00565791130065918, + 0.002552096964791417, + 0.0038495606277137995, + 0.0026631527580320835, + 0.0032657727133482695, + 0.0035882075317204, + 0.004427643492817879, + 0.0029792464338243008, + 0.002974608913064003, + 0.0033760294318199158, + 0.004495091270655394, + 0.003944294061511755, + 0.004551813937723637, + 0.0028107885736972094, + 0.00550260953605175, + 0.0064324685372412205, + 0.0038308233488351107, + 0.004717923700809479, + 0.004507738631218672, + 0.005561418365687132, + 0.006274515762925148, + 0.005907671991735697, + 0.0023853268940001726, + 0.001249460387043655, + 0.002326142741367221, + 0.0019007482333108783, + 0.0017748225945979357, + 0.001351817511022091, + 0.0019038668833673, + 0.0017884518019855022, + 0.007083189208060503, + 0.006023381371051073, + 0.008466294966638088, + 0.008874118328094482, + 0.007258490659296513, + 0.005087126512080431, + 0.008918918669223785, + 0.0057134623639285564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 9 + ], + "selected_reward": -0.005945615470409393 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_14/index_tensor([1063, 1063])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_14/index_tensor([1063, 1063])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..091ecfb34bd0d18d8b1ea36f3b1c3faae82d8772 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_14/index_tensor([1063, 1063])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic tracking shot, filmed from a low angle, keeps close to the ground, alongside a nimble golden retriever as it gallops across a vast expanse of rolling sand dunes with explosive energy. Its magnificent golden fur sways in the warm golden afternoon sun, its ears fluttering joyfully as it runs, its tongue lolling out of its mouth, its eyes filled with focus and pure excitement. Its target is a bright red tennis ball bouncing erratically along the dune slope ahead. With each powerful kick, its paws dig deep into the soft, fine, pale sand, stirring up a shower of tiny, sunlit grains. The entire scene is bathed in the warm hues of the setting sun, casting long, dramatic shadows across the rolling dunes. In the distance, the horizon is where the deep blue sea meets the clear sky, creating a dynamic atmosphere of boundless freedom and pure joy.", + "seed": 27568, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_14/index_1063_seed_27568_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.018730613682419062, + "vlm_logprob": -0.018730613682419062, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.018730613682419062, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog galloping alongside a bright red tennis ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast expanse of rolling sand dunes with pale, fine sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a deep blue sea meeting a clear sky at the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005324350204318762 + ], + "opposite_logprob_tokens": [ + -6.380324363708496 + ], + "logprob_reward": -0.005324350204318762, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.005324350204318762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16260932385921478, + 0.06702511757612228, + 0.08130605518817902, + 0.07352063804864883, + 0.042529232800006866, + 0.11572762578725815, + 0.11985848098993301, + 0.08076736330986023, + 0.10616680234670639, + 0.1504894345998764 + ], + "credit_map": [ + 0.015488086268305779, + 0.02523798868060112, + 0.014243891462683678, + 0.01661127433180809, + 0.03407900780439377, + 0.012625939212739468, + 0.02906796894967556, + 0.015255160629749298, + 0.006783161778002977, + 0.007608053740113974, + 0.008273898623883724, + 0.008222715929150581, + 0.01259850524365902, + 0.007571259047836065, + 0.009422379545867443, + 0.006545140407979488, + 0.00564555823802948, + 0.01047180499881506, + 0.013143040239810944, + 0.014812511391937733, + 0.009457733482122421, + 0.00555616058409214, + 0.012498103082180023, + 0.009721139445900917, + 0.006308898329734802, + 0.007438416127115488, + 0.01442116778343916, + 0.012618972919881344, + 0.006839945912361145, + 0.006755456794053316, + 0.009423403069376945, + 0.009714368730783463, + 0.004321628715842962, + 0.007210992742329836, + 0.003677658038213849, + 0.00485596340149641, + 0.004616432823240757, + 0.007016138173639774, + 0.006195005029439926, + 0.004635415505617857, + 0.010206122882664204, + 0.021255847066640854, + 0.008286111988127232, + 0.016020016744732857, + 0.025105711072683334, + 0.00901892688125372, + 0.015910880640149117, + 0.009924013167619705, + 0.011590356007218361, + 0.01512889750301838, + 0.01749890297651291, + 0.013260209001600742, + 0.0109530845656991, + 0.012119009159505367, + 0.02604653872549534, + 0.013261480256915092, + 0.006916218437254429, + 0.009081230498850346, + 0.005679979920387268, + 0.022091064602136612, + 0.00563077675178647, + 0.011350140906870365, + 0.014299674890935421, + 0.0057182712480425835, + 0.01215367391705513, + 0.016867877915501595, + 0.009657008573412895, + 0.014360614120960236, + 0.010198106057941914, + 0.015078375115990639, + 0.017480261623859406, + 0.010370882228016853, + 0.01653890311717987, + 0.02738298289477825, + 0.016157062724232674, + 0.01842786930501461, + 0.012324788607656956, + 0.022065652534365654, + 0.025295276194810867, + 0.012296893633902073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.005324350204318762 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog galloping alongside a bright red tennis ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast expanse of rolling sand dunes with pale, fine sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a deep blue sea meeting a clear sky at the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054956551641225815 + ], + "opposite_logprob_tokens": [ + -2.9299564361572266 + ], + "logprob_reward": -0.054956551641225815, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054956551641225815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14517593383789062, + 0.03516705334186554, + 0.09205100685358047, + 0.02551138959825039, + 0.08448725938796997, + 0.18666106462478638, + 0.11572910845279694, + 0.06353992223739624, + 0.12097933143377304, + 0.13069792091846466 + ], + "credit_map": [ + 0.012328741140663624, + 0.01755804754793644, + 0.022757308557629585, + 0.014774926006793976, + 0.023003490641713142, + 0.012988168746232986, + 0.024214820936322212, + 0.01755044236779213, + 0.004114429000765085, + 0.0037932174745947123, + 0.003991583362221718, + 0.004968852270394564, + 0.004819292575120926, + 0.0056165531277656555, + 0.004177788272500038, + 0.003685340052470565, + 0.00589653430506587, + 0.012623426504433155, + 0.012211107648909092, + 0.01871379464864731, + 0.009624752216041088, + 0.006958932615816593, + 0.016500817611813545, + 0.009521638043224812, + 0.002765896962955594, + 0.003955083899199963, + 0.002579817082732916, + 0.004006829112768173, + 0.002596901496872306, + 0.0026419067289680243, + 0.004046768415719271, + 0.0029181870631873608, + 0.0069733052514493465, + 0.011355779133737087, + 0.004885262809693813, + 0.012141578830778599, + 0.006952446419745684, + 0.004976024851202965, + 0.021099064499139786, + 0.016103794798254967, + 0.01737719401717186, + 0.030291514471173286, + 0.013763570226728916, + 0.030619073659181595, + 0.03713954985141754, + 0.015730533748865128, + 0.02909279428422451, + 0.012646835297346115, + 0.010953293181955814, + 0.012417640537023544, + 0.014143755659461021, + 0.014235151000320911, + 0.011562795378267765, + 0.01432722806930542, + 0.02685912325978279, + 0.011230130679905415, + 0.005338246002793312, + 0.0080993277952075, + 0.006668997462838888, + 0.007396644912660122, + 0.006425109226256609, + 0.012623883783817291, + 0.01079789362847805, + 0.006189816165715456, + 0.011054597795009613, + 0.02124100551009178, + 0.012804793193936348, + 0.013995603658258915, + 0.01384799275547266, + 0.015269886702299118, + 0.020504534244537354, + 0.01226091105490923, + 0.016991963610053062, + 0.020766617730259895, + 0.014010516926646233, + 0.01648472622036934, + 0.012183126993477345, + 0.017554759979248047, + 0.021233921870589256, + 0.011472279205918312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.054956551641225815 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog galloping alongside a bright red tennis ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast expanse of rolling sand dunes with pale, fine sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a deep blue sea meeting a clear sky at the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011067195795476437 + ], + "opposite_logprob_tokens": [ + -4.5110673904418945 + ], + "logprob_reward": -0.011067195795476437, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011067195795476437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11161714047193527, + 0.05897555127739906, + 0.09583064913749695, + 0.13373565673828125, + 0.24768613278865814, + 0.04726804792881012, + 0.13556566834449768, + 0.08627000451087952, + 0.05505092069506645, + 0.0280002448707819 + ], + "credit_map": [ + 0.009573105722665787, + 0.017711952328681946, + 0.016974851489067078, + 0.008711667731404305, + 0.018579863011837006, + 0.013003697618842125, + 0.01735728047788143, + 0.009704720228910446, + 0.010474126785993576, + 0.005373538006097078, + 0.0065984479151666164, + 0.007637170143425465, + 0.007111359853297472, + 0.008139207027852535, + 0.006170534994453192, + 0.007471165154129267, + 0.00728094344958663, + 0.014158280566334724, + 0.01122557744383812, + 0.011887912638485432, + 0.009029502049088478, + 0.008671177551150322, + 0.015695767477154732, + 0.017881492152810097, + 0.016250746324658394, + 0.014252078719437122, + 0.018848123028874397, + 0.022509386762976646, + 0.011386443860828876, + 0.015955770388245583, + 0.020249998196959496, + 0.014283119700849056, + 0.02156379632651806, + 0.025358939543366432, + 0.031652312725782394, + 0.029692569747567177, + 0.02634000964462757, + 0.028969909995794296, + 0.056347813457250595, + 0.027760788798332214, + 0.004106747917830944, + 0.009219901636242867, + 0.0032123613636940718, + 0.007274184376001358, + 0.009746989235281944, + 0.0035183981526643038, + 0.006867906544357538, + 0.0033215591683983803, + 0.014047444798052311, + 0.017627926543354988, + 0.01340887974947691, + 0.01328305434435606, + 0.01504419930279255, + 0.01915379799902439, + 0.026349781081080437, + 0.01665058732032776, + 0.008326111361384392, + 0.01069527305662632, + 0.011227747425436974, + 0.011413928121328354, + 0.009386014193296432, + 0.012056705541908741, + 0.013967029750347137, + 0.009197196923196316, + 0.005353654734790325, + 0.0054999846033751965, + 0.0057409279979765415, + 0.008833162486553192, + 0.005422684364020824, + 0.006417536176741123, + 0.007463417015969753, + 0.01031955424696207, + 0.0027625877410173416, + 0.004280432127416134, + 0.0035622422583401203, + 0.004499116446822882, + 0.002251236466690898, + 0.0028568520210683346, + 0.005132611840963364, + 0.002655166434124112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 3 + ], + "selected_reward": -0.011067195795476437 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog galloping alongside a bright red tennis ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast expanse of rolling sand dunes with pale, fine sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a deep blue sea meeting a clear sky at the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009775033220648766 + ], + "opposite_logprob_tokens": [ + -4.634775161743164 + ], + "logprob_reward": -0.009775033220648766, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009775033220648766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15367525815963745, + 0.0391349270939827, + 0.10096192359924316, + 0.06525889784097672, + 0.09976061433553696, + 0.16021178662776947, + 0.09437455981969833, + 0.09718993306159973, + 0.1277342289686203, + 0.06169779226183891 + ], + "credit_map": [ + 0.011736825108528137, + 0.020070195198059082, + 0.03258150443434715, + 0.01126914843916893, + 0.02946363389492035, + 0.01065490860491991, + 0.02593064494431019, + 0.011968406848609447, + 0.005305761471390724, + 0.004802570678293705, + 0.005828119348734617, + 0.005118577741086483, + 0.0042548696510493755, + 0.0033662545029073954, + 0.00488863792270422, + 0.0055701397359371185, + 0.009168910793960094, + 0.012481973506510258, + 0.012118018232285976, + 0.01598619669675827, + 0.015097007155418396, + 0.010405406355857849, + 0.014299612492322922, + 0.011404814198613167, + 0.0062825544737279415, + 0.010917727835476398, + 0.008648797869682312, + 0.007198471110314131, + 0.007402567192912102, + 0.008272765204310417, + 0.009047274477779865, + 0.007488748989999294, + 0.00900948978960514, + 0.0077258008532226086, + 0.0061296122148633, + 0.012713689357042313, + 0.007309428881853819, + 0.009002378210425377, + 0.028429212048649788, + 0.019441010430455208, + 0.016212206333875656, + 0.023471791297197342, + 0.013363314792513847, + 0.027063991874456406, + 0.0220753476023674, + 0.016616148874163628, + 0.02580939047038555, + 0.015599619597196579, + 0.00866832584142685, + 0.010923155583441257, + 0.011857513338327408, + 0.010923375375568867, + 0.010748498141765594, + 0.009877934120595455, + 0.021994808688759804, + 0.009380960837006569, + 0.010189509019255638, + 0.01043474581092596, + 0.009186125360429287, + 0.018801994621753693, + 0.010819405317306519, + 0.014636380597949028, + 0.01560843177139759, + 0.007513349410146475, + 0.012098011560738087, + 0.012887437827885151, + 0.011297393590211868, + 0.010861326940357685, + 0.020811066031455994, + 0.02517535910010338, + 0.017902925610542297, + 0.016700729727745056, + 0.009194550104439259, + 0.009229420684278011, + 0.0065011377446353436, + 0.009111004881560802, + 0.005981627851724625, + 0.007241449784487486, + 0.010133510455489159, + 0.004305094946175814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.009775033220648766 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog galloping alongside a bright red tennis ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast expanse of rolling sand dunes with pale, fine sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a deep blue sea meeting a clear sky at the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01252993755042553 + ], + "opposite_logprob_tokens": [ + -4.3875298500061035 + ], + "logprob_reward": -0.01252993755042553, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01252993755042553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15281975269317627, + 0.03651757165789604, + 0.05161149054765701, + 0.2377031445503235, + 0.30246701836586, + 0.06295205652713776, + 0.04686671122908592, + 0.03187691792845726, + 0.02607669308781624, + 0.0511087030172348 + ], + "credit_map": [ + 0.013518452644348145, + 0.029670553281903267, + 0.019245851784944534, + 0.012352569960057735, + 0.03075428679585457, + 0.013984408229589462, + 0.022683149203658104, + 0.010610479861497879, + 0.005557843949645758, + 0.004494913388043642, + 0.004390976391732693, + 0.0042357430793344975, + 0.004449605476111174, + 0.003636943409219384, + 0.006241742521524429, + 0.003509803442284465, + 0.0061280070804059505, + 0.00691485870629549, + 0.005846984218806028, + 0.0061507150530815125, + 0.0078085362911224365, + 0.005015564616769552, + 0.00820702314376831, + 0.005539798177778721, + 0.04098580777645111, + 0.022600730881094933, + 0.03186240792274475, + 0.02217012643814087, + 0.025852523744106293, + 0.04026366397738457, + 0.030841531231999397, + 0.02312636375427246, + 0.02908233180642128, + 0.01825234666466713, + 0.018534602597355843, + 0.02850978821516037, + 0.01786351203918457, + 0.023252038285136223, + 0.09562958031892776, + 0.0713428482413292, + 0.005329689476639032, + 0.011285410262644291, + 0.005225257016718388, + 0.00867830403149128, + 0.012110686860978603, + 0.00552387023344636, + 0.009919194504618645, + 0.004879643674939871, + 0.0039716861210763454, + 0.006120722740888596, + 0.004194333683699369, + 0.007541189901530743, + 0.005542993079870939, + 0.0029058309737592936, + 0.012746565043926239, + 0.00384338921867311, + 0.0032654982060194016, + 0.004605108872056007, + 0.0035569246392697096, + 0.004682421684265137, + 0.0035210668575018644, + 0.003716424573212862, + 0.005635500885546207, + 0.002893972909078002, + 0.00510706752538681, + 0.0027538887225091457, + 0.0028100614435970783, + 0.003832523711025715, + 0.0020793024450540543, + 0.003982882481068373, + 0.0030763118993490934, + 0.0024346543941646814, + 0.005712560378015041, + 0.00760295893996954, + 0.006689975969493389, + 0.006951395887881517, + 0.004192066844552755, + 0.006175678689032793, + 0.007644550874829292, + 0.0061395177617669106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.01252993755042553 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic tracking shot, filmed from a low angle, keeps close to the ground, alongside a nimble golden retriever as it gallops across a vast expanse of rolling sand dunes with explosive energy. Its magnificent golden fur sways in the warm golden afternoon sun, its ears fluttering joyfully as it runs, its tongue lolling out of its mouth, its eyes filled with focus and pure excitement. Its target is a bright red tennis ball bouncing erratically along the dune slope ahead. With each powerful kick, its paws dig deep into the soft, fine, pale sand, stirring up a shower of tiny, sunlit grains. The entire scene is bathed in the warm hues of the setting sun, casting long, dramatic shadows across the rolling dunes. In the distance, the horizon is where the deep blue sea meets the clear sky, creating a dynamic atmosphere of boundless freedom and pure joy.", + "seed": 27569, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_14/index_1063_seed_27569_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.13724711525719613, + "vlm_logprob": -0.13724711525719613, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.13724711525719613, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog galloping alongside a bright red tennis ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast expanse of rolling sand dunes with pale, fine sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a deep blue sea meeting a clear sky at the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10302657634019852 + ], + "opposite_logprob_tokens": [ + -2.3530266284942627 + ], + "logprob_reward": -0.10302657634019852, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10302657634019852, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10415016114711761, + 0.0214792862534523, + 0.05429977923631668, + 0.05070233345031738, + 0.04814701899886131, + 0.14386068284511566, + 0.3053593039512634, + 0.055684469640254974, + 0.10455646365880966, + 0.11176052689552307 + ], + "credit_map": [ + 0.008848064579069614, + 0.014409055933356285, + 0.014011731371283531, + 0.02436009980738163, + 0.005676417611539364, + 0.008031093515455723, + 0.012223747558891773, + 0.016589952632784843, + 0.0017911893082782626, + 0.0018849180778488517, + 0.002784949028864503, + 0.0035155185032635927, + 0.0017297378508374095, + 0.0028984711971133947, + 0.0037556232418864965, + 0.0031188796274363995, + 0.004805644974112511, + 0.004983788356184959, + 0.0055817230604588985, + 0.015887396410107613, + 0.0028478503227233887, + 0.004975948017090559, + 0.005582278128713369, + 0.009635146707296371, + 0.0029298646841198206, + 0.008220440708100796, + 0.006852633785456419, + 0.00796768069267273, + 0.004560786299407482, + 0.004623479675501585, + 0.008784415200352669, + 0.0067630307748913765, + 0.0030846830923110247, + 0.004312002100050449, + 0.008073211647570133, + 0.009481056593358517, + 0.002738736569881439, + 0.004519618581980467, + 0.008768645115196705, + 0.007169063668698072, + 0.008525130338966846, + 0.014055595733225346, + 0.024343108758330345, + 0.03211599215865135, + 0.009676621295511723, + 0.010222976095974445, + 0.021586958318948746, + 0.023334307596087456, + 0.02482890710234642, + 0.045658502727746964, + 0.046604931354522705, + 0.058579329401254654, + 0.02252330258488655, + 0.032113343477249146, + 0.034088362008333206, + 0.04096259921789169, + 0.0032706710044294596, + 0.00371679593808949, + 0.012253541499376297, + 0.01393356453627348, + 0.006247894838452339, + 0.004263104405254126, + 0.006498073693364859, + 0.005500822328031063, + 0.005510908551514149, + 0.012480644509196281, + 0.009985939599573612, + 0.016605528071522713, + 0.01736867055296898, + 0.013861233368515968, + 0.018381575122475624, + 0.010361967608332634, + 0.007699803449213505, + 0.008431192487478256, + 0.015928979963064194, + 0.020564153790473938, + 0.011420436203479767, + 0.013696283102035522, + 0.021212173625826836, + 0.01280749961733818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 9 + ], + "selected_reward": -0.10302657634019852 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog galloping alongside a bright red tennis ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast expanse of rolling sand dunes with pale, fine sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a deep blue sea meeting a clear sky at the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233149528503418 + ], + "opposite_logprob_tokens": [ + -0.8983149528503418 + ], + "logprob_reward": -0.5233149528503418, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233149528503418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07290415465831757, + 0.12351968139410019, + 0.12331126630306244, + 0.12742139399051666, + 0.07568749785423279, + 0.06778595596551895, + 0.1301860213279724, + 0.1607949584722519, + 0.05897299572825432, + 0.059416063129901886 + ], + "credit_map": [ + 0.009297995828092098, + 0.009077847935259342, + 0.007317391224205494, + 0.012183844111859798, + 0.006289265118539333, + 0.010632345452904701, + 0.007978067733347416, + 0.010127393528819084, + 0.008903121575713158, + 0.015141758136451244, + 0.011531811207532883, + 0.03615276888012886, + 0.00699700228869915, + 0.01003488339483738, + 0.012906978838145733, + 0.02185135893523693, + 0.008464856073260307, + 0.011098942719399929, + 0.012188228778541088, + 0.038448866456747055, + 0.005667587276548147, + 0.012368491850793362, + 0.01115463674068451, + 0.023919658735394478, + 0.008176653645932674, + 0.01703268475830555, + 0.014066820032894611, + 0.0268117506057024, + 0.008983423002064228, + 0.012574716471135616, + 0.021895386278629303, + 0.017879951745271683, + 0.005738801322877407, + 0.009534426964819431, + 0.009142129682004452, + 0.011267264373600483, + 0.00525525351986289, + 0.005996671039611101, + 0.01621423102915287, + 0.012538718990981579, + 0.0036445281002670527, + 0.004602205008268356, + 0.009774367325007915, + 0.010293293744325638, + 0.004353785887360573, + 0.006572733633220196, + 0.018682774156332016, + 0.009862272068858147, + 0.011200682260096073, + 0.007239350583404303, + 0.03982902318239212, + 0.015567781403660774, + 0.00521065155044198, + 0.007959593087434769, + 0.024326378479599953, + 0.018852563574910164, + 0.009497730061411858, + 0.00982559472322464, + 0.045307617634534836, + 0.037044089287519455, + 0.01853901706635952, + 0.015444768592715263, + 0.013432232663035393, + 0.011703907512128353, + 0.0028442174661904573, + 0.004601180553436279, + 0.006835369858890772, + 0.009476828388869762, + 0.00950437132269144, + 0.00886593945324421, + 0.009698264300823212, + 0.007146822288632393, + 0.003585339291021228, + 0.00399838387966156, + 0.008061496540904045, + 0.008494826033711433, + 0.009452973492443562, + 0.004960242658853531, + 0.014283702708780766, + 0.006579095497727394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 3 + ], + "selected_reward": -0.5233149528503418 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog galloping alongside a bright red tennis ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast expanse of rolling sand dunes with pale, fine sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a deep blue sea meeting a clear sky at the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001335801207460463 + ], + "opposite_logprob_tokens": [ + -6.626335620880127 + ], + "logprob_reward": -0.001335801207460463, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001335801207460463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22922901809215546, + 0.06575556844472885, + 0.038816507905721664, + 0.038580961525440216, + 0.03285716101527214, + 0.06167216971516609, + 0.030890922993421555, + 0.30342260003089905, + 0.040329914540052414, + 0.15844516456127167 + ], + "credit_map": [ + 0.019846193492412567, + 0.037372902035713196, + 0.03333069756627083, + 0.037064142525196075, + 0.016357053071260452, + 0.020400799810886383, + 0.03091694973409176, + 0.033940281718969345, + 0.005155584774911404, + 0.004469778388738632, + 0.006571554113179445, + 0.010677160695195198, + 0.0064723901450634, + 0.015973173081874847, + 0.0063025630079209805, + 0.01013336144387722, + 0.0028221295215189457, + 0.003920563962310553, + 0.00556880934163928, + 0.006997447460889816, + 0.003986572381108999, + 0.0049664415419101715, + 0.005110094789415598, + 0.00544445076957345, + 0.0028163702227175236, + 0.005245797801762819, + 0.005036420654505491, + 0.007219642400741577, + 0.003315179143100977, + 0.004519685171544552, + 0.004874533042311668, + 0.005553330294787884, + 0.0023072194308042526, + 0.0028958905022591352, + 0.0057052504271268845, + 0.005655128974467516, + 0.0020576801616698503, + 0.0052663180977106094, + 0.0039770957082509995, + 0.004992576315999031, + 0.0037574106827378273, + 0.006186541169881821, + 0.008236953988671303, + 0.009908515959978104, + 0.003468321403488517, + 0.011618790216743946, + 0.008669666945934296, + 0.00982597004622221, + 0.0020439994987100363, + 0.003694994607940316, + 0.0043386672623455524, + 0.00577180553227663, + 0.0022479193285107613, + 0.002403713297098875, + 0.006247983314096928, + 0.00414184108376503, + 0.01334398239850998, + 0.013991123996675014, + 0.11271696537733078, + 0.08271848410367966, + 0.021028488874435425, + 0.0183409396559, + 0.020966658368706703, + 0.020315948873758316, + 0.0044068689458072186, + 0.0034680345561355352, + 0.005177836865186691, + 0.007422298192977905, + 0.0029188848566263914, + 0.005945479962974787, + 0.004544379189610481, + 0.006446133367717266, + 0.01800040528178215, + 0.0124130267649889, + 0.027496764436364174, + 0.04172129929065704, + 0.014446967281401157, + 0.009725522249937057, + 0.020208965986967087, + 0.01443222351372242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.001335801207460463 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog galloping alongside a bright red tennis ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast expanse of rolling sand dunes with pale, fine sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a deep blue sea meeting a clear sky at the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003614442190155387 + ], + "opposite_logprob_tokens": [ + -5.62861442565918 + ], + "logprob_reward": -0.003614442190155387, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003614442190155387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08868507295846939, + 0.07356411218643188, + 0.03285329416394234, + 0.031773582100868225, + 0.021441809833049774, + 0.035402730107307434, + 0.06907268613576889, + 0.1032366007566452, + 0.4133472740650177, + 0.13062286376953125 + ], + "credit_map": [ + 0.00782200787216425, + 0.013887617737054825, + 0.014888763427734375, + 0.013234518468379974, + 0.007234145421534777, + 0.010413270443677902, + 0.008299754001200199, + 0.012904997915029526, + 0.005651120562106371, + 0.005854668561369181, + 0.007063053082674742, + 0.009764865972101688, + 0.009583620354533195, + 0.017088187858462334, + 0.0076497141271829605, + 0.010908878408372402, + 0.002803198527544737, + 0.004244853742420673, + 0.0032482228707522154, + 0.006667846813797951, + 0.002564883092418313, + 0.005367612466216087, + 0.0033968547359108925, + 0.0045598228462040424, + 0.002451026812195778, + 0.002702611032873392, + 0.004578327294439077, + 0.004107228014618158, + 0.0015662861987948418, + 0.0036197465378791094, + 0.008032601326704025, + 0.004715756978839636, + 0.0016342339804396033, + 0.0022228744346648455, + 0.003722703317180276, + 0.004117853939533234, + 0.0013200673274695873, + 0.0028915258590132, + 0.002760368399322033, + 0.0027721829246729612, + 0.0018656537868082523, + 0.0025035603903234005, + 0.005369039252400398, + 0.008484878577291965, + 0.0023932589683681726, + 0.0038104066625237465, + 0.0049028354696929455, + 0.006073095835745335, + 0.006454163230955601, + 0.00410226546227932, + 0.02151334099471569, + 0.009405896998941898, + 0.003134043188765645, + 0.004279221873730421, + 0.010853714309632778, + 0.009330039843916893, + 0.00573852751404047, + 0.007249135058373213, + 0.025769632309675217, + 0.022385112941265106, + 0.010136505588889122, + 0.010534302331507206, + 0.01048884354531765, + 0.010934541001915932, + 0.03705492243170738, + 0.05199035257101059, + 0.05239745229482651, + 0.0809388980269432, + 0.031204693019390106, + 0.05040080472826958, + 0.05339598283171654, + 0.055964190512895584, + 0.014784089289605618, + 0.010519706644117832, + 0.022280046716332436, + 0.027372512966394424, + 0.014204158447682858, + 0.009904749691486359, + 0.019087575376033783, + 0.012470024637877941 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.003614442190155387 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog galloping alongside a bright red tennis ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast expanse of rolling sand dunes with pale, fine sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a deep blue sea meeting a clear sky at the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494380369782448 + ], + "opposite_logprob_tokens": [ + -2.929943799972534 + ], + "logprob_reward": -0.05494380369782448, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494380369782448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20994338393211365, + 0.10090736299753189, + 0.07532524317502975, + 0.046340566128492355, + 0.03193734586238861, + 0.0774664655327797, + 0.08108280599117279, + 0.11256351321935654, + 0.06602197140455246, + 0.19841133058071136 + ], + "credit_map": [ + 0.019676871597766876, + 0.0385793074965477, + 0.029252154752612114, + 0.0326116606593132, + 0.015665970742702484, + 0.014751373790204525, + 0.02523348294198513, + 0.034172553569078445, + 0.007148717064410448, + 0.008079190738499165, + 0.010896447114646435, + 0.015837201848626137, + 0.012571978382766247, + 0.020691178739070892, + 0.010080451145768166, + 0.015602195635437965, + 0.0055403877049684525, + 0.005728091578930616, + 0.008549406193196774, + 0.01748672127723694, + 0.005562301259487867, + 0.011644226498901844, + 0.009591625072062016, + 0.011222481727600098, + 0.0033237584866583347, + 0.003818143391981721, + 0.006767856422811747, + 0.0069048963487148285, + 0.0028307754546403885, + 0.007039771415293217, + 0.00918565597385168, + 0.006469709798693657, + 0.003108417149633169, + 0.0037521570920944214, + 0.0042691403068602085, + 0.004576306790113449, + 0.002241980517283082, + 0.003851843299344182, + 0.005537653341889381, + 0.004599845968186855, + 0.0049982392229139805, + 0.007498154416680336, + 0.011081475764513016, + 0.011364741250872612, + 0.004391741007566452, + 0.0150809520855546, + 0.010566038079559803, + 0.012485130690038204, + 0.007497530430555344, + 0.004003543872386217, + 0.030249638482928276, + 0.009790035896003246, + 0.0029415846802294254, + 0.004220231436192989, + 0.010562455281615257, + 0.01181778497993946, + 0.005461821332573891, + 0.004725414793938398, + 0.042432673275470734, + 0.031025663018226624, + 0.006692379247397184, + 0.006167761515825987, + 0.008261688984930515, + 0.007796114310622215, + 0.00636251037940383, + 0.009212885983288288, + 0.010035983286798, + 0.01113983429968357, + 0.0068621947430074215, + 0.006183438468724489, + 0.007574084680527449, + 0.008651032112538815, + 0.02330610156059265, + 0.016395561397075653, + 0.03357221931219101, + 0.047001827508211136, + 0.01716810092329979, + 0.015097652561962605, + 0.02414083108305931, + 0.021729035302996635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.05494380369782448 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic tracking shot, filmed from a low angle, keeps close to the ground, alongside a nimble golden retriever as it gallops across a vast expanse of rolling sand dunes with explosive energy. Its magnificent golden fur sways in the warm golden afternoon sun, its ears fluttering joyfully as it runs, its tongue lolling out of its mouth, its eyes filled with focus and pure excitement. Its target is a bright red tennis ball bouncing erratically along the dune slope ahead. With each powerful kick, its paws dig deep into the soft, fine, pale sand, stirring up a shower of tiny, sunlit grains. The entire scene is bathed in the warm hues of the setting sun, casting long, dramatic shadows across the rolling dunes. In the distance, the horizon is where the deep blue sea meets the clear sky, creating a dynamic atmosphere of boundless freedom and pure joy.", + "seed": 27570, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_14/index_1063_seed_27570_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.18620394645258784, + "vlm_logprob": -0.18620394645258784, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.18620394645258784, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog galloping alongside a bright red tennis ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast expanse of rolling sand dunes with pale, fine sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a deep blue sea meeting a clear sky at the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015709219500422478 + ], + "opposite_logprob_tokens": [ + -4.265709400177002 + ], + "logprob_reward": -0.015709219500422478, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.015709219500422478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10967795550823212, + 0.06268474459648132, + 0.01697484217584133, + 0.025354918092489243, + 0.08734925091266632, + 0.04749169200658798, + 0.05155324563384056, + 0.04820713400840759, + 0.07119534909725189, + 0.4795109033584595 + ], + "credit_map": [ + 0.017457209527492523, + 0.008808127604424953, + 0.01786920800805092, + 0.013662214390933514, + 0.009305269457399845, + 0.007212033029645681, + 0.01868053525686264, + 0.016683358699083328, + 0.005505041219294071, + 0.006568974815309048, + 0.007626094855368137, + 0.006856758613139391, + 0.008802125230431557, + 0.01388532668352127, + 0.006193188484758139, + 0.007247233763337135, + 0.0016505667008459568, + 0.0021062970627099276, + 0.001287944265641272, + 0.0018747696885839105, + 0.003756328020244837, + 0.0026379977352917194, + 0.0022960021160542965, + 0.0013649365864694118, + 0.002940637059509754, + 0.0032697846181690693, + 0.00299524562433362, + 0.004295951221138239, + 0.0025648237206041813, + 0.0028573849704116583, + 0.004035958554595709, + 0.0023951309267431498, + 0.008716261945664883, + 0.006115821190178394, + 0.010708427987992764, + 0.014606391079723835, + 0.011363600380718708, + 0.009802186861634254, + 0.011717739515006542, + 0.014318819157779217, + 0.0037707495503127575, + 0.005904058925807476, + 0.0039156703278422356, + 0.005684714298695326, + 0.006319476757198572, + 0.009731300175189972, + 0.007325064856559038, + 0.004840657580643892, + 0.0039053393993526697, + 0.006345948204398155, + 0.007259114645421505, + 0.0066412705928087234, + 0.005799801088869572, + 0.007704068440943956, + 0.006002673413604498, + 0.007895026355981827, + 0.0034398173447698355, + 0.0035121000837534666, + 0.006908982526510954, + 0.007501143496483564, + 0.007115907035768032, + 0.005699143745005131, + 0.006521956529468298, + 0.007508082315325737, + 0.004541842266917229, + 0.0057063535787165165, + 0.006846945267170668, + 0.010085908696055412, + 0.017206449061632156, + 0.008665152825415134, + 0.0107882646843791, + 0.007354439236223698, + 0.0351458340883255, + 0.048100609332323074, + 0.046137742698192596, + 0.05924495682120323, + 0.07680954039096832, + 0.08298231661319733, + 0.09478005021810532, + 0.03630988672375679 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.015709219500422478 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog galloping alongside a bright red tennis ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast expanse of rolling sand dunes with pale, fine sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a deep blue sea meeting a clear sky at the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8982682228088379 + ], + "opposite_logprob_tokens": [ + -0.5232682228088379 + ], + "logprob_reward": -0.8982682228088379, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8982682228088379, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.25214633345603943, + 0.07910400629043579, + 0.0407414473593235, + 0.029769496992230415, + 0.046914923936128616, + 0.04699857160449028, + 0.1532479077577591, + 0.09397668391466141, + 0.10628721117973328, + 0.15081344544887543 + ], + "credit_map": [ + 0.03023594245314598, + 0.021285921335220337, + 0.024776626378297806, + 0.026317277923226357, + 0.015709757804870605, + 0.011364821344614029, + 0.07425209134817123, + 0.04820389673113823, + 0.007161545101553202, + 0.008660771884024143, + 0.009330778382718563, + 0.009102653712034225, + 0.012589110992848873, + 0.013255237601697445, + 0.009225557558238506, + 0.009778350591659546, + 0.0028431888204067945, + 0.0036621103063225746, + 0.004023238550871611, + 0.006638322025537491, + 0.009262561798095703, + 0.006218980066478252, + 0.0046545108780264854, + 0.00343853491358459, + 0.003264468861743808, + 0.003920360002666712, + 0.0023445638362318277, + 0.0037624461110681295, + 0.004548420663923025, + 0.004343586042523384, + 0.005107400473207235, + 0.0024782519321888685, + 0.0046638306230306625, + 0.0032465083058923483, + 0.004327275790274143, + 0.006990984082221985, + 0.007615434005856514, + 0.006922345142811537, + 0.007371469866484404, + 0.005777076352387667, + 0.003137125400826335, + 0.004121970385313034, + 0.00466268602758646, + 0.007865391671657562, + 0.0036623075138777494, + 0.006092255935072899, + 0.01154540479183197, + 0.005911430809646845, + 0.011525261215865612, + 0.01712120696902275, + 0.02106013149023056, + 0.02063162811100483, + 0.0197767224162817, + 0.014020961709320545, + 0.017309268936514854, + 0.03180274739861488, + 0.008477091789245605, + 0.010689007118344307, + 0.009096608497202396, + 0.013333916664123535, + 0.0178314670920372, + 0.008326890878379345, + 0.013332338072359562, + 0.012889367528259754, + 0.012378549203276634, + 0.012089400552213192, + 0.011220677755773067, + 0.014139755629003048, + 0.014967890456318855, + 0.008412313647568226, + 0.019793594256043434, + 0.013285025954246521, + 0.010163755156099796, + 0.013058862648904324, + 0.012836607173085213, + 0.02020891010761261, + 0.022369664162397385, + 0.022585950791835785, + 0.039597708731889725, + 0.009991989471018314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.8982682228088379 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog galloping alongside a bright red tennis ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast expanse of rolling sand dunes with pale, fine sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a deep blue sea meeting a clear sky at the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001335205975919962 + ], + "opposite_logprob_tokens": [ + -6.626335144042969 + ], + "logprob_reward": -0.001335205975919962, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001335205975919962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21562324464321136, + 0.07455302774906158, + 0.10129570960998535, + 0.03246886655688286, + 0.2443002164363861, + 0.05066952481865883, + 0.07241720706224442, + 0.06262163072824478, + 0.050078801810741425, + 0.09597175568342209 + ], + "credit_map": [ + 0.031244229525327682, + 0.020365949720144272, + 0.026825381442904472, + 0.022696971893310547, + 0.014829691499471664, + 0.013907303102314472, + 0.045214127749204636, + 0.040539611130952835, + 0.006730033550411463, + 0.00632965425029397, + 0.010599392466247082, + 0.014067462645471096, + 0.009845650754868984, + 0.007720191031694412, + 0.011917470023036003, + 0.007343168370425701, + 0.00951672438532114, + 0.010151919908821583, + 0.011807672679424286, + 0.0217982716858387, + 0.013962462544441223, + 0.01087080966681242, + 0.014294309541583061, + 0.008893541060388088, + 0.0036076423712074757, + 0.00408277427777648, + 0.0023297383449971676, + 0.004372446797788143, + 0.0044829160906374454, + 0.004552656318992376, + 0.005986184347420931, + 0.0030545059125870466, + 0.01908002234995365, + 0.017132943496108055, + 0.057283807545900345, + 0.03302709013223648, + 0.019936177879571915, + 0.019381804391741753, + 0.028707453981041908, + 0.049750916659832, + 0.004198999144136906, + 0.005550561007112265, + 0.00991357583552599, + 0.0068738991394639015, + 0.004175685346126556, + 0.00509952288120985, + 0.009340427815914154, + 0.005516855977475643, + 0.005438691005110741, + 0.006441160570830107, + 0.009415381588041782, + 0.01101713441312313, + 0.009366778656840324, + 0.005212351679801941, + 0.009007212705910206, + 0.016518495976924896, + 0.006419979967176914, + 0.008817088790237904, + 0.006053105462342501, + 0.009141342714428902, + 0.010296778753399849, + 0.006152295973151922, + 0.008145884610712528, + 0.007595154456794262, + 0.004794174339622259, + 0.007232305593788624, + 0.006083284039050341, + 0.005914460867643356, + 0.007031801622360945, + 0.0038077549543231726, + 0.009221654385328293, + 0.005993365775793791, + 0.008858355693519115, + 0.014264569617807865, + 0.01060627680271864, + 0.012499019503593445, + 0.01114344596862793, + 0.013593699783086777, + 0.018534444272518158, + 0.0064719486981630325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.001335205975919962 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog galloping alongside a bright red tennis ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast expanse of rolling sand dunes with pale, fine sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a deep blue sea meeting a clear sky at the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011073797941207886 + ], + "opposite_logprob_tokens": [ + -4.511073589324951 + ], + "logprob_reward": -0.011073797941207886, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011073797941207886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3687495291233063, + 0.048872627317905426, + 0.042770545929670334, + 0.08493116497993469, + 0.07776635140180588, + 0.05545554310083389, + 0.08936131745576859, + 0.10743822157382965, + 0.05600202828645706, + 0.0686526745557785 + ], + "credit_map": [ + 0.06535661220550537, + 0.0308151226490736, + 0.05441107973456383, + 0.027268828824162483, + 0.018823545426130295, + 0.026056958362460136, + 0.045568883419036865, + 0.10044852644205093, + 0.00412547355517745, + 0.004427018575370312, + 0.0058157481253147125, + 0.0053259097039699554, + 0.008623671717941761, + 0.00703846663236618, + 0.007554203737527132, + 0.005962133873254061, + 0.004359206184744835, + 0.0037632982712239027, + 0.0047932942397892475, + 0.006183639634400606, + 0.006752736866474152, + 0.0057500810362398624, + 0.006776571273803711, + 0.004391718655824661, + 0.010579438880085945, + 0.010411069728434086, + 0.008372966200113297, + 0.014287597499787807, + 0.009721874259412289, + 0.008624264039099216, + 0.014186283573508263, + 0.008747673593461514, + 0.00767061160877347, + 0.0060839406214654446, + 0.01632831059396267, + 0.007581873331218958, + 0.010910146869719028, + 0.007770245429128408, + 0.011120152659714222, + 0.010301071219146252, + 0.005879971198737621, + 0.005654765293002129, + 0.00807708129286766, + 0.00698454026132822, + 0.005939203780144453, + 0.00592532753944397, + 0.009678537957370281, + 0.007316112983971834, + 0.008283552713692188, + 0.01058268267661333, + 0.014690620824694633, + 0.01072382740676403, + 0.011692710220813751, + 0.00818100105971098, + 0.009973780252039433, + 0.015233136713504791, + 0.007176908198744059, + 0.009869348257780075, + 0.01245145220309496, + 0.015184120275080204, + 0.021064678207039833, + 0.01412122044712305, + 0.01303546316921711, + 0.014535042457282543, + 0.004669690039008856, + 0.005669260397553444, + 0.0059331683441996574, + 0.007249458692967892, + 0.010570391081273556, + 0.0061192926950752735, + 0.009571877308189869, + 0.006218887399882078, + 0.0044813151471316814, + 0.007120165042579174, + 0.006890085060149431, + 0.01121353916823864, + 0.012333076447248459, + 0.009893265552818775, + 0.011767379008233547, + 0.004953848663717508 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.011073797941207886 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog galloping alongside a bright red tennis ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast expanse of rolling sand dunes with pale, fine sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a deep blue sea meeting a clear sky at the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633286036550999 + ], + "opposite_logprob_tokens": [ + -5.37963342666626 + ], + "logprob_reward": -0.004633286036550999, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633286036550999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17235389351844788, + 0.21682335436344147, + 0.05624764412641525, + 0.03420216217637062, + 0.037832025438547134, + 0.0926879420876503, + 0.08731977641582489, + 0.08746843785047531, + 0.08455883711576462, + 0.13050591945648193 + ], + "credit_map": [ + 0.024375800043344498, + 0.016267020255327225, + 0.030534040182828903, + 0.014346937648952007, + 0.010377536527812481, + 0.018080487847328186, + 0.02929454855620861, + 0.029077520594000816, + 0.023090118542313576, + 0.026173148304224014, + 0.02355918101966381, + 0.022338813170790672, + 0.03372256085276604, + 0.03458905220031738, + 0.02385232225060463, + 0.029498161748051643, + 0.004011781886219978, + 0.005052557680755854, + 0.005681686569005251, + 0.009809501469135284, + 0.01087792869657278, + 0.005693037062883377, + 0.00863944087177515, + 0.006481708027422428, + 0.004463380202651024, + 0.004107352811843157, + 0.0032550788018852472, + 0.005477054510265589, + 0.0029254057444632053, + 0.004573536105453968, + 0.0061018625274300575, + 0.0032984900753945112, + 0.00247383676469326, + 0.002905044471845031, + 0.005794058088213205, + 0.0056689027696847916, + 0.00517582893371582, + 0.004049262497574091, + 0.0053081875666975975, + 0.006456905510276556, + 0.007855592295527458, + 0.010288200341165066, + 0.006364347413182259, + 0.014084948226809502, + 0.008571367710828781, + 0.017294498160481453, + 0.017052404582500458, + 0.011176577769219875, + 0.005888964049518108, + 0.00903067085891962, + 0.012290171347558498, + 0.012955715879797935, + 0.011044887825846672, + 0.0070167300291359425, + 0.0103362537920475, + 0.018756387755274773, + 0.009212599135935307, + 0.01297940593212843, + 0.00876486673951149, + 0.013772613368928432, + 0.01116316020488739, + 0.006259740795940161, + 0.01724177412688732, + 0.008074280805885792, + 0.011881857179105282, + 0.013186657801270485, + 0.00819169171154499, + 0.009815368801355362, + 0.011000401340425014, + 0.009943162091076374, + 0.010179312899708748, + 0.010360379703342915, + 0.01600945182144642, + 0.022609766572713852, + 0.017452824860811234, + 0.015942253172397614, + 0.012896208092570305, + 0.012598478235304356, + 0.019401835277676582, + 0.013595102354884148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.004633286036550999 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic tracking shot, filmed from a low angle, keeps close to the ground, alongside a nimble golden retriever as it gallops across a vast expanse of rolling sand dunes with explosive energy. Its magnificent golden fur sways in the warm golden afternoon sun, its ears fluttering joyfully as it runs, its tongue lolling out of its mouth, its eyes filled with focus and pure excitement. Its target is a bright red tennis ball bouncing erratically along the dune slope ahead. With each powerful kick, its paws dig deep into the soft, fine, pale sand, stirring up a shower of tiny, sunlit grains. The entire scene is bathed in the warm hues of the setting sun, casting long, dramatic shadows across the rolling dunes. In the distance, the horizon is where the deep blue sea meets the clear sky, creating a dynamic atmosphere of boundless freedom and pure joy.", + "seed": 27571, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_14/index_1063_seed_27571_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03762109987437725, + "vlm_logprob": -0.03762109987437725, + "vlm_yes_no_margin": 4.124999976158142, + "vlm_token_credit": -0.03762109987437725, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog galloping alongside a bright red tennis ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast expanse of rolling sand dunes with pale, fine sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a deep blue sea meeting a clear sky at the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021381352096796036 + ], + "opposite_logprob_tokens": [ + -4.021381378173828 + ], + "logprob_reward": -0.021381352096796036, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.021381352096796036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.257788747549057, + 0.03888704627752304, + 0.032529182732105255, + 0.08382114768028259, + 0.037403494119644165, + 0.12687553465366364, + 0.044714078307151794, + 0.09077166765928268, + 0.20844879746437073, + 0.0787602886557579 + ], + "credit_map": [ + 0.0318383164703846, + 0.021138226613402367, + 0.01752256602048874, + 0.033855799585580826, + 0.036111969500780106, + 0.056504812091588974, + 0.03370935469865799, + 0.02710769884288311, + 0.004665490239858627, + 0.006101884879171848, + 0.0033421963453292847, + 0.0042106411419808865, + 0.005023679696023464, + 0.005049244035035372, + 0.0055306400172412395, + 0.004963272251188755, + 0.0028772172518074512, + 0.003636910580098629, + 0.0029351776465773582, + 0.004035484045743942, + 0.005888734944164753, + 0.002846106421202421, + 0.0048980796709656715, + 0.005411472637206316, + 0.010339947417378426, + 0.006567107047885656, + 0.012292343191802502, + 0.01617131382226944, + 0.008425346575677395, + 0.01270064152777195, + 0.009669658727943897, + 0.007654793094843626, + 0.0038224956952035427, + 0.003126023802906275, + 0.004509876947849989, + 0.00357192219235003, + 0.0030125021003186703, + 0.0034658515360206366, + 0.004651285707950592, + 0.01124353613704443, + 0.015510435216128826, + 0.017434999346733093, + 0.011331233195960522, + 0.009976035915315151, + 0.010449162684381008, + 0.011693413369357586, + 0.015347589738667011, + 0.035132668912410736, + 0.005854491610080004, + 0.0030082371085882187, + 0.0023850712459534407, + 0.002665542531758547, + 0.0065472484566271305, + 0.003239148063585162, + 0.013287395238876343, + 0.007726945914328098, + 0.006259390618652105, + 0.00835771020501852, + 0.008061952888965607, + 0.006970701273530722, + 0.0191790834069252, + 0.0091157341375947, + 0.022697435691952705, + 0.010129662230610847, + 0.03456420078873634, + 0.04464038833975792, + 0.009499660693109035, + 0.01272391714155674, + 0.025933707132935524, + 0.01951548084616661, + 0.04299469292163849, + 0.018576741218566895, + 0.010170373134315014, + 0.0064726658165454865, + 0.0036738531198352575, + 0.0038564095739275217, + 0.010595908388495445, + 0.004425603896379471, + 0.028641588985919952, + 0.010923884809017181 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.021381352096796036 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog galloping alongside a bright red tennis ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast expanse of rolling sand dunes with pale, fine sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a deep blue sea meeting a clear sky at the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14282381534576416 + ], + "opposite_logprob_tokens": [ + -2.0178236961364746 + ], + "logprob_reward": -0.14282381534576416, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14282381534576416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04652271419763565, + 0.034655336290597916, + 0.03698822483420372, + 0.10133586078882217, + 0.06728145480155945, + 0.48189082741737366, + 0.06107218191027641, + 0.05489096790552139, + 0.07687095552682877, + 0.03849149867892265 + ], + "credit_map": [ + 0.004566799849271774, + 0.005679816007614136, + 0.0037114336155354977, + 0.006740154232829809, + 0.0060005891136825085, + 0.006434815004467964, + 0.008081907406449318, + 0.005307201761752367, + 0.00375307141803205, + 0.0038953700568526983, + 0.0032660476863384247, + 0.005738504230976105, + 0.0038325863424688578, + 0.004776407033205032, + 0.005581314209848642, + 0.003812036942690611, + 0.006366330664604902, + 0.004628047347068787, + 0.002406365005299449, + 0.0033527049235999584, + 0.0033296935725957155, + 0.008229616098105907, + 0.004991911351680756, + 0.003683556104078889, + 0.025009488686919212, + 0.006345133297145367, + 0.00876801460981369, + 0.007835187949240208, + 0.00799973588436842, + 0.028287023305892944, + 0.008123934268951416, + 0.008967339992523193, + 0.008180261589586735, + 0.005784664303064346, + 0.006010128650814295, + 0.009544060565531254, + 0.010139127261936665, + 0.007605771534144878, + 0.011988808400928974, + 0.008028626441955566, + 0.06193994730710983, + 0.0708218440413475, + 0.03836984559893608, + 0.03375659137964249, + 0.04207634925842285, + 0.04401939734816551, + 0.05628521367907524, + 0.13462163507938385, + 0.005638555157929659, + 0.004887229297310114, + 0.0033919906709343195, + 0.0054577975533902645, + 0.015098904259502888, + 0.0051521011628210545, + 0.007962708361446857, + 0.013482891954481602, + 0.006375831086188555, + 0.00922870822250843, + 0.0034809161443263292, + 0.004060433711856604, + 0.011034098453819752, + 0.004568792413920164, + 0.009714752435684204, + 0.006427435204386711, + 0.01213410496711731, + 0.015588595531880856, + 0.0037164788227528334, + 0.005310860462486744, + 0.00942393857985735, + 0.0063800509087741375, + 0.016405625268816948, + 0.007911298424005508, + 0.005462757311761379, + 0.004960485268384218, + 0.002886146539822221, + 0.0035047687124460936, + 0.004606029950082302, + 0.0026721961330622435, + 0.009633084759116173, + 0.004766028840094805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 8 + ], + "selected_reward": -0.14282381534576416 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog galloping alongside a bright red tennis ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast expanse of rolling sand dunes with pale, fine sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a deep blue sea meeting a clear sky at the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004634710028767586 + ], + "opposite_logprob_tokens": [ + -5.379634857177734 + ], + "logprob_reward": -0.004634710028767586, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004634710028767586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13770397007465363, + 0.08852225542068481, + 0.07970179617404938, + 0.09366815537214279, + 0.11216985434293747, + 0.16235953569412231, + 0.03631272166967392, + 0.07221148908138275, + 0.1481855809688568, + 0.06916466355323792 + ], + "credit_map": [ + 0.02169875055551529, + 0.011707755737006664, + 0.01725073531270027, + 0.00929670874029398, + 0.009011588990688324, + 0.037757739424705505, + 0.02033149264752865, + 0.010649193078279495, + 0.011164319701492786, + 0.015205801464617252, + 0.010125712491571903, + 0.010555526241660118, + 0.008520536124706268, + 0.011764882132411003, + 0.011162273585796356, + 0.010023202747106552, + 0.01500290259718895, + 0.010132602415978909, + 0.006266459822654724, + 0.006483423989266157, + 0.0065439739264547825, + 0.016970505937933922, + 0.007872250862419605, + 0.010429676622152328, + 0.011685492470860481, + 0.007558146025985479, + 0.012032832019031048, + 0.0164806991815567, + 0.009684366174042225, + 0.014550082385540009, + 0.011288292706012726, + 0.010388238355517387, + 0.011486086994409561, + 0.010509255342185497, + 0.014322708360850811, + 0.011959894560277462, + 0.00742189958691597, + 0.011069664731621742, + 0.013461815193295479, + 0.03193852677941322, + 0.02441605180501938, + 0.028566066175699234, + 0.01221819780766964, + 0.015731262043118477, + 0.016037434339523315, + 0.023807428777217865, + 0.015309557318687439, + 0.026273543015122414, + 0.004070758353918791, + 0.003274444257840514, + 0.0018435284728184342, + 0.0029527863953262568, + 0.005771963857114315, + 0.003779456950724125, + 0.0074371593073010445, + 0.007182624656707048, + 0.007569709327071905, + 0.006450631655752659, + 0.010700126178562641, + 0.008894904516637325, + 0.011765026487410069, + 0.008409863337874413, + 0.01170123741030693, + 0.0067199887707829475, + 0.01831289567053318, + 0.02729172073304653, + 0.011753784492611885, + 0.01506813708692789, + 0.011824432760477066, + 0.00967599917203188, + 0.030847648158669472, + 0.023410966619849205, + 0.006986901629716158, + 0.004614294972270727, + 0.004826672840863466, + 0.005595851223915815, + 0.009318030439317226, + 0.007217625621706247, + 0.01986710913479328, + 0.010738176293671131 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.004634710028767586 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog galloping alongside a bright red tennis ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast expanse of rolling sand dunes with pale, fine sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a deep blue sea meeting a clear sky at the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535471469163895 + ], + "opposite_logprob_tokens": [ + -4.387535572052002 + ], + "logprob_reward": -0.012535471469163895, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535471469163895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12236561626195908, + 0.11899042874574661, + 0.09453760832548141, + 0.08900163322687149, + 0.06308532506227493, + 0.1694282740354538, + 0.13403233885765076, + 0.06057966500520706, + 0.07734972983598709, + 0.07062936574220657 + ], + "credit_map": [ + 0.019933536648750305, + 0.013887906447052956, + 0.011204523965716362, + 0.013149439357221127, + 0.0179580170661211, + 0.017024731263518333, + 0.01751774176955223, + 0.011689719744026661, + 0.015082404948771, + 0.024983631446957588, + 0.010372506454586983, + 0.015441499650478363, + 0.010441617108881474, + 0.0182789359241724, + 0.014408526010811329, + 0.00998130813241005, + 0.015603813342750072, + 0.009768546558916569, + 0.006114039104431868, + 0.010099874809384346, + 0.008588623255491257, + 0.02282819151878357, + 0.01425101701170206, + 0.0072834971360862255, + 0.017771219834685326, + 0.005428394302725792, + 0.009390889666974545, + 0.01048740278929472, + 0.006658874452114105, + 0.019093995913863182, + 0.008261619135737419, + 0.011909239925444126, + 0.008930441923439503, + 0.005257660057395697, + 0.006409646011888981, + 0.006607753224670887, + 0.008121520280838013, + 0.009292012080550194, + 0.010322864167392254, + 0.008143424056470394, + 0.02985849604010582, + 0.02813531830906868, + 0.013181555084884167, + 0.015465056523680687, + 0.017639128491282463, + 0.01643216982483864, + 0.017160363495349884, + 0.031556181609630585, + 0.01606828346848488, + 0.008127497509121895, + 0.011960643343627453, + 0.017190488055348396, + 0.0165602657943964, + 0.010953309945762157, + 0.032474201172590256, + 0.020697643980383873, + 0.006845599040389061, + 0.009252915158867836, + 0.006658314261585474, + 0.0068716444075107574, + 0.009153928607702255, + 0.005184107460081577, + 0.010159299708902836, + 0.006453854031860828, + 0.007414914667606354, + 0.008915405720472336, + 0.004866424482315779, + 0.008436810225248337, + 0.015189461410045624, + 0.012909059412777424, + 0.012166586704552174, + 0.0074510639533400536, + 0.007740833330899477, + 0.0062990980222821236, + 0.004215457942336798, + 0.005713098682463169, + 0.010704799555242062, + 0.009201006032526493, + 0.015983352437615395, + 0.010771723464131355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.012535471469163895 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog galloping alongside a bright red tennis ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast expanse of rolling sand dunes with pale, fine sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene filmed from a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a deep blue sea meeting a clear sky at the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006730150431394577 + ], + "opposite_logprob_tokens": [ + -5.006730079650879 + ], + "logprob_reward": -0.006730150431394577, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006730150431394577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12715213000774384, + 0.05665910243988037, + 0.10923073440790176, + 0.11817663162946701, + 0.022790128365159035, + 0.0839972198009491, + 0.11057676374912262, + 0.057504039257764816, + 0.11049765348434448, + 0.2034156173467636 + ], + "credit_map": [ + 0.021431175991892815, + 0.013306887820363045, + 0.008071227930486202, + 0.010493442416191101, + 0.012187869288027287, + 0.036395058035850525, + 0.013834851793944836, + 0.011431614868342876, + 0.010464091785252094, + 0.014222505502402782, + 0.0047870720736682415, + 0.004155566915869713, + 0.004638687241822481, + 0.008642103523015976, + 0.00478132301941514, + 0.004967748187482357, + 0.021446140483021736, + 0.01615128479897976, + 0.008366389200091362, + 0.01031168457120657, + 0.00592100340873003, + 0.026473643258213997, + 0.011373592540621758, + 0.009186988696455956, + 0.015674883499741554, + 0.008953803218901157, + 0.017215263098478317, + 0.02078971266746521, + 0.012469462119042873, + 0.02025163732469082, + 0.013788578100502491, + 0.00903328787535429, + 0.0027361896354705095, + 0.002031786600127816, + 0.0027764015831053257, + 0.0026519703678786755, + 0.0030179161112755537, + 0.0028295766096562147, + 0.002831194782629609, + 0.003915090579539537, + 0.009172651916742325, + 0.0064733815379440784, + 0.007535601034760475, + 0.00556203443557024, + 0.0076880124397575855, + 0.009290501475334167, + 0.011116484180092812, + 0.027158550918102264, + 0.013839575462043285, + 0.007062029093503952, + 0.009070335887372494, + 0.01164122112095356, + 0.017447195947170258, + 0.007490759249776602, + 0.029825404286384583, + 0.014200234785676003, + 0.006135261617600918, + 0.008414858020842075, + 0.004531707614660263, + 0.004142859950661659, + 0.01072068139910698, + 0.004745025187730789, + 0.013372626155614853, + 0.005441016051918268, + 0.0076416609808802605, + 0.008902379311621189, + 0.007758056279271841, + 0.010629420168697834, + 0.027685901150107384, + 0.02324661985039711, + 0.015118892304599285, + 0.009514715522527695, + 0.027623960748314857, + 0.017467571422457695, + 0.01015437114983797, + 0.013296709395945072, + 0.025420906022191048, + 0.015289104543626308, + 0.06815160065889359, + 0.026011373847723007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.006730150431394577 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_15/index_tensor([8954, 8954])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_15/index_tensor([8954, 8954])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b61203344407b0b39ade7cc14d63d8da19232339 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_15/index_tensor([8954, 8954])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Set in a beautiful Ghibli-inspired art style, a low-angle medium shot unfolds across a sun-dappled forest clearing as two magnificent Senju chickens engage in a dramatic duel. Their feathers, a painterly cascade of iridescent greens, blues, and purples, shimmer with every twitch, creating a stark aggression against their proud crimson plumage. They leap and soar with astonishing agility, their powerful legs kicking up swirling clouds of fine dust and dry leaves, capturing the soft, warm light filtering through the dense forest canopy. As one of the combatants crouches, then explosively soars into the air for a dramatic attack, the camera pans fluidly upward and right, following its graceful arc. Against a backdrop of lush green foliage and hazy golden sunlight, the action captures the fluid movement of its outstretched wings and extended claws, creating a vibrant and captivating atmosphere that is both tense and beautiful.", + "seed": 12634, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_15/index_8954_seed_12634_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02917245470453054, + "vlm_logprob": -0.02917245470453054, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.02917245470453054, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two chickens visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the chickens display feathers with iridescent greens, blues, and purples mixed with crimson plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the chickens performing leaping and soaring motions while kicking up dust and dry leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-dappled forest clearing with lush green foliage and hazy golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan fluidly upward and right to follow the chickens' movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04029287025332451 + ], + "opposite_logprob_tokens": [ + -3.290292978286743 + ], + "logprob_reward": -0.04029287025332451, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04029287025332451, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05099862441420555, + 0.03759673982858658, + 0.1737707257270813, + 0.05504379794001579, + 0.14300335943698883, + 0.030624497681856155, + 0.16454222798347473, + 0.08922004699707031, + 0.20602495968341827, + 0.049174997955560684 + ], + "credit_map": [ + 0.0030287345871329308, + 0.004522043280303478, + 0.0056413342244923115, + 0.011411079205572605, + 0.008493777364492416, + 0.006864701397716999, + 0.006282090675085783, + 0.004754864610731602, + 0.0022434175480157137, + 0.003576472867280245, + 0.0036700873170048, + 0.007646988611668348, + 0.006015049759298563, + 0.005152978468686342, + 0.004973209463059902, + 0.00431853486225009, + 0.0088414391502738, + 0.013716952875256538, + 0.01647782139480114, + 0.04430617764592171, + 0.025161627680063248, + 0.022889329120516777, + 0.02207304909825325, + 0.020304329693317413, + 0.004041918087750673, + 0.004542200826108456, + 0.00578713184222579, + 0.007527545560151339, + 0.00590816605836153, + 0.006899415981024504, + 0.010850085876882076, + 0.009487333707511425, + 0.009988782927393913, + 0.00888906978070736, + 0.051939770579338074, + 0.015309428796172142, + 0.005862363148480654, + 0.03134595602750778, + 0.006241535302251577, + 0.01342645101249218, + 0.002197688678279519, + 0.004231318831443787, + 0.0030035125091671944, + 0.0049823978915810585, + 0.0030756962951272726, + 0.004277672152966261, + 0.005065555218607187, + 0.003790657501667738, + 0.012179522775113583, + 0.025874920189380646, + 0.02270190604031086, + 0.02689393237233162, + 0.022585023194551468, + 0.016000626608729362, + 0.020846478641033173, + 0.017459820955991745, + 0.02524915710091591, + 0.010697939433157444, + 0.01524122804403305, + 0.005725907627493143, + 0.0042428248561918736, + 0.006453806534409523, + 0.0075784604996442795, + 0.01403072476387024, + 0.04914755001664162, + 0.016093548387289047, + 0.03568386659026146, + 0.019109265878796577, + 0.01388577837496996, + 0.025657031685113907, + 0.020178496837615967, + 0.02626940608024597, + 0.00557754235342145, + 0.010250629857182503, + 0.007254874333739281, + 0.005502235144376755, + 0.004918929189443588, + 0.005372270941734314, + 0.0038053591269999743, + 0.006493162363767624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 6 + ], + "selected_reward": -0.04029287025332451 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two chickens visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the chickens display feathers with iridescent greens, blues, and purples mixed with crimson plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the chickens performing leaping and soaring motions while kicking up dust and dry leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-dappled forest clearing with lush green foliage and hazy golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan fluidly upward and right to follow the chickens' movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608537696301937 + ], + "opposite_logprob_tokens": [ + -4.141085147857666 + ], + "logprob_reward": -0.01608537696301937, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608537696301937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15524806082248688, + 0.15487737953662872, + 0.10430317372083664, + 0.08830693364143372, + 0.19278249144554138, + 0.03855304792523384, + 0.07436562329530716, + 0.0757729783654213, + 0.06492913514375687, + 0.05086112394928932 + ], + "credit_map": [ + 0.008513517677783966, + 0.020714111626148224, + 0.019930081441998482, + 0.02022348716855049, + 0.024663979187607765, + 0.02423911727964878, + 0.01954491436481476, + 0.01741885580122471, + 0.00974349770694971, + 0.019534679129719734, + 0.013844514265656471, + 0.039220742881298065, + 0.01846539042890072, + 0.017951594665646553, + 0.018300442025065422, + 0.017816517502069473, + 0.006017529405653477, + 0.008641672320663929, + 0.009057491086423397, + 0.027239125221967697, + 0.017363514751195908, + 0.0128984609618783, + 0.011883207596838474, + 0.011202176101505756, + 0.007916112430393696, + 0.010901344008743763, + 0.01083013042807579, + 0.00914039183408022, + 0.012400343082845211, + 0.011835102923214436, + 0.012561697512865067, + 0.012721817009150982, + 0.012413978576660156, + 0.02586774341762066, + 0.046583496034145355, + 0.02042626589536667, + 0.017945606261491776, + 0.024198167026042938, + 0.015868013724684715, + 0.029479214921593666, + 0.0035257828421890736, + 0.007516281213611364, + 0.004327225033193827, + 0.005537156015634537, + 0.003095855237916112, + 0.005604862235486507, + 0.004512263927608728, + 0.004433621186763048, + 0.006041266955435276, + 0.013879375532269478, + 0.007239065133035183, + 0.011737210676074028, + 0.008624833077192307, + 0.00709168566390872, + 0.01060404535382986, + 0.009148144163191319, + 0.018258240073919296, + 0.008908395655453205, + 0.01131361722946167, + 0.006209217943251133, + 0.006873088423162699, + 0.006089142989367247, + 0.006810553837567568, + 0.01131072174757719, + 0.010635272599756718, + 0.005373178981244564, + 0.00907611008733511, + 0.010628174059092999, + 0.005716890096664429, + 0.005682060960680246, + 0.009928434155881405, + 0.007889017462730408, + 0.005221528001129627, + 0.010288619436323643, + 0.007566513959318399, + 0.007554009091109037, + 0.004889441654086113, + 0.005572942551225424, + 0.004453487228602171, + 0.005314583424478769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.01608537696301937 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two chickens visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the chickens display feathers with iridescent greens, blues, and purples mixed with crimson plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the chickens performing leaping and soaring motions while kicking up dust and dry leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-dappled forest clearing with lush green foliage and hazy golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan fluidly upward and right to follow the chickens' movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018175480887293816 + ], + "opposite_logprob_tokens": [ + -4.0181756019592285 + ], + "logprob_reward": -0.018175480887293816, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018175480887293816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15573927760124207, + 0.03408249467611313, + 0.07060904800891876, + 0.03762289509177208, + 0.36361533403396606, + 0.03666901960968971, + 0.04425510764122009, + 0.12476760149002075, + 0.09716066718101501, + 0.035478584468364716 + ], + "credit_map": [ + 0.009844102896749973, + 0.02399536408483982, + 0.01756337471306324, + 0.012606200762093067, + 0.02422611229121685, + 0.023631101474165916, + 0.022214969620108604, + 0.021658048033714294, + 0.0025774084497243166, + 0.0031671796459704638, + 0.0034945784136652946, + 0.009406521916389465, + 0.004313124343752861, + 0.0044282106682658195, + 0.00367765873670578, + 0.003017810173332691, + 0.003383359871804714, + 0.006035888101905584, + 0.006815305445343256, + 0.014003744348883629, + 0.010937312617897987, + 0.009428014978766441, + 0.009070058353245258, + 0.010935366153717041, + 0.0032838392071425915, + 0.0033463858999311924, + 0.005314746871590614, + 0.005135539453476667, + 0.004560764413326979, + 0.004685353022068739, + 0.005744286812841892, + 0.005551980342715979, + 0.02530781924724579, + 0.01583603210747242, + 0.14099855720996857, + 0.039619967341423035, + 0.013069882057607174, + 0.08051330596208572, + 0.01229599118232727, + 0.03597377985715866, + 0.003710698802024126, + 0.007202316075563431, + 0.004059758502990007, + 0.004403818864375353, + 0.00255002873018384, + 0.0058453017845749855, + 0.003270916873589158, + 0.005626183934509754, + 0.00582512142136693, + 0.006991139147430658, + 0.00451320968568325, + 0.005127111915498972, + 0.0047126756981015205, + 0.005836331285536289, + 0.005155053921043873, + 0.006094465032219887, + 0.024810809642076492, + 0.014419248327612877, + 0.024612147361040115, + 0.009138203226029873, + 0.010605880059301853, + 0.008868524804711342, + 0.013781175017356873, + 0.018531616777181625, + 0.02464008703827858, + 0.007264517247676849, + 0.016866914927959442, + 0.008012455888092518, + 0.0063688019290566444, + 0.011974973604083061, + 0.009417104534804821, + 0.0126158082857728, + 0.005104676820337772, + 0.005266362801194191, + 0.005474624689668417, + 0.003836618969216943, + 0.0027305802796036005, + 0.004258412402123213, + 0.0041119870729744434, + 0.0046953242272138596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.018175480887293816 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two chickens visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the chickens display feathers with iridescent greens, blues, and purples mixed with crimson plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the chickens performing leaping and soaring motions while kicking up dust and dry leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-dappled forest clearing with lush green foliage and hazy golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan fluidly upward and right to follow the chickens' movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013316344702616334 + ], + "opposite_logprob_tokens": [ + -6.626331806182861 + ], + "logprob_reward": -0.0013316344702616334, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013316344702616334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17158116400241852, + 0.09082573652267456, + 0.11148769408464432, + 0.10581064224243164, + 0.09741527587175369, + 0.03164228796958923, + 0.0403217077255249, + 0.1128864660859108, + 0.18582916259765625, + 0.052199844270944595 + ], + "credit_map": [ + 0.008848665282130241, + 0.018244067206978798, + 0.02124028280377388, + 0.019493836909532547, + 0.028486493974924088, + 0.023556753993034363, + 0.02546943910419941, + 0.026241634041070938, + 0.006256607826799154, + 0.012162333354353905, + 0.009707678109407425, + 0.014636877924203873, + 0.016316132619976997, + 0.013855533674359322, + 0.008216938003897667, + 0.009673641063272953, + 0.006793460343033075, + 0.01010676845908165, + 0.012171008624136448, + 0.03148414567112923, + 0.012318354099988937, + 0.013846861198544502, + 0.012909467332065105, + 0.011857641860842705, + 0.013946608640253544, + 0.013682587072253227, + 0.016352742910385132, + 0.01682119444012642, + 0.007487798109650612, + 0.014670588076114655, + 0.010169985704123974, + 0.012679144740104675, + 0.0073462678119540215, + 0.004583190660923719, + 0.03565942868590355, + 0.01076912134885788, + 0.0037340489216148853, + 0.022521544247865677, + 0.004006311297416687, + 0.008795369416475296, + 0.0035543644335120916, + 0.004859938286244869, + 0.0058067310601472855, + 0.0032248618081212044, + 0.0027947358321398497, + 0.0032507460564374924, + 0.0034999747294932604, + 0.004650937393307686, + 0.004691904876381159, + 0.0062914458103477955, + 0.0036442065611481667, + 0.004728841129690409, + 0.004023578483611345, + 0.006725763902068138, + 0.004503198899328709, + 0.005712774116545916, + 0.022157618775963783, + 0.012078073807060719, + 0.02176675945520401, + 0.009098727256059647, + 0.007445116527378559, + 0.010833315551280975, + 0.012198581360280514, + 0.017308278009295464, + 0.04522489383816719, + 0.014833001419901848, + 0.03282711282372475, + 0.015614750795066357, + 0.012053309008479118, + 0.023398764431476593, + 0.01755988970398903, + 0.02431745082139969, + 0.006037868559360504, + 0.009604032151401043, + 0.007941676303744316, + 0.0061090304516255856, + 0.005215588957071304, + 0.007011102978140116, + 0.004094713367521763, + 0.006185837090015411 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.0013316344702616334 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two chickens visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the chickens display feathers with iridescent greens, blues, and purples mixed with crimson plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the chickens performing leaping and soaring motions while kicking up dust and dry leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-dappled forest clearing with lush green foliage and hazy golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan fluidly upward and right to follow the chickens' movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06997691094875336 + ], + "opposite_logprob_tokens": [ + -2.694976806640625 + ], + "logprob_reward": -0.06997691094875336, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06997691094875336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08884201943874359, + 0.05574994906783104, + 0.15311196446418762, + 0.051537372171878815, + 0.3607960343360901, + 0.03252951800823212, + 0.03156657889485359, + 0.08739931881427765, + 0.09965554624795914, + 0.03881172463297844 + ], + "credit_map": [ + 0.007574263960123062, + 0.01817288249731064, + 0.008168108761310577, + 0.01117625366896391, + 0.01628762297332287, + 0.013034321367740631, + 0.006641840096563101, + 0.007786730770021677, + 0.004079100210219622, + 0.007355086505413055, + 0.006019165739417076, + 0.009374062530696392, + 0.007688029203563929, + 0.0071183620020747185, + 0.006997762247920036, + 0.007118379697203636, + 0.007697987370193005, + 0.011349817737936974, + 0.0134337292984128, + 0.04245253652334213, + 0.02427728846669197, + 0.019648144021630287, + 0.017190061509609222, + 0.01706239953637123, + 0.006842410191893578, + 0.007168943993747234, + 0.00568352360278368, + 0.00722894724458456, + 0.006448747590184212, + 0.006836384069174528, + 0.005534446332603693, + 0.0057939705438911915, + 0.02713293395936489, + 0.008524383418262005, + 0.14597876369953156, + 0.04065409675240517, + 0.007953055202960968, + 0.09008276462554932, + 0.008329699747264385, + 0.03214031830430031, + 0.0028956991154700518, + 0.004908043425530195, + 0.00513293594121933, + 0.003212766256183386, + 0.002917738864198327, + 0.004470022860914469, + 0.003272365778684616, + 0.0057199448347091675, + 0.0031138325575739145, + 0.005178576800972223, + 0.0031208712607622147, + 0.0036717441398650408, + 0.0037071621045470238, + 0.004421847406774759, + 0.004317105282098055, + 0.004035440273582935, + 0.025771765038371086, + 0.010969643481075764, + 0.01631803624331951, + 0.00389486295171082, + 0.004826773889362812, + 0.005793881602585316, + 0.00600800896063447, + 0.013816342689096928, + 0.02411949820816517, + 0.0077201551757752895, + 0.01709943264722824, + 0.00848936103284359, + 0.006507248617708683, + 0.012299547903239727, + 0.00993609894067049, + 0.01348420325666666, + 0.00482014799490571, + 0.005333106033504009, + 0.006208705250173807, + 0.0047944458201527596, + 0.0030962745659053326, + 0.005166071001440287, + 0.0031317102257162333, + 0.006261263508349657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 8 + ], + "selected_reward": -0.06997691094875336 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Set in a beautiful Ghibli-inspired art style, a low-angle medium shot unfolds across a sun-dappled forest clearing as two magnificent Senju chickens engage in a dramatic duel. Their feathers, a painterly cascade of iridescent greens, blues, and purples, shimmer with every twitch, creating a stark aggression against their proud crimson plumage. They leap and soar with astonishing agility, their powerful legs kicking up swirling clouds of fine dust and dry leaves, capturing the soft, warm light filtering through the dense forest canopy. As one of the combatants crouches, then explosively soars into the air for a dramatic attack, the camera pans fluidly upward and right, following its graceful arc. Against a backdrop of lush green foliage and hazy golden sunlight, the action captures the fluid movement of its outstretched wings and extended claws, creating a vibrant and captivating atmosphere that is both tense and beautiful.", + "seed": 12635, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_15/index_8954_seed_12635_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03169603426940739, + "vlm_logprob": -0.03169603426940739, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.03169603426940739, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two chickens visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the chickens display feathers with iridescent greens, blues, and purples mixed with crimson plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the chickens performing leaping and soaring motions while kicking up dust and dry leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-dappled forest clearing with lush green foliage and hazy golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan fluidly upward and right to follow the chickens' movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08066009730100632 + ], + "opposite_logprob_tokens": [ + -2.580660104751587 + ], + "logprob_reward": -0.08066009730100632, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08066009730100632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14859282970428467, + 0.0832536369562149, + 0.17741595208644867, + 0.07209053635597229, + 0.1469258964061737, + 0.15454457700252533, + 0.04123791679739952, + 0.040240488946437836, + 0.06816187500953674, + 0.06753630936145782 + ], + "credit_map": [ + 0.014395184814929962, + 0.026103710755705833, + 0.02913310006260872, + 0.012359937652945518, + 0.01805642805993557, + 0.027899449691176414, + 0.012319903820753098, + 0.008325108326971531, + 0.009049972519278526, + 0.012521733529865742, + 0.012560255825519562, + 0.007712204474955797, + 0.010889959521591663, + 0.007388101425021887, + 0.013149090111255646, + 0.009982322342693806, + 0.010403341613709927, + 0.023253822699189186, + 0.019510382786393166, + 0.018801206722855568, + 0.027351543307304382, + 0.026357946917414665, + 0.034591615200042725, + 0.01714608445763588, + 0.006527990102767944, + 0.012925022281706333, + 0.010708166286349297, + 0.007395962253212929, + 0.01174100860953331, + 0.00741208903491497, + 0.006736749317497015, + 0.008643544279038906, + 0.014469780027866364, + 0.016754312440752983, + 0.022989947348833084, + 0.027185432612895966, + 0.017767639830708504, + 0.01493312418460846, + 0.020147638395428658, + 0.012678032740950584, + 0.009426997043192387, + 0.013631961308419704, + 0.023145953193306923, + 0.027781110256910324, + 0.01673036441206932, + 0.023362750187516212, + 0.023320728912949562, + 0.017144721001386642, + 0.0032491490710526705, + 0.005125524941831827, + 0.006130429450422525, + 0.006633580196648836, + 0.006025382317602634, + 0.00422727782279253, + 0.00587335554882884, + 0.003973218612372875, + 0.004706155974417925, + 0.004334636032581329, + 0.005885187536478043, + 0.005123713985085487, + 0.00577673502266407, + 0.007921937853097916, + 0.0034335001837462187, + 0.0030586211942136288, + 0.004733857698738575, + 0.009744818322360516, + 0.011184990406036377, + 0.009167131967842579, + 0.008678790181875229, + 0.008419139310717583, + 0.008776089176535606, + 0.007457063067704439, + 0.005790082272142172, + 0.005604834761470556, + 0.007319977972656488, + 0.0119701586663723, + 0.005258271936327219, + 0.008258404210209846, + 0.009915228933095932, + 0.013419345952570438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.08066009730100632 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two chickens visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the chickens display feathers with iridescent greens, blues, and purples mixed with crimson plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the chickens performing leaping and soaring motions while kicking up dust and dry leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-dappled forest clearing with lush green foliage and hazy golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan fluidly upward and right to follow the chickens' movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012557132169604301 + ], + "opposite_logprob_tokens": [ + -4.387557029724121 + ], + "logprob_reward": -0.012557132169604301, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012557132169604301, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11354456841945648, + 0.08508102595806122, + 0.08681896328926086, + 0.03409229591488838, + 0.12326089292764664, + 0.10773206502199173, + 0.16300536692142487, + 0.08314787596464157, + 0.15428540110588074, + 0.04903160408139229 + ], + "credit_map": [ + 0.011275744065642357, + 0.016688724979758263, + 0.020253702998161316, + 0.010794196277856827, + 0.014308251440525055, + 0.01911817491054535, + 0.011935877613723278, + 0.009169886820018291, + 0.009737242013216019, + 0.022094357758760452, + 0.008781509473919868, + 0.010626775212585926, + 0.010050376877188683, + 0.00552341528236866, + 0.010633152909576893, + 0.007634185254573822, + 0.007052547764033079, + 0.01208092924207449, + 0.010484152473509312, + 0.009097877889871597, + 0.013690538704395294, + 0.012220937758684158, + 0.01304344180971384, + 0.009148536249995232, + 0.003167020855471492, + 0.007624675985425711, + 0.0042203133925795555, + 0.0039687068201601505, + 0.004673661198467016, + 0.004131615627557039, + 0.0030869075562804937, + 0.003219394013285637, + 0.01612210087478161, + 0.012666508555412292, + 0.019145745784044266, + 0.02308526076376438, + 0.015499381348490715, + 0.012607413344085217, + 0.014415989629924297, + 0.009718481451272964, + 0.009241658262908459, + 0.008260298520326614, + 0.017919359728693962, + 0.015892598778009415, + 0.012737229466438293, + 0.012022409588098526, + 0.023096082732081413, + 0.008562419563531876, + 0.013928025960922241, + 0.016884060576558113, + 0.035762399435043335, + 0.042987801134586334, + 0.017884865403175354, + 0.00849222857505083, + 0.01820664294064045, + 0.008859326131641865, + 0.007047266233712435, + 0.011586552485823631, + 0.012218748219311237, + 0.012888397090137005, + 0.009087767452001572, + 0.011554133147001266, + 0.009974964894354343, + 0.008790040388703346, + 0.01692703738808632, + 0.025056807324290276, + 0.021941766142845154, + 0.015762731432914734, + 0.020053984597325325, + 0.018695326521992683, + 0.021549265831708908, + 0.014298460446298122, + 0.005600511562079191, + 0.006450584623962641, + 0.007058968301862478, + 0.005228708498179913, + 0.004576167557388544, + 0.00577659672126174, + 0.006966328714042902, + 0.0073737334460020065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 4 + ], + "selected_reward": -0.012557132169604301 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two chickens visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the chickens display feathers with iridescent greens, blues, and purples mixed with crimson plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the chickens performing leaping and soaring motions while kicking up dust and dry leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-dappled forest clearing with lush green foliage and hazy golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan fluidly upward and right to follow the chickens' movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008631540462374687 + ], + "opposite_logprob_tokens": [ + -4.758631706237793 + ], + "logprob_reward": -0.008631540462374687, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008631540462374687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1626439094543457, + 0.10325894504785538, + 0.09351303428411484, + 0.10718502104282379, + 0.13919247686862946, + 0.06920155137777328, + 0.05926137417554855, + 0.1603754162788391, + 0.050385795533657074, + 0.05498252436518669 + ], + "credit_map": [ + 0.014229372143745422, + 0.013882521539926529, + 0.03579620644450188, + 0.016384027898311615, + 0.021025000140070915, + 0.03392641991376877, + 0.013981489464640617, + 0.01341887004673481, + 0.014604039490222931, + 0.01630496419966221, + 0.014353756792843342, + 0.010578162968158722, + 0.014704804867506027, + 0.008769765496253967, + 0.013633383437991142, + 0.010310067795217037, + 0.007209753151983023, + 0.013277554884552956, + 0.011232469230890274, + 0.010580132715404034, + 0.014324490912258625, + 0.012426321394741535, + 0.013335928320884705, + 0.011126389726996422, + 0.010630689561367035, + 0.015600129961967468, + 0.018278200179338455, + 0.015248097479343414, + 0.013092631474137306, + 0.01163685042411089, + 0.012532921507954597, + 0.010165495797991753, + 0.015877358615398407, + 0.029212825000286102, + 0.01914672553539276, + 0.018815411254763603, + 0.02062910795211792, + 0.01365793589502573, + 0.011983366683125496, + 0.009869751520454884, + 0.008197860792279243, + 0.008394096978008747, + 0.00824046041816473, + 0.007478086743503809, + 0.009046577848494053, + 0.00813050102442503, + 0.010104083456099033, + 0.009609891101717949, + 0.00734207546338439, + 0.007217923179268837, + 0.00971202552318573, + 0.012332048267126083, + 0.007358100730925798, + 0.003695515915751457, + 0.006888757459819317, + 0.004714928567409515, + 0.01629207469522953, + 0.021392887458205223, + 0.02176864631474018, + 0.019732249900698662, + 0.020037855952978134, + 0.03648252785205841, + 0.010997645556926727, + 0.013671528548002243, + 0.003931599669158459, + 0.006238420028239489, + 0.006814403925091028, + 0.005806682165712118, + 0.006693016737699509, + 0.008059613406658173, + 0.007699964568018913, + 0.005142098292708397, + 0.005231078714132309, + 0.005182809196412563, + 0.007396874949336052, + 0.012284867465496063, + 0.003918429836630821, + 0.005602186545729637, + 0.008256186731159687, + 0.0071100895293056965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.008631540462374687 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two chickens visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the chickens display feathers with iridescent greens, blues, and purples mixed with crimson plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the chickens performing leaping and soaring motions while kicking up dust and dry leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-dappled forest clearing with lush green foliage and hazy golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan fluidly upward and right to follow the chickens' movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001708358759060502 + ], + "opposite_logprob_tokens": [ + -6.376708507537842 + ], + "logprob_reward": -0.001708358759060502, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001708358759060502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12155521661043167, + 0.05980776622891426, + 0.08285801857709885, + 0.05455463007092476, + 0.26006442308425903, + 0.13329312205314636, + 0.09557854384183884, + 0.07689830660820007, + 0.03703523799777031, + 0.07835476845502853 + ], + "credit_map": [ + 0.015984924510121346, + 0.0191523227840662, + 0.021499788388609886, + 0.0099300192669034, + 0.014513357542455196, + 0.02057621255517006, + 0.007767823524773121, + 0.012130767107009888, + 0.013847917318344116, + 0.00672499556094408, + 0.010199373587965965, + 0.0046618361957371235, + 0.007134489249438047, + 0.0064102341420948505, + 0.004412406589835882, + 0.006416514050215483, + 0.007200457155704498, + 0.010143780149519444, + 0.012490496039390564, + 0.00498427776619792, + 0.0124840522184968, + 0.018681352958083153, + 0.007032035384327173, + 0.009841559454798698, + 0.005449885968118906, + 0.007552634458988905, + 0.008762951008975506, + 0.006254794076085091, + 0.00951629038900137, + 0.006274393759667873, + 0.0054417806677520275, + 0.005301894154399633, + 0.024664977565407753, + 0.026812708005309105, + 0.042410194873809814, + 0.04782763496041298, + 0.0359862744808197, + 0.028254173696041107, + 0.032018546015024185, + 0.022089878097176552, + 0.016984710469841957, + 0.017292173579335213, + 0.023538943380117416, + 0.018596656620502472, + 0.012647216208279133, + 0.015604381449520588, + 0.013141618110239506, + 0.01548741851001978, + 0.01244804635643959, + 0.009328977204859257, + 0.0168911162763834, + 0.016817498952150345, + 0.012371508404612541, + 0.010194544680416584, + 0.009759923443198204, + 0.007766926661133766, + 0.007303535006940365, + 0.012190969660878181, + 0.010826141573488712, + 0.011042993515729904, + 0.008126535452902317, + 0.010812257416546345, + 0.008352185599505901, + 0.008243681862950325, + 0.0038327088113874197, + 0.004916743841022253, + 0.004936061333864927, + 0.00513850012794137, + 0.005808243528008461, + 0.004885286558419466, + 0.003838136326521635, + 0.0036795560736209154, + 0.011633699759840965, + 0.007367792539298534, + 0.012704998254776001, + 0.011459436267614365, + 0.006423449143767357, + 0.008635958656668663, + 0.010991017334163189, + 0.009138419292867184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.001708358759060502 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two chickens visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the chickens display feathers with iridescent greens, blues, and purples mixed with crimson plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the chickens performing leaping and soaring motions while kicking up dust and dry leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-dappled forest clearing with lush green foliage and hazy golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan fluidly upward and right to follow the chickens' movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05492304265499115 + ], + "opposite_logprob_tokens": [ + -2.9299230575561523 + ], + "logprob_reward": -0.05492304265499115, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05492304265499115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10497961938381195, + 0.1119128167629242, + 0.050019752234220505, + 0.037411391735076904, + 0.12706269323825836, + 0.11037041991949081, + 0.12191479653120041, + 0.1783577799797058, + 0.09780333936214447, + 0.06016741693019867 + ], + "credit_map": [ + 0.010883287526667118, + 0.011450057849287987, + 0.024998506531119347, + 0.009356217458844185, + 0.01057609636336565, + 0.023957816883921623, + 0.00680092116817832, + 0.00695671234279871, + 0.014690850861370564, + 0.013275973498821259, + 0.01711958833038807, + 0.014868250116705894, + 0.011857674457132816, + 0.008125619031488895, + 0.01589387096464634, + 0.016080981120467186, + 0.004419779404997826, + 0.008013434708118439, + 0.006599795538932085, + 0.0043134912848472595, + 0.0075281704775989056, + 0.008508766070008278, + 0.004563345108181238, + 0.0060729640536010265, + 0.0038107531145215034, + 0.00797631498426199, + 0.004258110653609037, + 0.004515974782407284, + 0.0041969092562794685, + 0.0037564088124781847, + 0.004856233485043049, + 0.004040681757032871, + 0.012822674587368965, + 0.017554188147187233, + 0.01993132010102272, + 0.023880986496806145, + 0.018463896587491035, + 0.01233639009296894, + 0.01259101927280426, + 0.009482205845415592, + 0.010239826515316963, + 0.012712480500340462, + 0.020474201068282127, + 0.013468252494931221, + 0.013631943613290787, + 0.010553400963544846, + 0.017448995262384415, + 0.011841312050819397, + 0.014047964476048946, + 0.012440656311810017, + 0.02499944530427456, + 0.026632534340023994, + 0.016190793365240097, + 0.007599339820444584, + 0.012795154005289078, + 0.007208900526165962, + 0.017535237595438957, + 0.0201824102550745, + 0.0317494198679924, + 0.02568761073052883, + 0.021198751404881477, + 0.016192641109228134, + 0.02123364992439747, + 0.02457805909216404, + 0.007722063455730677, + 0.016507158055901527, + 0.012501939199864864, + 0.00984177365899086, + 0.012641950510442257, + 0.0099760377779603, + 0.01762506179511547, + 0.010987351648509502, + 0.007561213336884975, + 0.007157725282013416, + 0.007334382273256779, + 0.007454054430127144, + 0.004883154761046171, + 0.006169293075799942, + 0.007640084717422724, + 0.011967500671744347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 6 + ], + "selected_reward": -0.05492304265499115 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Set in a beautiful Ghibli-inspired art style, a low-angle medium shot unfolds across a sun-dappled forest clearing as two magnificent Senju chickens engage in a dramatic duel. Their feathers, a painterly cascade of iridescent greens, blues, and purples, shimmer with every twitch, creating a stark aggression against their proud crimson plumage. They leap and soar with astonishing agility, their powerful legs kicking up swirling clouds of fine dust and dry leaves, capturing the soft, warm light filtering through the dense forest canopy. As one of the combatants crouches, then explosively soars into the air for a dramatic attack, the camera pans fluidly upward and right, following its graceful arc. Against a backdrop of lush green foliage and hazy golden sunlight, the action captures the fluid movement of its outstretched wings and extended claws, creating a vibrant and captivating atmosphere that is both tense and beautiful.", + "seed": 12636, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_15/index_8954_seed_12636_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.017643930856138468, + "vlm_logprob": -0.017643930856138468, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.017643930856138468, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two chickens visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the chickens display feathers with iridescent greens, blues, and purples mixed with crimson plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the chickens performing leaping and soaring motions while kicking up dust and dry leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-dappled forest clearing with lush green foliage and hazy golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan fluidly upward and right to follow the chickens' movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009817530401051044 + ], + "opposite_logprob_tokens": [ + -4.759817600250244 + ], + "logprob_reward": -0.009817530401051044, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.009817530401051044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08328047394752502, + 0.14316263794898987, + 0.07865999639034271, + 0.2930563986301422, + 0.060267895460128784, + 0.040205858647823334, + 0.08054469525814056, + 0.10771967470645905, + 0.05760576203465462, + 0.05549658089876175 + ], + "credit_map": [ + 0.006804573815315962, + 0.012306789867579937, + 0.01084953173995018, + 0.016655059531331062, + 0.010364055633544922, + 0.010107899084687233, + 0.008331581950187683, + 0.007860979065299034, + 0.01446838490664959, + 0.009620200842618942, + 0.027104729786515236, + 0.022187232971191406, + 0.01188400387763977, + 0.010769994929432869, + 0.031246589496731758, + 0.015881508588790894, + 0.008089633658528328, + 0.010449085384607315, + 0.010019710287451744, + 0.011151356622576714, + 0.015826009213924408, + 0.008211340755224228, + 0.0075308033265173435, + 0.007382057141512632, + 0.018542995676398277, + 0.01136142946779728, + 0.1040586531162262, + 0.11230749636888504, + 0.007644450757652521, + 0.008558541536331177, + 0.013933288864791393, + 0.016649562865495682, + 0.00619472237303853, + 0.004621513653546572, + 0.014843027107417583, + 0.0066756680607795715, + 0.007823720574378967, + 0.00940592959523201, + 0.004171739332377911, + 0.006531578954309225, + 0.0029733916744589806, + 0.003442958928644657, + 0.0032382705248892307, + 0.005761187057942152, + 0.004691772628575563, + 0.007115645334124565, + 0.004870886914432049, + 0.008111748844385147, + 0.006670758593827486, + 0.005884423851966858, + 0.005909809377044439, + 0.007269649766385555, + 0.0054933144710958, + 0.02432389184832573, + 0.004614970646798611, + 0.0203778687864542, + 0.009565407410264015, + 0.012121899984776974, + 0.009227558970451355, + 0.01052191574126482, + 0.009088597260415554, + 0.015897277742624283, + 0.02197689190506935, + 0.01932012289762497, + 0.007061778102070093, + 0.006589109543710947, + 0.005064176395535469, + 0.0062920572236180305, + 0.013441150076687336, + 0.007218281272798777, + 0.007392258383333683, + 0.00454695476219058, + 0.005221369676291943, + 0.006035267375409603, + 0.005708742421120405, + 0.00633586198091507, + 0.007837842218577862, + 0.005558324512094259, + 0.0083885807543993, + 0.010410591028630733 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 7 + ], + "selected_reward": -0.009817530401051044 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two chickens visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the chickens display feathers with iridescent greens, blues, and purples mixed with crimson plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the chickens performing leaping and soaring motions while kicking up dust and dry leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-dappled forest clearing with lush green foliage and hazy golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan fluidly upward and right to follow the chickens' movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059604281559586525 + ], + "opposite_logprob_tokens": [ + -5.130960464477539 + ], + "logprob_reward": -0.0059604281559586525, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059604281559586525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12258098274469376, + 0.09241300821304321, + 0.13571541011333466, + 0.2608788013458252, + 0.024720579385757446, + 0.031374234706163406, + 0.07029763609170914, + 0.08630955964326859, + 0.12154101580381393, + 0.054168831557035446 + ], + "credit_map": [ + 0.011787190102040768, + 0.01887110434472561, + 0.01499902456998825, + 0.017636749893426895, + 0.019366269931197166, + 0.015233603306114674, + 0.011803007684648037, + 0.012884017080068588, + 0.00927495677024126, + 0.00804444495588541, + 0.016714010387659073, + 0.01401960663497448, + 0.006902182009071112, + 0.009117117151618004, + 0.016674397513270378, + 0.011666289530694485, + 0.017083853483200073, + 0.012388254515826702, + 0.023922163993120193, + 0.012439501471817493, + 0.026130786165595055, + 0.019531194120645523, + 0.013987855985760689, + 0.010231790132820606, + 0.016367681324481964, + 0.0086802514269948, + 0.09321869164705276, + 0.10040517151355743, + 0.00586526095867157, + 0.007526226807385683, + 0.013568327762186527, + 0.01524715218693018, + 0.0028453234117478132, + 0.0021553465630859137, + 0.004724034108221531, + 0.0030506569892168045, + 0.0021636346355080605, + 0.0031498649623245, + 0.00277117732912302, + 0.0038605383597314358, + 0.002584208967164159, + 0.0025877405423671007, + 0.0025646660942584276, + 0.003039236878976226, + 0.002450754400342703, + 0.008202029392123222, + 0.002363999607041478, + 0.0075815957970917225, + 0.008097211830317974, + 0.006612114142626524, + 0.007203100249171257, + 0.009801851585507393, + 0.00750234629958868, + 0.015541448257863522, + 0.005895583890378475, + 0.00964396819472313, + 0.004923524800688028, + 0.004581778775900602, + 0.008774016983807087, + 0.006786647718399763, + 0.008506101556122303, + 0.007674792315810919, + 0.025677617639303207, + 0.019385067746043205, + 0.013023060746490955, + 0.00883530080318451, + 0.015947960317134857, + 0.015053892508149147, + 0.013480783440172672, + 0.02212505228817463, + 0.02042972296476364, + 0.012645230628550053, + 0.0044310870580375195, + 0.006145409308373928, + 0.005591562483459711, + 0.009839583188295364, + 0.005576587747782469, + 0.006812356878072023, + 0.00892008375376463, + 0.006852156948298216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.0059604281559586525 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two chickens visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the chickens display feathers with iridescent greens, blues, and purples mixed with crimson plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the chickens performing leaping and soaring motions while kicking up dust and dry leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-dappled forest clearing with lush green foliage and hazy golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan fluidly upward and right to follow the chickens' movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007621141150593758 + ], + "opposite_logprob_tokens": [ + -4.8826212882995605 + ], + "logprob_reward": -0.007621141150593758, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007621141150593758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16921724379062653, + 0.10106644779443741, + 0.03535062074661255, + 0.1346830427646637, + 0.06417600065469742, + 0.04117361828684807, + 0.06100061908364296, + 0.13634975254535675, + 0.14507244527339935, + 0.11191019415855408 + ], + "credit_map": [ + 0.011660361662507057, + 0.01719597354531288, + 0.03107593208551407, + 0.022179745137691498, + 0.02420632168650627, + 0.024102114140987396, + 0.015142702497541904, + 0.023654095828533173, + 0.009368019178509712, + 0.010134943760931492, + 0.014910670928657055, + 0.012916449457406998, + 0.011964719742536545, + 0.01131884753704071, + 0.019854864105582237, + 0.01059793122112751, + 0.003309668740257621, + 0.0041105239652097225, + 0.00587843544781208, + 0.005719107110053301, + 0.004261918365955353, + 0.004376557655632496, + 0.004413176793605089, + 0.00328123290091753, + 0.010899231769144535, + 0.01043360773473978, + 0.028325721621513367, + 0.0360158309340477, + 0.014047203585505486, + 0.011096624657511711, + 0.014416512101888657, + 0.009448305703699589, + 0.005879663396626711, + 0.0057995934039354324, + 0.018629062920808792, + 0.006769775412976742, + 0.007155959960073233, + 0.006070917472243309, + 0.007238838355988264, + 0.006632195319980383, + 0.002273973310366273, + 0.003551698988303542, + 0.0037346442695707083, + 0.003931541461497545, + 0.004451720044016838, + 0.01047839131206274, + 0.0044826907105743885, + 0.008268958888947964, + 0.0055250669829547405, + 0.005350899882614613, + 0.006896196864545345, + 0.010312308557331562, + 0.005650591570883989, + 0.011201390065252781, + 0.006633705925196409, + 0.009430458769202232, + 0.010642735287547112, + 0.013524416834115982, + 0.012054496444761753, + 0.014154302887618542, + 0.010239139199256897, + 0.01987760327756405, + 0.029112158343195915, + 0.026744907721877098, + 0.017667043954133987, + 0.012118300423026085, + 0.01597999967634678, + 0.03056824952363968, + 0.014642294496297836, + 0.011404431425035, + 0.02338644675910473, + 0.019305691123008728, + 0.014688967727124691, + 0.00951271690428257, + 0.020130053162574768, + 0.015127803198993206, + 0.010242597199976444, + 0.009589439257979393, + 0.011226505041122437, + 0.02139211818575859 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.007621141150593758 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two chickens visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the chickens display feathers with iridescent greens, blues, and purples mixed with crimson plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the chickens performing leaping and soaring motions while kicking up dust and dry leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-dappled forest clearing with lush green foliage and hazy golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan fluidly upward and right to follow the chickens' movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028127175755798817 + ], + "opposite_logprob_tokens": [ + -5.87781286239624 + ], + "logprob_reward": -0.0028127175755798817, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028127175755798817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11514081060886383, + 0.051943402737379074, + 0.017435749992728233, + 0.16048571467399597, + 0.023498302325606346, + 0.03537219762802124, + 0.06811431050300598, + 0.06358956545591354, + 0.35347306728363037, + 0.11094685643911362 + ], + "credit_map": [ + 0.008062214590609074, + 0.019171206280589104, + 0.014379351399838924, + 0.014946607872843742, + 0.021917056292295456, + 0.014137358404695988, + 0.009585381485521793, + 0.012941633351147175, + 0.005409739445894957, + 0.006726815365254879, + 0.007242344785481691, + 0.006605531554669142, + 0.004999780561774969, + 0.007214633747935295, + 0.007607151288539171, + 0.006137405056506395, + 0.0020747443195432425, + 0.0018502543680369854, + 0.0024418551474809647, + 0.002180727431550622, + 0.0030396191868931055, + 0.0016572894528508186, + 0.002173246582970023, + 0.002018013270571828, + 0.009634933434426785, + 0.005849756766110659, + 0.05588633194565773, + 0.0629328042268753, + 0.004290787503123283, + 0.004785132594406605, + 0.007459726184606552, + 0.009646235965192318, + 0.0034527021925896406, + 0.0024036720860749483, + 0.0031597381457686424, + 0.0034300461411476135, + 0.002385453786700964, + 0.0031790537759661674, + 0.002786036115139723, + 0.0027016010135412216, + 0.0020882769022136927, + 0.002305376809090376, + 0.0031456612050533295, + 0.002989088883623481, + 0.0031123445369303226, + 0.00915564689785242, + 0.0032501553650945425, + 0.009325649589300156, + 0.004604116547852755, + 0.002703930251300335, + 0.005265612620860338, + 0.003650007303804159, + 0.004845710471272469, + 0.02288343943655491, + 0.0036504962481558323, + 0.020511001348495483, + 0.005523182917386293, + 0.007911969907581806, + 0.006324035115540028, + 0.007840309292078018, + 0.005367489065974951, + 0.011291882954537868, + 0.01125193852931261, + 0.008078756742179394, + 0.03135077282786369, + 0.02435261569917202, + 0.09139899164438248, + 0.03971344977617264, + 0.03051074594259262, + 0.038621362298727036, + 0.04434284195303917, + 0.053182266652584076, + 0.013669698499143124, + 0.009524691849946976, + 0.027180705219507217, + 0.014993221499025822, + 0.00878423172980547, + 0.008023673668503761, + 0.010571378283202648, + 0.01819925755262375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.0028127175755798817 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two chickens visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the chickens display feathers with iridescent greens, blues, and purples mixed with crimson plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the chickens performing leaping and soaring motions while kicking up dust and dry leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-dappled forest clearing with lush green foliage and hazy golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan fluidly upward and right to follow the chickens' movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062007836997509 + ], + "opposite_logprob_tokens": [ + -2.8120079040527344 + ], + "logprob_reward": -0.062007836997509, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062007836997509, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054273106157779694, + 0.04236144572496414, + 0.037338659167289734, + 0.5440296530723572, + 0.04222538694739342, + 0.024485602974891663, + 0.04494352638721466, + 0.0438794381916523, + 0.10525653511285782, + 0.061206646263599396 + ], + "credit_map": [ + 0.0038013129960745573, + 0.006136934272944927, + 0.009716139174997807, + 0.007062151562422514, + 0.00878194160759449, + 0.006930687464773655, + 0.005084343254566193, + 0.006759593263268471, + 0.004279167857021093, + 0.003287440398707986, + 0.008416038937866688, + 0.00587735278531909, + 0.003584302728995681, + 0.0039594052359461784, + 0.007084115408360958, + 0.005873620510101318, + 0.0034116103779524565, + 0.003737878752872348, + 0.008237773552536964, + 0.006279745604842901, + 0.003925515338778496, + 0.003006178420037031, + 0.00415587704628706, + 0.004584084264934063, + 0.036335770040750504, + 0.017419179901480675, + 0.20400018990039825, + 0.20816458761692047, + 0.01161556039005518, + 0.016043785959482193, + 0.020901331678032875, + 0.029549280181527138, + 0.004975283518433571, + 0.002739272080361843, + 0.015411061234772205, + 0.00527295982465148, + 0.0028752414509654045, + 0.004049831070005894, + 0.0033051464706659317, + 0.0035965910647064447, + 0.0014821331715211272, + 0.0018148041563108563, + 0.002055661752820015, + 0.0025899934116750956, + 0.002140131313353777, + 0.00616892334073782, + 0.0022902744822204113, + 0.005943681579083204, + 0.0036095604300498962, + 0.0020508458837866783, + 0.004011454060673714, + 0.0033241803757846355, + 0.00386639847420156, + 0.012875551357865334, + 0.003810603404417634, + 0.011394931003451347, + 0.00339974625967443, + 0.004947933834046125, + 0.004301202949136496, + 0.004766433499753475, + 0.0036775374319404364, + 0.004898515064269304, + 0.008617267943918705, + 0.009270802140235901, + 0.011178719811141491, + 0.007483727298676968, + 0.01901310868561268, + 0.013302600011229515, + 0.011664948426187038, + 0.017165370285511017, + 0.014300907962024212, + 0.011147145181894302, + 0.005145899020135403, + 0.004912128672003746, + 0.017191627994179726, + 0.008675393648445606, + 0.006104223895817995, + 0.003617333946749568, + 0.006930295377969742, + 0.00862974114716053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 9 + ], + "selected_reward": -0.062007836997509 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Set in a beautiful Ghibli-inspired art style, a low-angle medium shot unfolds across a sun-dappled forest clearing as two magnificent Senju chickens engage in a dramatic duel. Their feathers, a painterly cascade of iridescent greens, blues, and purples, shimmer with every twitch, creating a stark aggression against their proud crimson plumage. They leap and soar with astonishing agility, their powerful legs kicking up swirling clouds of fine dust and dry leaves, capturing the soft, warm light filtering through the dense forest canopy. As one of the combatants crouches, then explosively soars into the air for a dramatic attack, the camera pans fluidly upward and right, following its graceful arc. Against a backdrop of lush green foliage and hazy golden sunlight, the action captures the fluid movement of its outstretched wings and extended claws, creating a vibrant and captivating atmosphere that is both tense and beautiful.", + "seed": 12637, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_15/index_8954_seed_12637_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06415535006672143, + "vlm_logprob": -0.06415535006672143, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.06415535006672143, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two chickens visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the chickens display feathers with iridescent greens, blues, and purples mixed with crimson plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the chickens performing leaping and soaring motions while kicking up dust and dry leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-dappled forest clearing with lush green foliage and hazy golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan fluidly upward and right to follow the chickens' movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20261801779270172 + ], + "opposite_logprob_tokens": [ + -1.7026180028915405 + ], + "logprob_reward": -0.20261801779270172, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20261801779270172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5365328192710876, + 0.014165602624416351, + 0.04110173508524895, + 0.04308808967471123, + 0.012843250297009945, + 0.019911078736186028, + 0.13899701833724976, + 0.03665700927376747, + 0.14005912840366364, + 0.016644343733787537 + ], + "credit_map": [ + 0.21727444231510162, + 0.015618420206010342, + 0.03223607689142227, + 0.012506750412285328, + 0.019075363874435425, + 0.21350306272506714, + 0.009817373007535934, + 0.016501273959875107, + 0.0014116493985056877, + 0.0010428414680063725, + 0.00266400002874434, + 0.0023016654886305332, + 0.001300552859902382, + 0.0017932644113898277, + 0.001423361012712121, + 0.0022282677236944437, + 0.0046035489067435265, + 0.003901202930137515, + 0.006998736876994371, + 0.0066647701896727085, + 0.0051759639754891396, + 0.005116750951856375, + 0.004340429324656725, + 0.0043003251776099205, + 0.002310926793143153, + 0.0023188276682049036, + 0.004716400057077408, + 0.004452264867722988, + 0.0026202809531241655, + 0.0026975381188094616, + 0.014180309139192104, + 0.009791539050638676, + 0.001052842941135168, + 0.0015910342335700989, + 0.0017915032804012299, + 0.002250164281576872, + 0.0014659560984000564, + 0.0012861988507211208, + 0.0017841918161138892, + 0.001621357980184257, + 0.002943615196272731, + 0.00334198703058064, + 0.0015717045171186328, + 0.0024829849135130644, + 0.0024753790348768234, + 0.00233027758076787, + 0.002505105221644044, + 0.002260022796690464, + 0.013827416114509106, + 0.010794606059789658, + 0.016584429889917374, + 0.020428212359547615, + 0.024111399427056313, + 0.013731184415519238, + 0.018797043710947037, + 0.02072271890938282, + 0.005370128899812698, + 0.004130915272980928, + 0.0026692256797105074, + 0.003214845899492502, + 0.007378618698567152, + 0.004819069989025593, + 0.006022355519235134, + 0.0030518467538058758, + 0.01484651118516922, + 0.0027444828301668167, + 0.005246020387858152, + 0.006506417877972126, + 0.004782038740813732, + 0.04512815177440643, + 0.013458957895636559, + 0.04734652861952782, + 0.0022265359293669462, + 0.0019856207072734833, + 0.0035337815061211586, + 0.0026673986576497555, + 0.0014068292221054435, + 0.0011307954555377364, + 0.0020720697939395905, + 0.001621311530470848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.20261801779270172 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two chickens visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the chickens display feathers with iridescent greens, blues, and purples mixed with crimson plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the chickens performing leaping and soaring motions while kicking up dust and dry leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-dappled forest clearing with lush green foliage and hazy golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan fluidly upward and right to follow the chickens' movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005270277615636587 + ], + "opposite_logprob_tokens": [ + -5.255270481109619 + ], + "logprob_reward": -0.005270277615636587, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005270277615636587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25624072551727295, + 0.11868784576654434, + 0.0761789008975029, + 0.04646316543221474, + 0.026861529797315598, + 0.021343190222978592, + 0.09141905605792999, + 0.05876586586236954, + 0.2559404671192169, + 0.04809918999671936 + ], + "credit_map": [ + 0.09412311762571335, + 0.009610840119421482, + 0.01551064196974039, + 0.011297158896923065, + 0.011971351690590382, + 0.09293641895055771, + 0.009671319276094437, + 0.011119899339973927, + 0.015419474802911282, + 0.010798422619700432, + 0.014072667807340622, + 0.012949857860803604, + 0.013546005822718143, + 0.017502913251519203, + 0.009277978911995888, + 0.02512052096426487, + 0.007416107226163149, + 0.006369130685925484, + 0.013229911215603352, + 0.01729920320212841, + 0.007711596321314573, + 0.008672598749399185, + 0.00718065956607461, + 0.008299693465232849, + 0.002206334611400962, + 0.0034345060121268034, + 0.007037288509309292, + 0.0041024694219231606, + 0.0034044047351926565, + 0.004029292147606611, + 0.015395594760775566, + 0.006853276863694191, + 0.0023974296636879444, + 0.002140581840649247, + 0.004549272358417511, + 0.0021558094304054976, + 0.002612414536997676, + 0.005164609756320715, + 0.002433341694995761, + 0.005408068653196096, + 0.003151063108816743, + 0.003747721202671528, + 0.0018310839077457786, + 0.0023395298048853874, + 0.0023573911748826504, + 0.0021671168506145477, + 0.003175212536007166, + 0.0025740715209394693, + 0.007937104441225529, + 0.004640213679522276, + 0.010307824239134789, + 0.013671192340552807, + 0.015758121386170387, + 0.009159821085631847, + 0.017678389325737953, + 0.012266390956938267, + 0.013971804641187191, + 0.005040284711867571, + 0.004543527029454708, + 0.004684388637542725, + 0.0059832320548594, + 0.014021027833223343, + 0.0067972601391375065, + 0.0037243380211293697, + 0.025676146149635315, + 0.0064029092900455, + 0.012684788554906845, + 0.012235214933753014, + 0.011397810652852058, + 0.08074487745761871, + 0.022927239537239075, + 0.08387147635221481, + 0.0063674794510006905, + 0.004298435058444738, + 0.009541318751871586, + 0.009681847877800465, + 0.003756301011890173, + 0.0036197688896209, + 0.00554504431784153, + 0.005288995336741209 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.005270277615636587 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two chickens visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the chickens display feathers with iridescent greens, blues, and purples mixed with crimson plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the chickens performing leaping and soaring motions while kicking up dust and dry leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-dappled forest clearing with lush green foliage and hazy golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan fluidly upward and right to follow the chickens' movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026329822838306427 + ], + "opposite_logprob_tokens": [ + -3.651329755783081 + ], + "logprob_reward": -0.026329822838306427, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026329822838306427, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5625802278518677, + 0.13086169958114624, + 0.018286189064383507, + 0.05347947031259537, + 0.007070052437484264, + 0.01714611053466797, + 0.03826180472970009, + 0.01889851503074169, + 0.12214680761098862, + 0.03126910701394081 + ], + "credit_map": [ + 0.2267880141735077, + 0.01618940755724907, + 0.03375080227851868, + 0.014517875388264656, + 0.019886663183569908, + 0.22331111133098602, + 0.010130008682608604, + 0.018006332218647003, + 0.01601361483335495, + 0.01072049979120493, + 0.014819877222180367, + 0.014799579977989197, + 0.015986520797014236, + 0.019060086458921432, + 0.010367951355874538, + 0.029093563556671143, + 0.0016462682979181409, + 0.002368250396102667, + 0.003034269670024514, + 0.0028175469487905502, + 0.0031136528123170137, + 0.0015950860688462853, + 0.0015167947858572006, + 0.002194320084527135, + 0.002018837258219719, + 0.0021799011155962944, + 0.00816071592271328, + 0.00392443360760808, + 0.0026754967402666807, + 0.0034826770424842834, + 0.02053704485297203, + 0.010500364005565643, + 0.0007248157635331154, + 0.0008020971436053514, + 0.0012905100593343377, + 0.0011166564654558897, + 0.000649957568384707, + 0.000977442367002368, + 0.0006146893138065934, + 0.0008938838145695627, + 0.003220177488401532, + 0.002217415953055024, + 0.0018613451393321157, + 0.001914221327751875, + 0.0019009198294952512, + 0.0018141588661819696, + 0.0018793968483805656, + 0.0023384750820696354, + 0.0034527417737990618, + 0.0027697558980435133, + 0.004343094769865274, + 0.006165811326354742, + 0.006320056039839983, + 0.004551338031888008, + 0.0054580774158239365, + 0.005200927145779133, + 0.00362238846719265, + 0.0019602247048169374, + 0.001845723483711481, + 0.0017365427920594811, + 0.0026254733093082905, + 0.003733560210093856, + 0.002188514219596982, + 0.0011860879603773355, + 0.012973125092685223, + 0.0020830293651670218, + 0.004926867783069611, + 0.005523771978914738, + 0.0034649220760911703, + 0.03999572619795799, + 0.011340562254190445, + 0.04183880612254143, + 0.0036568341311067343, + 0.003670960431918502, + 0.006032111588865519, + 0.005508712027221918, + 0.00270330673083663, + 0.0022869519889354706, + 0.0037593436427414417, + 0.003650888567790389 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.026329822838306427 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two chickens visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the chickens display feathers with iridescent greens, blues, and purples mixed with crimson plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the chickens performing leaping and soaring motions while kicking up dust and dry leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-dappled forest clearing with lush green foliage and hazy golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan fluidly upward and right to follow the chickens' movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007618538569658995 + ], + "opposite_logprob_tokens": [ + -4.882618427276611 + ], + "logprob_reward": -0.007618538569658995, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007618538569658995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05958638712763786, + 0.10695626586675644, + 0.052362702786922455, + 0.08714912086725235, + 0.017742104828357697, + 0.03213239461183548, + 0.08231063187122345, + 0.09012270718812943, + 0.44255268573760986, + 0.029084933921694756 + ], + "credit_map": [ + 0.00956726260483265, + 0.006768601946532726, + 0.008334748446941376, + 0.007043376099318266, + 0.006628149189054966, + 0.009074988774955273, + 0.006258936133235693, + 0.005910331383347511, + 0.013772266916930676, + 0.009520086459815502, + 0.012055971659719944, + 0.012214952148497105, + 0.01291815284639597, + 0.015343270264565945, + 0.008588233962655067, + 0.022543340921401978, + 0.005511164199560881, + 0.0067159938625991344, + 0.006402883213013411, + 0.0068814815022051334, + 0.009630411863327026, + 0.006007102318108082, + 0.005259958561509848, + 0.005953710060566664, + 0.0037248258013278246, + 0.004843763075768948, + 0.011710556223988533, + 0.006683601997792721, + 0.0049498723819851875, + 0.005405970383435488, + 0.030094148591160774, + 0.019736386835575104, + 0.0018720024963840842, + 0.0018048519268631935, + 0.002938640071079135, + 0.0020030485466122627, + 0.0015926830237731338, + 0.0023678068537265062, + 0.0017167766345664859, + 0.003446298185735941, + 0.005176707170903683, + 0.0051816427148878574, + 0.002676385687664151, + 0.003847630927339196, + 0.0037332437932491302, + 0.004125329200178385, + 0.003983388654887676, + 0.0034080669283866882, + 0.006716822274029255, + 0.005027648061513901, + 0.012408385053277016, + 0.012283168733119965, + 0.013303211890161037, + 0.007684881798923016, + 0.012555957771837711, + 0.012330564670264721, + 0.015485808253288269, + 0.014131843112409115, + 0.009680126793682575, + 0.0077489037066698074, + 0.011578653939068317, + 0.014277522452175617, + 0.01112404465675354, + 0.006095814052969217, + 0.04723383113741875, + 0.006982077844440937, + 0.016458764672279358, + 0.020202498883008957, + 0.011375362984836102, + 0.145904541015625, + 0.04163749888539314, + 0.15275812149047852, + 0.003938158042728901, + 0.002546394243836403, + 0.005106976255774498, + 0.0059130024164915085, + 0.002378311241045594, + 0.002418814692646265, + 0.0032977324444800615, + 0.003485547611489892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 7 + ], + "selected_reward": -0.007618538569658995 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two chickens visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the chickens display feathers with iridescent greens, blues, and purples mixed with crimson plumage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the chickens performing leaping and soaring motions while kicking up dust and dry leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-dappled forest clearing with lush green foliage and hazy golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan fluidly upward and right to follow the chickens' movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07894009351730347 + ], + "opposite_logprob_tokens": [ + -2.5789401531219482 + ], + "logprob_reward": -0.07894009351730347, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07894009351730347, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08069223165512085, + 0.010747130028903484, + 0.02236821874976158, + 0.015022085048258305, + 0.009904922917485237, + 0.005551591515541077, + 0.01616705022752285, + 0.01777614653110504, + 0.7979958057403564, + 0.023774825036525726 + ], + "credit_map": [ + 0.030808279290795326, + 0.003109160577878356, + 0.004935296718031168, + 0.0030742704402655363, + 0.003525573294609785, + 0.030597293749451637, + 0.00195685843937099, + 0.0026854900643229485, + 0.0016199122183024883, + 0.0010436748852953315, + 0.0011989460326731205, + 0.001489071175456047, + 0.001451387070119381, + 0.0016887509264051914, + 0.0008733725408092141, + 0.0013820149470120668, + 0.0030600938480347395, + 0.0014364332892000675, + 0.005063328426331282, + 0.003393650520592928, + 0.0022864060010761023, + 0.002736393129453063, + 0.001505616819486022, + 0.0028862953186035156, + 0.0012164507061243057, + 0.000987425446510315, + 0.0018800081452354789, + 0.0014277924783527851, + 0.0017059426754713058, + 0.001283426070585847, + 0.0031990204006433487, + 0.0033220178447663784, + 0.0008650423842482269, + 0.0009144313517026603, + 0.0015274385223165154, + 0.0013715055538341403, + 0.000827629235573113, + 0.0015262403758242726, + 0.0011020073434337974, + 0.0017706274520605803, + 0.0007696840330027044, + 0.0005883858539164066, + 0.0005501447594724596, + 0.0006297871586866677, + 0.0006924553890712559, + 0.000537192274350673, + 0.0008950428455136716, + 0.0008888983284123242, + 0.0021427541505545378, + 0.0011711958795785904, + 0.002646669512614608, + 0.00185841741040349, + 0.0013408137019723654, + 0.002458187285810709, + 0.0019372140523046255, + 0.0026117952074855566, + 0.0037444094195961952, + 0.001582493307068944, + 0.002009232062846422, + 0.0015894168755039573, + 0.0023667579516768456, + 0.0036093660164624453, + 0.0016629548044875264, + 0.0012115134159103036, + 0.08573873341083527, + 0.012132604606449604, + 0.02933432161808014, + 0.036245569586753845, + 0.021100379526615143, + 0.26323121786117554, + 0.07531087100505829, + 0.2749020755290985, + 0.003516853554174304, + 0.0017423442332074046, + 0.0037385995965451, + 0.005530399736016989, + 0.0015942063182592392, + 0.0018597549060359597, + 0.003421283094212413, + 0.002371381502598524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.07894009351730347 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_2/index_tensor([3647, 3647])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_2/index_tensor([3647, 3647])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d9364d8e6f6d409da03104212b438088d1333555 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_2/index_tensor([3647, 3647])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, low-angle, medium shot, rendered in the dramatic, high-contrast art style of Todd McFarlane, captures a massive rat-tail eel, perfectly captured at the center of the frame, its massive, muscular body coiled in the dark, stagnant swamp water. Deep, inky black shadows, characteristic of the thick, comic book ink, carve out the powerful contours of its form, creating a terrifying silhouette that dominates the scene. Sharp highlights flicker fiercely from its damp, patterned scales, defining the beast's texture and immense size. A slender, forked black tongue swiftly springs from its massive head, savoring the humid air as ethereal vapor rises from the dark, oppressive water around it. The massive snake, still and perfectly still, is a terrifying symbol of primal power, its unblinking eyes lost in the shadows, making the entire composition a masterclass in suspense and eerie beauty.", + "seed": 9883, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_2/index_3647_seed_9883_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05248463889583945, + "vlm_logprob": -0.05248463889583945, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.05248463889583945, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rat-tail eel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eel's body coiled in dark, stagnant swamp water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, forked black tongue spring from the eel's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by high-contrast comic book ink and sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle medium shot with the eel centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1466682255268097 + ], + "opposite_logprob_tokens": [ + -2.0216681957244873 + ], + "logprob_reward": -0.1466682255268097, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1466682255268097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04972485452890396, + 0.030269604176282883, + 0.015402151271700859, + 0.02790694683790207, + 0.15369467437267303, + 0.2314857393503189, + 0.16574111580848694, + 0.07183310389518738, + 0.134926438331604, + 0.1190153956413269 + ], + "credit_map": [ + 0.003655023640021682, + 0.0031477301381528378, + 0.009128469042479992, + 0.007260886486619711, + 0.0031903779599815607, + 0.007257453165948391, + 0.01026234868913889, + 0.005822568666189909, + 0.0032612087670713663, + 0.0022206648718565702, + 0.003253891132771969, + 0.005569107830524445, + 0.0025979699566960335, + 0.004866719711571932, + 0.003352728206664324, + 0.0051473150961101055, + 0.0012707990827038884, + 0.0013791542733088136, + 0.00240309233777225, + 0.0020107352174818516, + 0.0014682123437523842, + 0.0022244222927838564, + 0.0022339592687785625, + 0.0024117762222886086, + 0.0030522362794727087, + 0.003282019402831793, + 0.0035371987614780664, + 0.004016517661511898, + 0.002511034021154046, + 0.003462170483544469, + 0.004762476775795221, + 0.0032832918222993612, + 0.00582900969311595, + 0.008044646121561527, + 0.0038147808518260717, + 0.05954034626483917, + 0.004556036554276943, + 0.00700201652944088, + 0.006507030222564936, + 0.058400798588991165, + 0.014484432525932789, + 0.008002134039998055, + 0.00832424033433199, + 0.0839829370379448, + 0.005602455232292414, + 0.011411309242248535, + 0.01386276539415121, + 0.08581546694040298, + 0.01239478774368763, + 0.013402692042291164, + 0.026991503313183784, + 0.02316945232450962, + 0.011342975310981274, + 0.027168530970811844, + 0.030816972255706787, + 0.020454203709959984, + 0.007892019115388393, + 0.009085114113986492, + 0.007822907529771328, + 0.006534295622259378, + 0.009228527545928955, + 0.0079191317781806, + 0.013445298187434673, + 0.009905807673931122, + 0.009213024750351906, + 0.018065329641103745, + 0.018388036638498306, + 0.012536592781543732, + 0.01764640398323536, + 0.018966596573591232, + 0.019294707104563713, + 0.020815741270780563, + 0.011194243095815182, + 0.015878817066550255, + 0.015176653861999512, + 0.01930163986980915, + 0.010962061583995819, + 0.018164929002523422, + 0.011797811836004257, + 0.016539234668016434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.1466682255268097 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rat-tail eel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eel's body coiled in dark, stagnant swamp water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, forked black tongue spring from the eel's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by high-contrast comic book ink and sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle medium shot with the eel centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062087833881378174 + ], + "opposite_logprob_tokens": [ + -2.8120877742767334 + ], + "logprob_reward": -0.062087833881378174, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062087833881378174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09703314304351807, + 0.028806280344724655, + 0.022161686792969704, + 0.08399283140897751, + 0.16394801437854767, + 0.12215250730514526, + 0.1359667330980301, + 0.10987426340579987, + 0.14186064898967743, + 0.09420393407344818 + ], + "credit_map": [ + 0.005091872066259384, + 0.003676480147987604, + 0.015891069546341896, + 0.020594749599695206, + 0.004444988910108805, + 0.008056744001805782, + 0.018054524436593056, + 0.021222705021500587, + 0.0025233766064047813, + 0.0021885314490646124, + 0.005397415719926357, + 0.004418776370584965, + 0.0023682047612965107, + 0.0038392608985304832, + 0.004498610273003578, + 0.0035721012391149998, + 0.0015315852360799909, + 0.0019496214808896184, + 0.003228450194001198, + 0.003072171239182353, + 0.0023798616603016853, + 0.0032429692801088095, + 0.004020586609840393, + 0.0027364399284124374, + 0.00833278987556696, + 0.006471019703894854, + 0.013959648087620735, + 0.0101942652836442, + 0.006441786885261536, + 0.013168958947062492, + 0.012748836539685726, + 0.012675520032644272, + 0.006142512895166874, + 0.008315816521644592, + 0.00563679588958621, + 0.06285656988620758, + 0.004536853171885014, + 0.008611306548118591, + 0.0065018972381949425, + 0.061346255242824554, + 0.0042583835311234, + 0.004048556089401245, + 0.004226840101182461, + 0.041657015681266785, + 0.004714460577815771, + 0.007788395509123802, + 0.008731622248888016, + 0.046727217733860016, + 0.01452095340937376, + 0.011475911363959312, + 0.018189413473010063, + 0.016388744115829468, + 0.009728638455271721, + 0.018939124420285225, + 0.029088428243994713, + 0.017635516822338104, + 0.007044555619359016, + 0.00808724481612444, + 0.008602204732596874, + 0.009445065632462502, + 0.009423776529729366, + 0.03107747621834278, + 0.014773396775126457, + 0.021420544013381004, + 0.010327483527362347, + 0.014381586574018002, + 0.017102858051657677, + 0.015039278194308281, + 0.02137257158756256, + 0.02002008818089962, + 0.022084899246692657, + 0.021531881764531136, + 0.004835889209061861, + 0.007466031238436699, + 0.006084381137043238, + 0.01288708858191967, + 0.014298683032393456, + 0.014547663740813732, + 0.02191433683037758, + 0.0121698547154665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 6 + ], + "selected_reward": -0.062087833881378174 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rat-tail eel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eel's body coiled in dark, stagnant swamp water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, forked black tongue spring from the eel's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by high-contrast comic book ink and sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle medium shot with the eel centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058744616806507 + ], + "opposite_logprob_tokens": [ + -3.89558744430542 + ], + "logprob_reward": -0.02058744616806507, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058744616806507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0817570611834526, + 0.023928193375468254, + 0.026781514286994934, + 0.05448688939213753, + 0.09885870665311813, + 0.05725068598985672, + 0.12837465107440948, + 0.0542474091053009, + 0.05776689574122429, + 0.4165480434894562 + ], + "credit_map": [ + 0.005609956569969654, + 0.003977540414780378, + 0.007541852071881294, + 0.021691836416721344, + 0.0037492243573069572, + 0.009424230083823204, + 0.008965251967310905, + 0.02079717069864273, + 0.002142653102055192, + 0.0018180874176323414, + 0.004565482027828693, + 0.003529473440721631, + 0.0016050114063546062, + 0.0024199187755584717, + 0.003872205736115575, + 0.003975361119955778, + 0.0018206847598776221, + 0.00203317916020751, + 0.0035560422111302614, + 0.005048438441008329, + 0.0022556036710739136, + 0.0033987455535680056, + 0.00409646425396204, + 0.004572355188429356, + 0.00395688321441412, + 0.0038176896050572395, + 0.005099268164485693, + 0.008402065373957157, + 0.00421407725661993, + 0.007680289447307587, + 0.011188934557139874, + 0.010127680376172066, + 0.00436627259477973, + 0.005645180121064186, + 0.005875375121831894, + 0.03425615653395653, + 0.004140263888984919, + 0.00665254658088088, + 0.005444339010864496, + 0.03247857466340065, + 0.00592999579384923, + 0.004634940065443516, + 0.00644707540050149, + 0.013139832764863968, + 0.0040736692026257515, + 0.006852342747151852, + 0.007406348362565041, + 0.008766476996243, + 0.009756227023899555, + 0.010980733670294285, + 0.020984861999750137, + 0.01767725683748722, + 0.011075329966843128, + 0.022274401038885117, + 0.01999472640454769, + 0.01563110761344433, + 0.0034298552200198174, + 0.006027547642588615, + 0.005903205368667841, + 0.00467276107519865, + 0.005369170568883419, + 0.01278243400156498, + 0.005595822352916002, + 0.010466612875461578, + 0.004457545466721058, + 0.008037271909415722, + 0.007905090227723122, + 0.006136082578450441, + 0.007938534952700138, + 0.0058181812055408955, + 0.010917609557509422, + 0.006556584499776363, + 0.022173335775732994, + 0.03927945718169212, + 0.032547686249017715, + 0.058989543467760086, + 0.06853573769330978, + 0.06282887607812881, + 0.07405216991901398, + 0.05814123526215553 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 4 + ], + "selected_reward": -0.02058744616806507 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rat-tail eel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eel's body coiled in dark, stagnant swamp water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, forked black tongue spring from the eel's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by high-contrast comic book ink and sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle medium shot with the eel centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734886672347784 + ], + "opposite_logprob_tokens": [ + -5.006734848022461 + ], + "logprob_reward": -0.006734886672347784, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734886672347784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2276323437690735, + 0.03627737611532211, + 0.07172030210494995, + 0.13022011518478394, + 0.07278016954660416, + 0.11945949494838715, + 0.13918225467205048, + 0.1012144461274147, + 0.06930315494537354, + 0.03221029043197632 + ], + "credit_map": [ + 0.013955035246908665, + 0.014971787109971046, + 0.03341931104660034, + 0.04262550175189972, + 0.014722481369972229, + 0.027344513684511185, + 0.033554721623659134, + 0.047039009630680084, + 0.0030569557566195726, + 0.0031918527092784643, + 0.006125188432633877, + 0.005226566456258297, + 0.00242882058955729, + 0.004663752857595682, + 0.005183848086744547, + 0.006400394253432751, + 0.005897878669202328, + 0.004493256565183401, + 0.007409729994833469, + 0.012775227427482605, + 0.006294990424066782, + 0.009032860398292542, + 0.014722713269293308, + 0.011093651875853539, + 0.011145842261612415, + 0.009150062687695026, + 0.015117766335606575, + 0.01866685040295124, + 0.010844854637980461, + 0.021164601668715477, + 0.026387503370642662, + 0.01774263195693493, + 0.003925932105630636, + 0.004080676008015871, + 0.005524501204490662, + 0.021195314824581146, + 0.003480389015749097, + 0.005464836023747921, + 0.007493909914046526, + 0.021614618599414825, + 0.02808486483991146, + 0.005068153142929077, + 0.011638865806162357, + 0.018388768658041954, + 0.008269033394753933, + 0.01429186575114727, + 0.013325403444468975, + 0.02039254829287529, + 0.015706965699791908, + 0.009299139492213726, + 0.014762932434678078, + 0.012338832952082157, + 0.013548636808991432, + 0.03001290000975132, + 0.02488509751856327, + 0.018627775833010674, + 0.009408420883119106, + 0.00881259236484766, + 0.009058309718966484, + 0.00776075990870595, + 0.008302910253405571, + 0.02039363607764244, + 0.015184314921498299, + 0.022293513640761375, + 0.0060194809921085835, + 0.008976010605692863, + 0.01153064426034689, + 0.007863807491958141, + 0.008843044750392437, + 0.006596538238227367, + 0.010995348915457726, + 0.008478287607431412, + 0.002246418735012412, + 0.003523898310959339, + 0.0038517892826348543, + 0.0043683042749762535, + 0.004274926148355007, + 0.005117336753755808, + 0.004156963434070349, + 0.004670656751841307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.006734886672347784 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rat-tail eel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eel's body coiled in dark, stagnant swamp water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, forked black tongue spring from the eel's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by high-contrast comic book ink and sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle medium shot with the eel centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026344802230596542 + ], + "opposite_logprob_tokens": [ + -3.6513447761535645 + ], + "logprob_reward": -0.026344802230596542, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026344802230596542, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0979776456952095, + 0.047114480286836624, + 0.035111382603645325, + 0.05283157899975777, + 0.2071947604417801, + 0.1378667652606964, + 0.07376505434513092, + 0.16221098601818085, + 0.05156846344470978, + 0.13435889780521393 + ], + "credit_map": [ + 0.009072087705135345, + 0.0050379084423184395, + 0.02215941809117794, + 0.014908005483448505, + 0.006276764906942844, + 0.01044551096856594, + 0.016514640301465988, + 0.013563312590122223, + 0.004725068807601929, + 0.0036620451137423515, + 0.005065262783318758, + 0.007943498902022839, + 0.003874817630276084, + 0.008015671744942665, + 0.0061379834078252316, + 0.007690132595598698, + 0.0035382446367293596, + 0.002278854139149189, + 0.004678973462432623, + 0.00508031016215682, + 0.003928935155272484, + 0.004300959408283234, + 0.005272767972201109, + 0.0060323383659124374, + 0.005420680157840252, + 0.0043534222058951855, + 0.008041009306907654, + 0.0075079589150846004, + 0.004465311765670776, + 0.008124002255499363, + 0.009345291182398796, + 0.005573899485170841, + 0.0075336722657084465, + 0.011571635492146015, + 0.0078372647985816, + 0.0776245966553688, + 0.006417911034077406, + 0.010302750393748283, + 0.008403075858950615, + 0.07750385254621506, + 0.036687325686216354, + 0.006342592183500528, + 0.01584767922759056, + 0.022670188918709755, + 0.008700462989509106, + 0.014089230448007584, + 0.013255598023533821, + 0.020273685455322266, + 0.008584240451455116, + 0.004287271294742823, + 0.011457796208560467, + 0.008129996247589588, + 0.005420240573585033, + 0.010555948130786419, + 0.014851693995296955, + 0.010477867908775806, + 0.008052865043282509, + 0.006944847293198109, + 0.00781609583646059, + 0.010650093667209148, + 0.005285680294036865, + 0.059265218675136566, + 0.008544706739485264, + 0.05565148964524269, + 0.003390286350622773, + 0.0050684502348303795, + 0.006739198230206966, + 0.0060762520879507065, + 0.008588366210460663, + 0.007874800823628902, + 0.006957927718758583, + 0.006873180624097586, + 0.010805453173816204, + 0.00984531082212925, + 0.011412924155592918, + 0.016083477064967155, + 0.02396199107170105, + 0.023074431344866753, + 0.021219750866293907, + 0.017955545336008072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 5 + ], + "selected_reward": -0.026344802230596542 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, medium shot, rendered in the dramatic, high-contrast art style of Todd McFarlane, captures a massive rat-tail eel, perfectly captured at the center of the frame, its massive, muscular body coiled in the dark, stagnant swamp water. Deep, inky black shadows, characteristic of the thick, comic book ink, carve out the powerful contours of its form, creating a terrifying silhouette that dominates the scene. Sharp highlights flicker fiercely from its damp, patterned scales, defining the beast's texture and immense size. A slender, forked black tongue swiftly springs from its massive head, savoring the humid air as ethereal vapor rises from the dark, oppressive water around it. The massive snake, still and perfectly still, is a terrifying symbol of primal power, its unblinking eyes lost in the shadows, making the entire composition a masterclass in suspense and eerie beauty.", + "seed": 9884, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_2/index_3647_seed_9884_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04288462549448013, + "vlm_logprob": -0.04288462549448013, + "vlm_yes_no_margin": 3.4999999523162844, + "vlm_token_credit": -0.04288462549448013, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rat-tail eel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eel's body coiled in dark, stagnant swamp water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, forked black tongue spring from the eel's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by high-contrast comic book ink and sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle medium shot with the eel centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13050861656665802 + ], + "opposite_logprob_tokens": [ + -2.1305086612701416 + ], + "logprob_reward": -0.13050861656665802, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13050861656665802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22681817412376404, + 0.1167512834072113, + 0.03229166939854622, + 0.1481839269399643, + 0.03100213222205639, + 0.07019219547510147, + 0.04963701590895653, + 0.031578436493873596, + 0.13256718218326569, + 0.1609780490398407 + ], + "credit_map": [ + 0.01783420331776142, + 0.013025541789829731, + 0.018460571765899658, + 0.009273327887058258, + 0.011187214404344559, + 0.06854395568370819, + 0.012985440902411938, + 0.07550792396068573, + 0.01101995725184679, + 0.00850099790841341, + 0.01240807119756937, + 0.008009100332856178, + 0.011204103007912636, + 0.03159036859869957, + 0.012044229544699192, + 0.021974457427859306, + 0.006261143367737532, + 0.0024806628935039043, + 0.00305475783534348, + 0.004162710625678301, + 0.002296470571309328, + 0.004465488251298666, + 0.0036165083292871714, + 0.005953927058726549, + 0.012288798578083515, + 0.004515735432505608, + 0.008265041746199131, + 0.0062445118092000484, + 0.008256672881543636, + 0.05229954048991203, + 0.005975787062197924, + 0.05033784732222557, + 0.004033870995044708, + 0.003099147230386734, + 0.004983444232493639, + 0.004264899995177984, + 0.003016793867573142, + 0.005286211613565683, + 0.0028123229276388884, + 0.003505439031869173, + 0.01013793982565403, + 0.01144337933510542, + 0.00786177534610033, + 0.009046354331076145, + 0.005909775383770466, + 0.009236742742359638, + 0.005140178836882114, + 0.011416050605475903, + 0.006424955558031797, + 0.0034163719974458218, + 0.0062602185644209385, + 0.008409355767071247, + 0.0032191264908760786, + 0.0069564105942845345, + 0.0029236851260066032, + 0.012026889249682426, + 0.0036414063069969416, + 0.0030435766093432903, + 0.004423374775797129, + 0.00596593227237463, + 0.0036361345555633307, + 0.003289316315203905, + 0.0032834617886692286, + 0.004295232705771923, + 0.011999049223959446, + 0.0023174379020929337, + 0.006308004260063171, + 0.002328973263502121, + 0.0055795833468437195, + 0.045617908239364624, + 0.0046941787004470825, + 0.05372205004096031, + 0.027492914348840714, + 0.012808571569621563, + 0.029301457107067108, + 0.014612220227718353, + 0.01479366421699524, + 0.016241857782006264, + 0.010255556553602219, + 0.035471800714731216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.13050861656665802 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rat-tail eel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eel's body coiled in dark, stagnant swamp water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, forked black tongue spring from the eel's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by high-contrast comic book ink and sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle medium shot with the eel centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018214929848909378 + ], + "opposite_logprob_tokens": [ + -4.018214702606201 + ], + "logprob_reward": -0.018214929848909378, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018214929848909378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07379063963890076, + 0.32378265261650085, + 0.10557595640420914, + 0.09180677682161331, + 0.013168028555810452, + 0.04521927982568741, + 0.024548029527068138, + 0.01661396399140358, + 0.19431106746196747, + 0.11118363589048386 + ], + "credit_map": [ + 0.00703246146440506, + 0.005984847899526358, + 0.01047150045633316, + 0.0043856254778802395, + 0.005632725078612566, + 0.013669077306985855, + 0.0066871787421405315, + 0.019927218556404114, + 0.0262433048337698, + 0.016642199829220772, + 0.024633152410387993, + 0.012966705486178398, + 0.014794966205954552, + 0.09950275719165802, + 0.01623016595840454, + 0.11276940256357193, + 0.009249914437532425, + 0.004990389104932547, + 0.009873449802398682, + 0.006110424641519785, + 0.004485426936298609, + 0.02887440100312233, + 0.006290640216320753, + 0.03570130839943886, + 0.009943140670657158, + 0.00239035626873374, + 0.003169855335727334, + 0.002501793671399355, + 0.0027505275793373585, + 0.03854852542281151, + 0.0022626062855124474, + 0.030239976942539215, + 0.0015662279911339283, + 0.0014412070158869028, + 0.0017603925662115216, + 0.002128826454281807, + 0.0013242446584627032, + 0.0019915206357836723, + 0.0014634665567427874, + 0.0014921437250450253, + 0.004295803140848875, + 0.001807939843274653, + 0.0027318533975631, + 0.0018893120577558875, + 0.0020840258803218603, + 0.016200464218854904, + 0.0023478216025978327, + 0.013862059451639652, + 0.003503733314573765, + 0.0024631302803754807, + 0.0032304890919476748, + 0.0037581860087811947, + 0.0024770437739789486, + 0.00409639673307538, + 0.00210983376018703, + 0.0029092179611325264, + 0.0020560368429869413, + 0.0016602188115939498, + 0.0018565591890364885, + 0.0018093402031809092, + 0.001566028455272317, + 0.0018442232394590974, + 0.001312774489633739, + 0.00450878357514739, + 0.0156322680413723, + 0.004223331343382597, + 0.01081835851073265, + 0.0046730502508580685, + 0.009817286394536495, + 0.0680294781923294, + 0.007013550493866205, + 0.07410375028848648, + 0.019816571846604347, + 0.009131746366620064, + 0.02089003287255764, + 0.011507421731948853, + 0.009222702123224735, + 0.012536855414509773, + 0.007144453469663858, + 0.020933859050273895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 9 + ], + "selected_reward": -0.018214929848909378 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rat-tail eel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eel's body coiled in dark, stagnant swamp water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, forked black tongue spring from the eel's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by high-contrast comic book ink and sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle medium shot with the eel centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023293664678931236 + ], + "opposite_logprob_tokens": [ + -3.7732937335968018 + ], + "logprob_reward": -0.023293664678931236, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023293664678931236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0945499911904335, + 0.07990725338459015, + 0.027196507900953293, + 0.46399497985839844, + 0.02075600065290928, + 0.1331222504377365, + 0.031261052936315536, + 0.02025407738983631, + 0.09958229959011078, + 0.029375553131103516 + ], + "credit_map": [ + 0.010321011766791344, + 0.01136432308703661, + 0.013043132610619068, + 0.008441434241831303, + 0.013460800983011723, + 0.013763348571956158, + 0.0153969032689929, + 0.008759044110774994, + 0.00873154029250145, + 0.0036452326457947493, + 0.007893768139183521, + 0.004107100889086723, + 0.004778600763529539, + 0.01828317530453205, + 0.005235875491052866, + 0.027231965214014053, + 0.005359117407351732, + 0.0017392985755577683, + 0.0029925017151981592, + 0.003218254307284951, + 0.0019159942166879773, + 0.007660761941224337, + 0.0018300032243132591, + 0.0024805788416415453, + 0.05125947669148445, + 0.011540370993316174, + 0.014185626991093159, + 0.012720948085188866, + 0.01329695526510477, + 0.1957571655511856, + 0.010809283703565598, + 0.15442515909671783, + 0.002518783789128065, + 0.003000512020662427, + 0.0025102554354816675, + 0.0029438405763357878, + 0.002287068171426654, + 0.002457359340041876, + 0.0023525776341557503, + 0.0026856069453060627, + 0.01805057004094124, + 0.008567290380597115, + 0.00863371230661869, + 0.0077598621137440205, + 0.005783080589026213, + 0.04851517453789711, + 0.00505234207957983, + 0.030760226771235466, + 0.0032668327912688255, + 0.0026113076601177454, + 0.0034799915738403797, + 0.0030534369871020317, + 0.002183479955419898, + 0.004794294945895672, + 0.0017431129235774279, + 0.010128598660230637, + 0.002822126029059291, + 0.0015486079500988126, + 0.002493769396096468, + 0.0030049511697143316, + 0.002084062434732914, + 0.002916493685916066, + 0.0016886345110833645, + 0.003695433959364891, + 0.01312280260026455, + 0.003298767376691103, + 0.007983343675732613, + 0.00415430124849081, + 0.005200997926294804, + 0.021248655393719673, + 0.005223914515227079, + 0.03934952989220619, + 0.0036870234180241823, + 0.0021812054328620434, + 0.0036173551343381405, + 0.0032906392589211464, + 0.0033236760646104813, + 0.0036380146630108356, + 0.002533117076382041, + 0.007104526273906231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 8 + ], + "selected_reward": -0.023293664678931236 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rat-tail eel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eel's body coiled in dark, stagnant swamp water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, forked black tongue spring from the eel's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by high-contrast comic book ink and sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle medium shot with the eel centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026342827826738358 + ], + "opposite_logprob_tokens": [ + -3.6513428688049316 + ], + "logprob_reward": -0.026342827826738358, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026342827826738358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14657340943813324, + 0.08211370557546616, + 0.033829644322395325, + 0.01538537535816431, + 0.048564061522483826, + 0.43382418155670166, + 0.10045958310365677, + 0.028167150914669037, + 0.09036145359277725, + 0.020721454173326492 + ], + "credit_map": [ + 0.015214944258332253, + 0.009992503561079502, + 0.01282742340117693, + 0.008854683488607407, + 0.009567472152411938, + 0.045781899243593216, + 0.010409487411379814, + 0.033924996852874756, + 0.006054349709302187, + 0.004796772263944149, + 0.007885786704719067, + 0.0042906953021883965, + 0.0053388262167572975, + 0.026223327964544296, + 0.0047616963274776936, + 0.022762257605791092, + 0.0040054176934063435, + 0.003017163136973977, + 0.005127223674207926, + 0.005342583637684584, + 0.002634678967297077, + 0.0067483726888895035, + 0.0029619268607348204, + 0.0039922744035720825, + 0.0020245041232556105, + 0.0012583148200064898, + 0.002838576678186655, + 0.0021994691342115402, + 0.001734154298901558, + 0.002063426887616515, + 0.001210299669764936, + 0.0020566298626363277, + 0.005342085845768452, + 0.0036818741355091333, + 0.00848762970417738, + 0.008020439185202122, + 0.0035941721871495247, + 0.006042331922799349, + 0.005593236070126295, + 0.007802292238920927, + 0.04451700672507286, + 0.012072436511516571, + 0.01866297423839569, + 0.011797419749200344, + 0.010183473117649555, + 0.17530615627765656, + 0.01182869914919138, + 0.14945603907108307, + 0.0107547203078866, + 0.0077616991475224495, + 0.012685791589319706, + 0.00922960601747036, + 0.006021338980644941, + 0.01522342674434185, + 0.005624406971037388, + 0.033158596605062485, + 0.0037775104865431786, + 0.0017469258746132255, + 0.003630691906437278, + 0.004375701770186424, + 0.0020426816772669554, + 0.0038598591927438974, + 0.002064518863335252, + 0.006669258698821068, + 0.009230501018464565, + 0.002059649908915162, + 0.0047484287060797215, + 0.001972336322069168, + 0.0035476461052894592, + 0.027761781588196754, + 0.0034896915312856436, + 0.03755142167210579, + 0.0032482147216796875, + 0.0013724548043683171, + 0.0026144457515329123, + 0.0015130458632484078, + 0.002506736433133483, + 0.0023654012475162745, + 0.001558061339892447, + 0.005543093662708998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.026342827826738358 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rat-tail eel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eel's body coiled in dark, stagnant swamp water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, forked black tongue spring from the eel's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by high-contrast comic book ink and sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle medium shot with the eel centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016063088551163673 + ], + "opposite_logprob_tokens": [ + -4.141063213348389 + ], + "logprob_reward": -0.016063088551163673, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016063088551163673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07984281331300735, + 0.11203951388597488, + 0.06884067505598068, + 0.12197553366422653, + 0.018211806192994118, + 0.19735224545001984, + 0.03910166770219803, + 0.04439069330692291, + 0.2518644630908966, + 0.06638070940971375 + ], + "credit_map": [ + 0.00943783950060606, + 0.005290118046104908, + 0.012213359586894512, + 0.004739530850201845, + 0.006014390382915735, + 0.02129046805202961, + 0.007191685494035482, + 0.013665401376783848, + 0.008755624294281006, + 0.00528140040114522, + 0.008170749060809612, + 0.0040029496885836124, + 0.0050893304869532585, + 0.037135377526283264, + 0.004579812753945589, + 0.03902424871921539, + 0.010992506518959999, + 0.002963984850794077, + 0.002803694922477007, + 0.004083692561835051, + 0.00263985269702971, + 0.029301078990101814, + 0.003098045941442251, + 0.012957804836332798, + 0.014029247686266899, + 0.003710878547281027, + 0.004957824479788542, + 0.0037720066029578447, + 0.0048852525651454926, + 0.04970495402812958, + 0.0033576947171241045, + 0.03755765035748482, + 0.002134694717824459, + 0.0018623527139425278, + 0.0022043155040591955, + 0.0025409474037587643, + 0.00205149594694376, + 0.0018181444611400366, + 0.002044635359197855, + 0.003555215895175934, + 0.026675162836909294, + 0.009669332765042782, + 0.010169357992708683, + 0.007687277626246214, + 0.006930455099791288, + 0.07750269770622253, + 0.006473435554653406, + 0.052244484424591064, + 0.005559887737035751, + 0.003001667559146881, + 0.004486724268645048, + 0.006596127059310675, + 0.0029845142271369696, + 0.004590120166540146, + 0.0025186375714838505, + 0.009363980032503605, + 0.004975477233529091, + 0.0043931384570896626, + 0.0060705337673425674, + 0.008518192917108536, + 0.004602761473506689, + 0.005586274899542332, + 0.0044937520287930965, + 0.00575055368244648, + 0.02122984454035759, + 0.00497030932456255, + 0.010455965064466, + 0.004685279447585344, + 0.012127327732741833, + 0.09224119782447815, + 0.00844988226890564, + 0.09770461171865463, + 0.010611405596137047, + 0.004292100667953491, + 0.010033241473138332, + 0.0040470608510077, + 0.007124847732484341, + 0.006916559301316738, + 0.004705438390374184, + 0.01865004189312458 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 3 + ], + "selected_reward": -0.016063088551163673 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, medium shot, rendered in the dramatic, high-contrast art style of Todd McFarlane, captures a massive rat-tail eel, perfectly captured at the center of the frame, its massive, muscular body coiled in the dark, stagnant swamp water. Deep, inky black shadows, characteristic of the thick, comic book ink, carve out the powerful contours of its form, creating a terrifying silhouette that dominates the scene. Sharp highlights flicker fiercely from its damp, patterned scales, defining the beast's texture and immense size. A slender, forked black tongue swiftly springs from its massive head, savoring the humid air as ethereal vapor rises from the dark, oppressive water around it. The massive snake, still and perfectly still, is a terrifying symbol of primal power, its unblinking eyes lost in the shadows, making the entire composition a masterclass in suspense and eerie beauty.", + "seed": 9885, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_2/index_3647_seed_9885_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03296602424234152, + "vlm_logprob": -0.03296602424234152, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.03296602424234152, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rat-tail eel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eel's body coiled in dark, stagnant swamp water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, forked black tongue spring from the eel's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by high-contrast comic book ink and sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle medium shot with the eel centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08175536245107651 + ], + "opposite_logprob_tokens": [ + -2.5817553997039795 + ], + "logprob_reward": -0.08175536245107651, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08175536245107651, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07594714313745499, + 0.014881515875458717, + 0.05212301388382912, + 0.04720565676689148, + 0.015742763876914978, + 0.12433337420225143, + 0.04592578113079071, + 0.04630705714225769, + 0.559719979763031, + 0.017813770100474358 + ], + "credit_map": [ + 0.0068305013701319695, + 0.0038002529181540012, + 0.012587804347276688, + 0.006370130926370621, + 0.013749266043305397, + 0.016623154282569885, + 0.006196384318172932, + 0.009789648465812206, + 0.0012158432509750128, + 0.001016179914586246, + 0.002882039872929454, + 0.0018091112142428756, + 0.0015081517631188035, + 0.0020443727262318134, + 0.0017803141381591558, + 0.002625502645969391, + 0.002146773273125291, + 0.0012269743019714952, + 0.019634829834103584, + 0.0032540648244321346, + 0.0028034665156155825, + 0.019335554912686348, + 0.0020212645176798105, + 0.0017000884981825948, + 0.006304146721959114, + 0.003918820526450872, + 0.00850608665496111, + 0.004986590705811977, + 0.005762407090514898, + 0.006210836581885815, + 0.004208661615848541, + 0.0073081073351204395, + 0.0022079800255596638, + 0.0013125102268531919, + 0.002226883778348565, + 0.002121230121701956, + 0.002169327111914754, + 0.002221126575022936, + 0.001537436037324369, + 0.0019462702330201864, + 0.009204071946442127, + 0.012810572981834412, + 0.023880213499069214, + 0.01000358909368515, + 0.022686585783958435, + 0.01645945943892002, + 0.016311796382069588, + 0.012977082282304764, + 0.004869160708039999, + 0.002784277079626918, + 0.007014568895101547, + 0.006847224663943052, + 0.007436675019562244, + 0.005171140190213919, + 0.006747446022927761, + 0.005055288318544626, + 0.0063595836982131, + 0.003551327856257558, + 0.008084526285529137, + 0.005645533557981253, + 0.006113032344728708, + 0.00545901944860816, + 0.006987837143242359, + 0.004106196109205484, + 0.10293329507112503, + 0.035384055227041245, + 0.051545437425374985, + 0.040925055742263794, + 0.08235256373882294, + 0.06787776201963425, + 0.09378594905138016, + 0.08491584658622742, + 0.0017859225627034903, + 0.0014405347174033523, + 0.0027687200345098972, + 0.0011370469583198428, + 0.0017417786875739694, + 0.0024825523141771555, + 0.0044941226951777935, + 0.001963092014193535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.08175536245107651 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rat-tail eel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eel's body coiled in dark, stagnant swamp water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, forked black tongue spring from the eel's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by high-contrast comic book ink and sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle medium shot with the eel centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02637440897524357 + ], + "opposite_logprob_tokens": [ + -3.651374340057373 + ], + "logprob_reward": -0.02637440897524357, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02637440897524357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17087717354297638, + 0.0963604524731636, + 0.17027372121810913, + 0.0940733402967453, + 0.038819320499897, + 0.06296753883361816, + 0.059958238154649734, + 0.10843677073717117, + 0.0986209511756897, + 0.09961256384849548 + ], + "credit_map": [ + 0.012774261645972729, + 0.00848193559795618, + 0.03596419468522072, + 0.014820980839431286, + 0.014852841384708881, + 0.058264296501874924, + 0.011318780481815338, + 0.014399890787899494, + 0.01098676398396492, + 0.005089061800390482, + 0.023384349420666695, + 0.010210431180894375, + 0.01014450378715992, + 0.014780585654079914, + 0.01095961220562458, + 0.010805146768689156, + 0.007523519918322563, + 0.002835767576470971, + 0.064896360039711, + 0.011517660692334175, + 0.006982064805924892, + 0.0665719211101532, + 0.004816108383238316, + 0.005130312871187925, + 0.011503161862492561, + 0.00734185753390193, + 0.01600746624171734, + 0.014220082201063633, + 0.009151885285973549, + 0.012200294993817806, + 0.011259534396231174, + 0.012389054521918297, + 0.0034706781152635813, + 0.0030585359781980515, + 0.007359582930803299, + 0.005447664763778448, + 0.004029749892652035, + 0.005852727685123682, + 0.004176157061010599, + 0.005424222443252802, + 0.008662266656756401, + 0.0053941188380122185, + 0.008566824719309807, + 0.005638530012220144, + 0.010270263068377972, + 0.00775781599804759, + 0.007152218371629715, + 0.009525503031909466, + 0.005219264887273312, + 0.004773275926709175, + 0.01056633424013853, + 0.010356592014431953, + 0.007614572532474995, + 0.00709566893056035, + 0.007610098458826542, + 0.006722434423863888, + 0.017254944890737534, + 0.0077830092050135136, + 0.016671154648065567, + 0.012285477481782436, + 0.01496872492134571, + 0.013234241865575314, + 0.014113950543105602, + 0.012125270441174507, + 0.02046014554798603, + 0.00598616898059845, + 0.00851515308022499, + 0.007222755346447229, + 0.013503030873835087, + 0.012698773294687271, + 0.01528273057192564, + 0.014952195808291435, + 0.018209582194685936, + 0.007726198527961969, + 0.013372935354709625, + 0.006679882761090994, + 0.011465054005384445, + 0.013150256127119064, + 0.01630074717104435, + 0.012707914225757122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.02637440897524357 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rat-tail eel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eel's body coiled in dark, stagnant swamp water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, forked black tongue spring from the eel's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by high-contrast comic book ink and sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle medium shot with the eel centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020582424476742744 + ], + "opposite_logprob_tokens": [ + -3.895582437515259 + ], + "logprob_reward": -0.020582424476742744, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020582424476742744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03599131107330322, + 0.027755681425333023, + 0.07853282988071442, + 0.11138560622930527, + 0.39954039454460144, + 0.08190689980983734, + 0.17425930500030518, + 0.029005195945501328, + 0.04660526290535927, + 0.015017497353255749 + ], + "credit_map": [ + 0.0023060610983520746, + 0.002746144076809287, + 0.007391021121293306, + 0.002917261328548193, + 0.004329077899456024, + 0.007392655126750469, + 0.0036883254069834948, + 0.005220766644924879, + 0.00264377030543983, + 0.0016080139903351665, + 0.0056939879432320595, + 0.003434437559917569, + 0.0022866304498165846, + 0.006168887484818697, + 0.0024054343812167645, + 0.003514517331495881, + 0.0028206249698996544, + 0.0014860752271488309, + 0.03066074103116989, + 0.0049984036013484, + 0.0025924702640622854, + 0.03001067228615284, + 0.00292068999260664, + 0.0030431593768298626, + 0.014664066955447197, + 0.009427485056221485, + 0.01857367716729641, + 0.011451792903244495, + 0.014963651075959206, + 0.015339759178459644, + 0.010800213553011417, + 0.01616496406495571, + 0.04126480594277382, + 0.03090151958167553, + 0.06867343932390213, + 0.05811823159456253, + 0.03775326907634735, + 0.06193022057414055, + 0.044195014983415604, + 0.05670391768217087, + 0.013933551497757435, + 0.0069880555383861065, + 0.01086611207574606, + 0.007407361641526222, + 0.011555438861250877, + 0.009928570128977299, + 0.008059068582952023, + 0.013168741948902607, + 0.010550561361014843, + 0.009328010492026806, + 0.02403387241065502, + 0.027481911703944206, + 0.012958469800651073, + 0.016900571063160896, + 0.05623776465654373, + 0.016768142580986023, + 0.003158074803650379, + 0.0021491774823516607, + 0.004931860137730837, + 0.002697315998375416, + 0.005168877076357603, + 0.00351185561157763, + 0.0038970194291323423, + 0.0034910154063254595, + 0.008312195539474487, + 0.003018626244738698, + 0.004260942339897156, + 0.00355199771001935, + 0.007695173844695091, + 0.005554506089538336, + 0.0076489634811878204, + 0.006562856957316399, + 0.0009621867793612182, + 0.0012663669185712934, + 0.00145951216109097, + 0.001233370858244598, + 0.0020089298486709595, + 0.002076418837532401, + 0.004015619400888681, + 0.001995092025026679 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 3 + ], + "selected_reward": -0.020582424476742744 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rat-tail eel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eel's body coiled in dark, stagnant swamp water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, forked black tongue spring from the eel's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by high-contrast comic book ink and sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle medium shot with the eel centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009778928942978382 + ], + "opposite_logprob_tokens": [ + -4.63477897644043 + ], + "logprob_reward": -0.009778928942978382, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009778928942978382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1476602852344513, + 0.12616871297359467, + 0.0931297242641449, + 0.05725488066673279, + 0.12106672674417496, + 0.08687491714954376, + 0.11748270690441132, + 0.06441599875688553, + 0.09786862879991531, + 0.08807743340730667 + ], + "credit_map": [ + 0.01532914862036705, + 0.011427934281527996, + 0.01937120407819748, + 0.014252376742661, + 0.030180474743247032, + 0.02169889397919178, + 0.014224011451005936, + 0.02117624506354332, + 0.010855762287974358, + 0.007582494523376226, + 0.029746662825345993, + 0.012002408504486084, + 0.012234407477080822, + 0.02569444850087166, + 0.012126869522035122, + 0.015925662592053413, + 0.004414634313434362, + 0.0026492192409932613, + 0.031264759600162506, + 0.00794933084398508, + 0.005334934219717979, + 0.029054604470729828, + 0.0065225097350776196, + 0.005939740687608719, + 0.005139862187206745, + 0.003739850129932165, + 0.010597999207675457, + 0.0084735918790102, + 0.006538390647619963, + 0.008176338858902454, + 0.006857316009700298, + 0.007731531746685505, + 0.01172822155058384, + 0.008474333211779594, + 0.02513127215206623, + 0.014487802051007748, + 0.02100200392305851, + 0.015382898971438408, + 0.010537357069551945, + 0.01432284340262413, + 0.009927907027304173, + 0.00726966792717576, + 0.013343787752091885, + 0.010489214211702347, + 0.009808444418013096, + 0.01077190786600113, + 0.012142824940383434, + 0.013121158815920353, + 0.010366145521402359, + 0.006185288541018963, + 0.014399630948901176, + 0.01576763391494751, + 0.012031721882522106, + 0.010975818149745464, + 0.03589572384953499, + 0.011860745027661324, + 0.005297652445733547, + 0.004984695464372635, + 0.008115723729133606, + 0.006003865972161293, + 0.012246789410710335, + 0.007770059630274773, + 0.00916904304176569, + 0.010828168131411076, + 0.01248759776353836, + 0.00588202802464366, + 0.013083860278129578, + 0.006489668972790241, + 0.014723950996994972, + 0.011764773167669773, + 0.019319334998726845, + 0.014117419719696045, + 0.010326456278562546, + 0.008182910270988941, + 0.013076968491077423, + 0.007016888353973627, + 0.010172249749302864, + 0.011736488901078701, + 0.016194790601730347, + 0.011370684020221233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.009778928942978382 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rat-tail eel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eel's body coiled in dark, stagnant swamp water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, forked black tongue spring from the eel's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by high-contrast comic book ink and sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle medium shot with the eel centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02633899636566639 + ], + "opposite_logprob_tokens": [ + -3.651339054107666 + ], + "logprob_reward": -0.02633899636566639, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02633899636566639, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23431287705898285, + 0.07610291987657547, + 0.0417451374232769, + 0.12879924476146698, + 0.06904693692922592, + 0.06070174649357796, + 0.17724652588367462, + 0.11271589249372482, + 0.06490164250135422, + 0.03442708030343056 + ], + "credit_map": [ + 0.01697228103876114, + 0.011504762805998325, + 0.045984864234924316, + 0.018320731818675995, + 0.029320379719138145, + 0.07505685091018677, + 0.01568182185292244, + 0.021471194922924042, + 0.0058753602206707, + 0.003786723595112562, + 0.017216041684150696, + 0.006866557523608208, + 0.006604624912142754, + 0.019084472209215164, + 0.006565145682543516, + 0.010103990323841572, + 0.004499209113419056, + 0.0022779323626309633, + 0.008588967844843864, + 0.005048987455666065, + 0.0032097450457513332, + 0.008528315462172031, + 0.004767495207488537, + 0.004824485629796982, + 0.01472113560885191, + 0.010026417672634125, + 0.02313905581831932, + 0.020830674096941948, + 0.012852364219725132, + 0.015155276283621788, + 0.015597241930663586, + 0.016477078199386597, + 0.008772892877459526, + 0.005982229486107826, + 0.010033448226749897, + 0.008186285383999348, + 0.008305522613227367, + 0.008690237067639828, + 0.0073195938020944595, + 0.011756723746657372, + 0.006086551118642092, + 0.005697815679013729, + 0.009663461707532406, + 0.00560079887509346, + 0.010915382765233517, + 0.0077293128706514835, + 0.008507234044373035, + 0.006501182448118925, + 0.01316971704363823, + 0.01023243460804224, + 0.02281925082206726, + 0.02494972012937069, + 0.017435817047953606, + 0.01699407771229744, + 0.0534043051302433, + 0.01824120059609413, + 0.018662847578525543, + 0.00896549317985773, + 0.014583641663193703, + 0.011013336479663849, + 0.014515566639602184, + 0.013254671357572079, + 0.016195712611079216, + 0.015524633228778839, + 0.010628140531480312, + 0.0041587622836232185, + 0.008513194508850574, + 0.004814166110008955, + 0.008906419388949871, + 0.007969442754983902, + 0.012004674412310123, + 0.007906842976808548, + 0.005720729939639568, + 0.0031591227743774652, + 0.004718065727502108, + 0.0030246784444898367, + 0.003807268338277936, + 0.004512438550591469, + 0.005987005308270454, + 0.0034977684263139963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.02633899636566639 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, medium shot, rendered in the dramatic, high-contrast art style of Todd McFarlane, captures a massive rat-tail eel, perfectly captured at the center of the frame, its massive, muscular body coiled in the dark, stagnant swamp water. Deep, inky black shadows, characteristic of the thick, comic book ink, carve out the powerful contours of its form, creating a terrifying silhouette that dominates the scene. Sharp highlights flicker fiercely from its damp, patterned scales, defining the beast's texture and immense size. A slender, forked black tongue swiftly springs from its massive head, savoring the humid air as ethereal vapor rises from the dark, oppressive water around it. The massive snake, still and perfectly still, is a terrifying symbol of primal power, its unblinking eyes lost in the shadows, making the entire composition a masterclass in suspense and eerie beauty.", + "seed": 9886, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_2/index_3647_seed_9886_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04290453866124153, + "vlm_logprob": -0.04290453866124153, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.04290453866124153, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rat-tail eel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eel's body coiled in dark, stagnant swamp water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, forked black tongue spring from the eel's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by high-contrast comic book ink and sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle medium shot with the eel centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08143234252929688 + ], + "opposite_logprob_tokens": [ + -2.581432342529297 + ], + "logprob_reward": -0.08143234252929688, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08143234252929688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06777598708868027, + 0.03937463462352753, + 0.04790754243731499, + 0.16379383206367493, + 0.170095294713974, + 0.1360437422990799, + 0.149961918592453, + 0.0788203626871109, + 0.08171842992305756, + 0.06450823694467545 + ], + "credit_map": [ + 0.0061551230028271675, + 0.007853091694414616, + 0.010513292625546455, + 0.006536803208291531, + 0.012873999774456024, + 0.011870238929986954, + 0.004591133911162615, + 0.0073823099955916405, + 0.003181957174092531, + 0.003098729532212019, + 0.008683057501912117, + 0.00370507943443954, + 0.005119657143950462, + 0.006520725321024656, + 0.0027867702301591635, + 0.006278662011027336, + 0.004408583045005798, + 0.006495058536529541, + 0.009829944930970669, + 0.0039009740576148033, + 0.007128970697522163, + 0.006296509876847267, + 0.003913505468517542, + 0.005933993961662054, + 0.008311170153319836, + 0.004422072321176529, + 0.06077060475945473, + 0.006761476863175631, + 0.0042415643110871315, + 0.06734661012887955, + 0.004829966463148594, + 0.00711038475856185, + 0.007708707358688116, + 0.0068469843827188015, + 0.05530550703406334, + 0.007972469553351402, + 0.006053600925952196, + 0.067377969622612, + 0.006058715749531984, + 0.012771336361765862, + 0.005342983640730381, + 0.00471913255751133, + 0.01193649135529995, + 0.00739644467830658, + 0.00885202456265688, + 0.04344052076339722, + 0.016313645988702774, + 0.038042500615119934, + 0.010480680502951145, + 0.009560597129166126, + 0.0303863026201725, + 0.010223089717328548, + 0.012132474221289158, + 0.041657522320747375, + 0.021036848425865173, + 0.0144844064489007, + 0.0077686314471066, + 0.005997755099087954, + 0.009726490825414658, + 0.006509227212518454, + 0.007821273058652878, + 0.011637983843684196, + 0.01911080628633499, + 0.010248194448649883, + 0.007073067128658295, + 0.008471288718283176, + 0.014775101095438004, + 0.004518762230873108, + 0.01130165345966816, + 0.01035979948937893, + 0.01417566929012537, + 0.011043093167245388, + 0.00930024217814207, + 0.005367955658584833, + 0.011171081103384495, + 0.00921740010380745, + 0.005625630728900433, + 0.006384215317666531, + 0.009541235864162445, + 0.007900473661720753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 6 + ], + "selected_reward": -0.08143234252929688 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rat-tail eel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eel's body coiled in dark, stagnant swamp water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, forked black tongue spring from the eel's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by high-contrast comic book ink and sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle medium shot with the eel centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373602032661438 + ], + "opposite_logprob_tokens": [ + -3.408735990524292 + ], + "logprob_reward": -0.03373602032661438, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373602032661438, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1011284664273262, + 0.03990361839532852, + 0.06816058605909348, + 0.02432057447731495, + 0.23693779110908508, + 0.2115059494972229, + 0.18515925109386444, + 0.07064823806285858, + 0.028875192627310753, + 0.033360302448272705 + ], + "credit_map": [ + 0.008186939172446728, + 0.011746379546821117, + 0.017965035513043404, + 0.0062355645932257175, + 0.016381459310650826, + 0.023216458037495613, + 0.007118146866559982, + 0.010278473608195782, + 0.0032118498347699642, + 0.002915615914389491, + 0.01001774612814188, + 0.0027359644882380962, + 0.004429513122886419, + 0.008035050705075264, + 0.0025618411600589752, + 0.005996038671582937, + 0.005994260776787996, + 0.0062279971316456795, + 0.014557288028299809, + 0.005333437584340572, + 0.008667473681271076, + 0.012011720798909664, + 0.005686605349183083, + 0.009681805036962032, + 0.0023688520304858685, + 0.0021923331078141928, + 0.004258436616510153, + 0.0022196518257260323, + 0.002518504625186324, + 0.004108614753931761, + 0.0037663436960428953, + 0.002887838752940297, + 0.009131604805588722, + 0.007885776460170746, + 0.08030646294355392, + 0.0094484006986022, + 0.0070414491929113865, + 0.09845646470785141, + 0.00802387110888958, + 0.01664375327527523, + 0.008575938642024994, + 0.005908534862101078, + 0.01724848710000515, + 0.011600974015891552, + 0.010789954103529453, + 0.07233273983001709, + 0.019696613773703575, + 0.06535272300243378, + 0.012986262328922749, + 0.01025647483766079, + 0.036740824580192566, + 0.012048997916281223, + 0.016858385875821114, + 0.048317693173885345, + 0.030176224187016487, + 0.017774375155568123, + 0.006305722054094076, + 0.00818394124507904, + 0.009911620058119297, + 0.006619304418563843, + 0.008071748539805412, + 0.010456855408847332, + 0.01143381092697382, + 0.009665233083069324, + 0.0025668719317764044, + 0.003524412401020527, + 0.006419655866920948, + 0.0020285588689148426, + 0.0032301428727805614, + 0.0030533194076269865, + 0.0051682027988135815, + 0.0028840298764407635, + 0.005087168421596289, + 0.002988085150718689, + 0.006290323566645384, + 0.004810837563127279, + 0.002561135683208704, + 0.002757670357823372, + 0.0039528971537947655, + 0.004912183620035648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.03373602032661438 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rat-tail eel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eel's body coiled in dark, stagnant swamp water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, forked black tongue spring from the eel's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by high-contrast comic book ink and sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle medium shot with the eel centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819327473640442 + ], + "opposite_logprob_tokens": [ + -4.018193244934082 + ], + "logprob_reward": -0.01819327473640442, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819327473640442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0406971238553524, + 0.03976049646735191, + 0.02644207887351513, + 0.15571865439414978, + 0.35989853739738464, + 0.15298104286193848, + 0.1123417317867279, + 0.059900637716054916, + 0.022902851924300194, + 0.029356880113482475 + ], + "credit_map": [ + 0.004417569376528263, + 0.003115737810730934, + 0.008016915991902351, + 0.0028016986325383186, + 0.0051980274729430676, + 0.008758215233683586, + 0.002468705642968416, + 0.005920248571783304, + 0.0024837690871208906, + 0.0024940052535384893, + 0.009915407747030258, + 0.002745494246482849, + 0.0034650794696062803, + 0.008689909242093563, + 0.0029355993028730154, + 0.00703122653067112, + 0.002223377814516425, + 0.002960801823064685, + 0.004254327155649662, + 0.0026674852706491947, + 0.004892116412520409, + 0.0037972196005284786, + 0.0022441197652369738, + 0.0034026263747364283, + 0.010148163884878159, + 0.0062873829156160355, + 0.04907582327723503, + 0.007391304709017277, + 0.0053557842038571835, + 0.05568660423159599, + 0.01179691031575203, + 0.00997666735202074, + 0.012928007170557976, + 0.011918448843061924, + 0.12408389896154404, + 0.012959163635969162, + 0.008614007383584976, + 0.15323570370674133, + 0.00903849396854639, + 0.027120791375637054, + 0.005290903151035309, + 0.004426885861903429, + 0.012239976786077023, + 0.008311192505061626, + 0.009553895331919193, + 0.05124341696500778, + 0.01608521118760109, + 0.045829568058252335, + 0.008704354986548424, + 0.006755670998245478, + 0.018846239894628525, + 0.008410572074353695, + 0.011078927665948868, + 0.02542012929916382, + 0.020477745682001114, + 0.012648074887692928, + 0.0037198022473603487, + 0.003875110764056444, + 0.007696559652686119, + 0.0038609381299465895, + 0.0063528623431921005, + 0.008763795718550682, + 0.016348116099834442, + 0.00928344577550888, + 0.002119004260748625, + 0.0018890084465965629, + 0.004365431610494852, + 0.0012808656319975853, + 0.0034203657414764166, + 0.003137701191008091, + 0.004205835983157158, + 0.002484638476744294, + 0.003714597085490823, + 0.002659096848219633, + 0.0051499842666089535, + 0.004049214068800211, + 0.0019677181262522936, + 0.002959616482257843, + 0.003731679869815707, + 0.005124971736222506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.01819327473640442 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rat-tail eel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eel's body coiled in dark, stagnant swamp water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, forked black tongue spring from the eel's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by high-contrast comic book ink and sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle medium shot with the eel centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038099367171525955 + ], + "opposite_logprob_tokens": [ + -3.2880992889404297 + ], + "logprob_reward": -0.038099367171525955, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038099367171525955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061385251581668854, + 0.08800151199102402, + 0.035737939178943634, + 0.17340326309204102, + 0.3511033356189728, + 0.11252011358737946, + 0.05926281213760376, + 0.05469927936792374, + 0.03363174572587013, + 0.030254783108830452 + ], + "credit_map": [ + 0.005567075219005346, + 0.004860640969127417, + 0.013359876349568367, + 0.004686801228672266, + 0.00788998045027256, + 0.014665646478533745, + 0.0027250719722360373, + 0.007630159147083759, + 0.0058045401237905025, + 0.005089518614113331, + 0.02797909453511238, + 0.0050491709262132645, + 0.006217800080776215, + 0.018126918002963066, + 0.003782689804211259, + 0.015951773151755333, + 0.002761285752058029, + 0.002514695981517434, + 0.0074543217197060585, + 0.001885190256871283, + 0.00454491563141346, + 0.006949097383767366, + 0.0032119823154062033, + 0.006416449788957834, + 0.00890688132494688, + 0.004538990557193756, + 0.06446290761232376, + 0.007132166530936956, + 0.004640892148017883, + 0.07110672444105148, + 0.005474892444908619, + 0.007139808498322964, + 0.01265370100736618, + 0.011155934073030949, + 0.11868158727884293, + 0.01201937161386013, + 0.010618533007800579, + 0.1450541466474533, + 0.016381055116653442, + 0.024538995698094368, + 0.004312341101467609, + 0.004192471504211426, + 0.00989207997918129, + 0.006394393742084503, + 0.00860269833356142, + 0.03491999953985214, + 0.013572433032095432, + 0.030633702874183655, + 0.004416378680616617, + 0.0033240551128983498, + 0.010775210335850716, + 0.003994644619524479, + 0.005355011206120253, + 0.0144371148198843, + 0.010516777634620667, + 0.00644361786544323, + 0.005570902023464441, + 0.00698983995243907, + 0.007097139954566956, + 0.006039868574589491, + 0.004234112333506346, + 0.009463667869567871, + 0.007526075001806021, + 0.0077776722609996796, + 0.0041886731050908566, + 0.003347985912114382, + 0.0054251146502792835, + 0.0023094264324754477, + 0.004739017691463232, + 0.0036167525686323643, + 0.007017165422439575, + 0.002987609012052417, + 0.00394819863140583, + 0.0031087524257600307, + 0.006385992746800184, + 0.004221736919134855, + 0.0019523761002346873, + 0.0026048084255307913, + 0.0031397452112287283, + 0.004893172532320023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.038099367171525955 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rat-tail eel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eel's body coiled in dark, stagnant swamp water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, forked black tongue spring from the eel's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by high-contrast comic book ink and sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle medium shot with the eel centered?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306168854236603 + ], + "opposite_logprob_tokens": [ + -3.1680617332458496 + ], + "logprob_reward": -0.04306168854236603, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306168854236603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04664440453052521, + 0.04866348206996918, + 0.04806782305240631, + 0.10559400916099548, + 0.1701848804950714, + 0.0461304634809494, + 0.111234150826931, + 0.0657067745923996, + 0.040390923619270325, + 0.3173830807209015 + ], + "credit_map": [ + 0.003646472468972206, + 0.005740866996347904, + 0.007053009700030088, + 0.0034820064902305603, + 0.010398918762803078, + 0.005844056606292725, + 0.003914500121027231, + 0.006564573384821415, + 0.00348962121643126, + 0.0027369356248527765, + 0.013276124373078346, + 0.0035763217601925135, + 0.003011820837855339, + 0.009191824123263359, + 0.0035254445392638445, + 0.009855391457676888, + 0.003275488503277302, + 0.011563454754650593, + 0.00692397728562355, + 0.0038528454024344683, + 0.005020327400416136, + 0.007416427135467529, + 0.00350342970341444, + 0.006511871702969074, + 0.005691329017281532, + 0.004195149056613445, + 0.037360064685344696, + 0.004908417817205191, + 0.0033873433712869883, + 0.04071183502674103, + 0.004054130986332893, + 0.005285739433020353, + 0.007175251841545105, + 0.005161653738468885, + 0.05545934662222862, + 0.006406048778444529, + 0.005357971414923668, + 0.06884302198886871, + 0.008420752361416817, + 0.013360831886529922, + 0.002934990217909217, + 0.0025807353667914867, + 0.005115612875670195, + 0.004406246822327375, + 0.003521906677633524, + 0.012316139414906502, + 0.006061966996639967, + 0.009192866273224354, + 0.008116713725030422, + 0.006063522771000862, + 0.017431726679205894, + 0.008517559617757797, + 0.010194442234933376, + 0.02444472722709179, + 0.022105081006884575, + 0.014360370114445686, + 0.005972724407911301, + 0.005314739421010017, + 0.00857956800609827, + 0.005533942021429539, + 0.008030440658330917, + 0.007896932773292065, + 0.012082330882549286, + 0.012296101078391075, + 0.004166631959378719, + 0.004235792439430952, + 0.006729478482156992, + 0.003069604979828, + 0.0046397107653319836, + 0.007557886186987162, + 0.006610192358493805, + 0.0033816250506788492, + 0.043257664889097214, + 0.027879148721694946, + 0.06967252492904663, + 0.0390738807618618, + 0.01924149878323078, + 0.02849535085260868, + 0.03486492484807968, + 0.05489807948470116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 6 + ], + "selected_reward": -0.04306168854236603 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_3/index_tensor([2635, 2635])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_3/index_tensor([2635, 2635])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c9af981201e65d50ae7f81796dca53c4de5a4ec3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_3/index_tensor([2635, 2635])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up static shot captures a scene of intense sensual anticipation, rendered in dramatic black and white contrast lighting. A decadent, multi-layered chocolate fudge cake rests on a pristine white ceramic plate. Warm golden light from a single, sharp, off-camera source sculpts the dessert's shape, creating shimmering highlights on the smooth surface of the dark chocolate ganache and casting deep, velvety shadows that accentuate its rich texture. A figure, partially visible, slowly and reverently leans toward the cake. Their eyes are gently closed, long lashes resting on their cheeks, their full attention focused on the experience about to begin. A subtle, almost imperceptible tremor crosses their slightly parted lips, a physical expression of pure, unadulterated desire. The camera unwaveringly captures this delicately tense moment, just before the bite, creating an atmosphere of luxurious, solitary indulgence that is remarkably intimate and focused.", + "seed": 5904, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_3/index_2635_seed_5904_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03644125834107399, + "vlm_logprob": -0.03644125834107399, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.03644125834107399, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a multi-layered chocolate fudge cake on a white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting rendered in dramatic black and white contrast with warm golden highlights on the chocolate ganache?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure partially visible, leaning toward the cake with eyes gently closed and long lashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's mouth slightly parted with a subtle tremor visible on the lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, maintaining an unwavering frame on the scene without movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03092469647526741 + ], + "opposite_logprob_tokens": [ + -3.7809247970581055 + ], + "logprob_reward": -0.03092469647526741, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.03092469647526741, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15129666030406952, + 0.1607072651386261, + 0.10117892175912857, + 0.034166693687438965, + 0.0744837298989296, + 0.03219054266810417, + 0.09464043378829956, + 0.04280826076865196, + 0.053072042763233185, + 0.25545549392700195 + ], + "credit_map": [ + 0.014088181778788567, + 0.005281467456370592, + 0.013225956819951534, + 0.008996273390948772, + 0.04889582842588425, + 0.024844730272889137, + 0.024776717647910118, + 0.011187518015503883, + 0.013919325545430183, + 0.004343101754784584, + 0.011338611133396626, + 0.005467367824167013, + 0.06605752557516098, + 0.026879554614424706, + 0.02639586851000786, + 0.006305892486125231, + 0.01726653426885605, + 0.003461497137323022, + 0.03386303409934044, + 0.005298221483826637, + 0.009173226542770863, + 0.007676335982978344, + 0.020341871306300163, + 0.004098204895853996, + 0.004782015457749367, + 0.0023675651755183935, + 0.010486414656043053, + 0.0026761218905448914, + 0.00431416230276227, + 0.00322248344309628, + 0.00415527680888772, + 0.0021626546513289213, + 0.007416386157274246, + 0.003089139936491847, + 0.004928901791572571, + 0.003401785157620907, + 0.02691541612148285, + 0.012933630496263504, + 0.011973898857831955, + 0.0038245730102062225, + 0.0023603495210409164, + 0.0020186041947454214, + 0.002700023353099823, + 0.003199133090674877, + 0.007156629581004381, + 0.005385229364037514, + 0.006771324202418327, + 0.0025992474984377623, + 0.008645628578960896, + 0.006886543706059456, + 0.006547989789396524, + 0.008009822107851505, + 0.028467973694205284, + 0.014948620460927486, + 0.014774736948311329, + 0.006359116639941931, + 0.0027488875202834606, + 0.0034396005794405937, + 0.002271847566589713, + 0.004052659496665001, + 0.009951799176633358, + 0.006922584492713213, + 0.009764119051396847, + 0.0036567600909620523, + 0.0032087964937090874, + 0.007415181025862694, + 0.003608634462580085, + 0.00432051345705986, + 0.01220187358558178, + 0.010395055636763573, + 0.00834723375737667, + 0.003574755508452654, + 0.01465110294520855, + 0.021320881322026253, + 0.015956999734044075, + 0.02555128186941147, + 0.0776931643486023, + 0.029651295393705368, + 0.05191931873559952, + 0.018711460754275322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.03092469647526741 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a multi-layered chocolate fudge cake on a white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting rendered in dramatic black and white contrast with warm golden highlights on the chocolate ganache?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure partially visible, leaning toward the cake with eyes gently closed and long lashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's mouth slightly parted with a subtle tremor visible on the lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, maintaining an unwavering frame on the scene without movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012565961107611656 + ], + "opposite_logprob_tokens": [ + -4.387566089630127 + ], + "logprob_reward": -0.012565961107611656, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012565961107611656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16233213245868683, + 0.055952444672584534, + 0.07325039058923721, + 0.057634927332401276, + 0.1865755170583725, + 0.06893005222082138, + 0.05132627487182617, + 0.22775118052959442, + 0.049734339118003845, + 0.06651273369789124 + ], + "credit_map": [ + 0.010937314480543137, + 0.012216811068356037, + 0.013566993176937103, + 0.011806941591203213, + 0.05169791728258133, + 0.02453545108437538, + 0.02751629427075386, + 0.010054416954517365, + 0.006746744737029076, + 0.004089463967829943, + 0.006445055827498436, + 0.0046899509616196156, + 0.012515839189291, + 0.0094264205545187, + 0.008267453871667385, + 0.003771514166146517, + 0.008062458597123623, + 0.003479506354779005, + 0.013765929266810417, + 0.004133481066673994, + 0.017861293628811836, + 0.00832801777869463, + 0.014303374104201794, + 0.0033163223415613174, + 0.009042997844517231, + 0.0033051592763513327, + 0.01667468249797821, + 0.003286505350843072, + 0.00930324662476778, + 0.005804504733532667, + 0.007193656172603369, + 0.003024173667654395, + 0.02030888944864273, + 0.0076997121796011925, + 0.010244005359709263, + 0.007242766208946705, + 0.07224138081073761, + 0.029794547706842422, + 0.030705014243721962, + 0.00833920855075121, + 0.012841988354921341, + 0.011213576421141624, + 0.007885624654591084, + 0.005813192576169968, + 0.008893593214452267, + 0.0075370678678154945, + 0.00924499612301588, + 0.005500019993633032, + 0.003865013597533107, + 0.0035709720104932785, + 0.0038219799753278494, + 0.004096609074622393, + 0.01059105433523655, + 0.010580383241176605, + 0.008427426218986511, + 0.006372836418449879, + 0.021388471126556396, + 0.0279545821249485, + 0.030616750940680504, + 0.0181291364133358, + 0.05334800109267235, + 0.0300630833953619, + 0.0321381650865078, + 0.014112981036305428, + 0.004281500354409218, + 0.010617709718644619, + 0.0026286214124411345, + 0.0037513067945837975, + 0.011517317034304142, + 0.0047324178740382195, + 0.008598017506301403, + 0.0036074472591280937, + 0.009170710109174252, + 0.007207209710031748, + 0.0038140141405165195, + 0.0036455870140343904, + 0.019281141459941864, + 0.007913478650152683, + 0.010506902821362019, + 0.004973687697201967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.012565961107611656 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a multi-layered chocolate fudge cake on a white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting rendered in dramatic black and white contrast with warm golden highlights on the chocolate ganache?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure partially visible, leaning toward the cake with eyes gently closed and long lashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's mouth slightly parted with a subtle tremor visible on the lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, maintaining an unwavering frame on the scene without movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107603870332241 + ], + "opposite_logprob_tokens": [ + -4.511075973510742 + ], + "logprob_reward": -0.01107603870332241, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107603870332241, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08869829773902893, + 0.24059399962425232, + 0.27909642457962036, + 0.05115838721394539, + 0.028366615995764732, + 0.026471655815839767, + 0.068710096180439, + 0.16045288741588593, + 0.021541856229305267, + 0.034909702837467194 + ], + "credit_map": [ + 0.013358286581933498, + 0.010887817479670048, + 0.008541617542505264, + 0.0068227206356823444, + 0.01630811206996441, + 0.010808170773088932, + 0.01289895549416542, + 0.009072624146938324, + 0.01824292726814747, + 0.00913977064192295, + 0.018996113911271095, + 0.008406398817896843, + 0.0967133566737175, + 0.03865256905555725, + 0.04120814800262451, + 0.009234718047082424, + 0.03326300531625748, + 0.02468920312821865, + 0.07494273781776428, + 0.018684811890125275, + 0.035932302474975586, + 0.03958798199892044, + 0.037978608161211014, + 0.014017780311405659, + 0.007720147259533405, + 0.005324372090399265, + 0.012378867715597153, + 0.005054940935224295, + 0.007735412567853928, + 0.005049099214375019, + 0.005033482331782579, + 0.0028620644006878138, + 0.0034755549859255552, + 0.001891626976430416, + 0.0021439241245388985, + 0.0017864973051473498, + 0.007978061214089394, + 0.005048932041972876, + 0.004273212049156427, + 0.0017688083462417126, + 0.003452408593147993, + 0.003958201967179775, + 0.0022756776306778193, + 0.0025627436116337776, + 0.004816965665668249, + 0.0033182392362505198, + 0.0036662144120782614, + 0.0024212044663727283, + 0.005606975872069597, + 0.003560819197446108, + 0.003024838864803314, + 0.0040945205837488174, + 0.025787286460399628, + 0.01103949174284935, + 0.01254022866487503, + 0.0030559387523680925, + 0.014390409924089909, + 0.011827386915683746, + 0.00824880413711071, + 0.007429691031575203, + 0.05780395120382309, + 0.026092058047652245, + 0.028787845745682716, + 0.005872744135558605, + 0.0020089014433324337, + 0.005263917613774538, + 0.00169308518525213, + 0.0026287094224244356, + 0.0025843833573162556, + 0.0030363579280674458, + 0.0027375700883567333, + 0.0015889302594587207, + 0.002348787384107709, + 0.00363293313421309, + 0.0018226804677397013, + 0.003376643406227231, + 0.009550183080136776, + 0.003832293674349785, + 0.0077134049497544765, + 0.0026327772065997124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 7 + ], + "selected_reward": -0.01107603870332241 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a multi-layered chocolate fudge cake on a white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting rendered in dramatic black and white contrast with warm golden highlights on the chocolate ganache?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure partially visible, leaning toward the cake with eyes gently closed and long lashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's mouth slightly parted with a subtle tremor visible on the lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, maintaining an unwavering frame on the scene without movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866215959191322 + ], + "opposite_logprob_tokens": [ + -3.0486621856689453 + ], + "logprob_reward": -0.04866215959191322, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866215959191322, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13432148098945618, + 0.3078145384788513, + 0.1169523298740387, + 0.06513886153697968, + 0.07154340296983719, + 0.033909354358911514, + 0.09164674580097198, + 0.08054326474666595, + 0.06535442173480988, + 0.03277551010251045 + ], + "credit_map": [ + 0.015253067016601562, + 0.008699728175997734, + 0.0091409832239151, + 0.0061645349487662315, + 0.043386176228523254, + 0.022061843425035477, + 0.020539037883281708, + 0.009076119400560856, + 0.01974586769938469, + 0.00986700039356947, + 0.018583286553621292, + 0.011421812698245049, + 0.12688128650188446, + 0.04804075136780739, + 0.06214214488863945, + 0.011132454499602318, + 0.018338486552238464, + 0.010462257079780102, + 0.03161780536174774, + 0.007056175731122494, + 0.01213056966662407, + 0.013925286009907722, + 0.017244676128029823, + 0.00617708545178175, + 0.009650970809161663, + 0.006202782969921827, + 0.018828636035323143, + 0.005568067077547312, + 0.007727834861725569, + 0.005266609601676464, + 0.008225707337260246, + 0.0036682540085166693, + 0.007534284144639969, + 0.0028344965539872646, + 0.00488164322450757, + 0.0032551439944654703, + 0.02480844035744667, + 0.014133892953395844, + 0.010778033174574375, + 0.003317472292110324, + 0.005196701735258102, + 0.004937435034662485, + 0.004420633893460035, + 0.0028877160511910915, + 0.004708862863481045, + 0.004214930348098278, + 0.0044682081788778305, + 0.0030748709104955196, + 0.00684274360537529, + 0.006014048587530851, + 0.006098318845033646, + 0.00638276943936944, + 0.027803979814052582, + 0.017113592475652695, + 0.015602011233568192, + 0.005789290647953749, + 0.008361967280507088, + 0.004701904486864805, + 0.006021855864673853, + 0.0054115657694637775, + 0.022783171385526657, + 0.013258913531899452, + 0.014117367565631866, + 0.005886530969291925, + 0.005544107407331467, + 0.012533685192465782, + 0.0045836991630494595, + 0.005495765246450901, + 0.015104795806109905, + 0.007135716266930103, + 0.010349406860768795, + 0.004607247654348612, + 0.0032136114314198494, + 0.00449534272775054, + 0.0017546593444421887, + 0.002481399103999138, + 0.00876806303858757, + 0.003712868085131049, + 0.006220421753823757, + 0.0021291489247232676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.04866215959191322 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a multi-layered chocolate fudge cake on a white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting rendered in dramatic black and white contrast with warm golden highlights on the chocolate ganache?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure partially visible, leaning toward the cake with eyes gently closed and long lashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's mouth slightly parted with a subtle tremor visible on the lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, maintaining an unwavering frame on the scene without movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897743582725525 + ], + "opposite_logprob_tokens": [ + -2.578977346420288 + ], + "logprob_reward": -0.07897743582725525, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897743582725525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10011989623308182, + 0.10889586061239243, + 0.14117708802223206, + 0.04428839311003685, + 0.1973467916250229, + 0.04223984107375145, + 0.12968230247497559, + 0.07544201612472534, + 0.1156938299536705, + 0.04511403292417526 + ], + "credit_map": [ + 0.00969863310456276, + 0.006757794879376888, + 0.004488160368055105, + 0.004731628578156233, + 0.03405462205410004, + 0.016615580767393112, + 0.017454566434025764, + 0.006318907719105482, + 0.008539896458387375, + 0.0032085750717669725, + 0.0075743598863482475, + 0.00434854393824935, + 0.043251823633909225, + 0.016697945073246956, + 0.020733216777443886, + 0.004541496746242046, + 0.02261365018785, + 0.0045981211587786674, + 0.05282435938715935, + 0.0073915328830480576, + 0.022544607520103455, + 0.008943221531808376, + 0.017046062275767326, + 0.005215532146394253, + 0.006947618443518877, + 0.0024252349976450205, + 0.013117477297782898, + 0.0025799276772886515, + 0.00677353423088789, + 0.00436452915892005, + 0.004622879903763533, + 0.003457191865891218, + 0.01804543100297451, + 0.007018551230430603, + 0.013206945732235909, + 0.008216885849833488, + 0.07088728994131088, + 0.03873801603913307, + 0.030972598120570183, + 0.01026107370853424, + 0.005619582254439592, + 0.004650096409022808, + 0.0031822046730667353, + 0.0028844974003732204, + 0.009429323486983776, + 0.006595236714929342, + 0.006662355735898018, + 0.0032165406737476587, + 0.012718490324914455, + 0.006294190883636475, + 0.006753020454198122, + 0.009135613217949867, + 0.04534049332141876, + 0.02143443562090397, + 0.02144899033010006, + 0.0065570734441280365, + 0.007996910251677036, + 0.009267161600291729, + 0.007983717136085033, + 0.005236506927758455, + 0.01666751503944397, + 0.009669698774814606, + 0.011813337914645672, + 0.00680716335773468, + 0.011011561378836632, + 0.016949711367487907, + 0.0092901811003685, + 0.015467318706214428, + 0.01848277635872364, + 0.018656618893146515, + 0.015325432643294334, + 0.01051024254411459, + 0.007745038252323866, + 0.004719774704426527, + 0.004060755483806133, + 0.004091240931302309, + 0.007994333282113075, + 0.005565056577324867, + 0.006625639274716377, + 0.004312191624194384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 6 + ], + "selected_reward": -0.07897743582725525 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot captures a scene of intense sensual anticipation, rendered in dramatic black and white contrast lighting. A decadent, multi-layered chocolate fudge cake rests on a pristine white ceramic plate. Warm golden light from a single, sharp, off-camera source sculpts the dessert's shape, creating shimmering highlights on the smooth surface of the dark chocolate ganache and casting deep, velvety shadows that accentuate its rich texture. A figure, partially visible, slowly and reverently leans toward the cake. Their eyes are gently closed, long lashes resting on their cheeks, their full attention focused on the experience about to begin. A subtle, almost imperceptible tremor crosses their slightly parted lips, a physical expression of pure, unadulterated desire. The camera unwaveringly captures this delicately tense moment, just before the bite, creating an atmosphere of luxurious, solitary indulgence that is remarkably intimate and focused.", + "seed": 5905, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_3/index_2635_seed_5905_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03625199794769287, + "vlm_logprob": -0.03625199794769287, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.03625199794769287, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a multi-layered chocolate fudge cake on a white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting rendered in dramatic black and white contrast with warm golden highlights on the chocolate ganache?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure partially visible, leaning toward the cake with eyes gently closed and long lashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's mouth slightly parted with a subtle tremor visible on the lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, maintaining an unwavering frame on the scene without movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0413125604391098 + ], + "opposite_logprob_tokens": [ + -3.2913124561309814 + ], + "logprob_reward": -0.0413125604391098, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0413125604391098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6994761824607849, + 0.06880465894937515, + 0.02359587699174881, + 0.007772472221404314, + 0.02353139966726303, + 0.016652001067996025, + 0.01115754060447216, + 0.04789077490568161, + 0.01304864976555109, + 0.08807045966386795 + ], + "credit_map": [ + 0.16412769258022308, + 0.07005733251571655, + 0.1008574366569519, + 0.045463550835847855, + 0.052232977002859116, + 0.12384988367557526, + 0.09483160078525543, + 0.04805564135313034, + 0.011219113133847713, + 0.016143295913934708, + 0.005270266905426979, + 0.005202651955187321, + 0.009617471136152744, + 0.008357277140021324, + 0.008835184387862682, + 0.0041593932546675205, + 0.0029186843894422054, + 0.0030793691985309124, + 0.001976432278752327, + 0.003088693367317319, + 0.0018028480699285865, + 0.002541639842092991, + 0.005841930862516165, + 0.002346275607123971, + 0.001311455387622118, + 0.0017820022767409682, + 0.0007781842141412199, + 0.0009644777164794505, + 0.0005045965081080794, + 0.0006641495274379849, + 0.0010345210321247578, + 0.0007330847438424826, + 0.0031378089915961027, + 0.004594973288476467, + 0.002844379283487797, + 0.002524979878216982, + 0.0016947821713984013, + 0.003684528172016144, + 0.0030367334838956594, + 0.002013212302699685, + 0.0029684221372008324, + 0.0022645152639597654, + 0.0010646979790180922, + 0.0016084101516753435, + 0.0017927867593243718, + 0.0026949888560920954, + 0.0023712427355349064, + 0.0018869360210373998, + 0.0017202842282131314, + 0.0025583768729120493, + 0.0010416366858407855, + 0.0010864484356716275, + 0.0007495004101656377, + 0.0014708879170939326, + 0.001611152314580977, + 0.0009192518191412091, + 0.008973856456577778, + 0.01177028939127922, + 0.0044839209876954556, + 0.0036719900090247393, + 0.0034777980763465166, + 0.005118887405842543, + 0.005969601217657328, + 0.004424428101629019, + 0.002185015007853508, + 0.0018483888125047088, + 0.0012504853075370193, + 0.001018355949781835, + 0.0011372752487659454, + 0.002143991645425558, + 0.002227852586656809, + 0.0012372838100418448, + 0.014060177840292454, + 0.00930461473762989, + 0.00793385598808527, + 0.006845688447356224, + 0.006487411446869373, + 0.01227360125631094, + 0.019233474507927895, + 0.011931625194847584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.0413125604391098 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a multi-layered chocolate fudge cake on a white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting rendered in dramatic black and white contrast with warm golden highlights on the chocolate ganache?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure partially visible, leaning toward the cake with eyes gently closed and long lashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's mouth slightly parted with a subtle tremor visible on the lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, maintaining an unwavering frame on the scene without movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008660022169351578 + ], + "opposite_logprob_tokens": [ + -4.758659839630127 + ], + "logprob_reward": -0.008660022169351578, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008660022169351578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7415127754211426, + 0.05419154092669487, + 0.012099008075892925, + 0.029938600957393646, + 0.012642835266888142, + 0.012837205082178116, + 0.010783517733216286, + 0.088435597717762, + 0.012248593382537365, + 0.025310339406132698 + ], + "credit_map": [ + 0.17454035580158234, + 0.07378948479890823, + 0.10217361897230148, + 0.04530060291290283, + 0.054495006799697876, + 0.1403067409992218, + 0.09893091768026352, + 0.05197599157691002, + 0.008815135806798935, + 0.012561244890093803, + 0.00719251437112689, + 0.004139163997024298, + 0.0036163099575787783, + 0.006491396576166153, + 0.0065181259997189045, + 0.004857648629695177, + 0.0020770838018506765, + 0.0018916827393695712, + 0.000900204642675817, + 0.001015234855003655, + 0.0007725415052846074, + 0.001614163164049387, + 0.0019919148180633783, + 0.0018361835973337293, + 0.0070266216062009335, + 0.006507175974547863, + 0.0020230587106198072, + 0.002049816306680441, + 0.0014611807418987155, + 0.0026612901128828526, + 0.0033408317249268293, + 0.004868625197559595, + 0.0024082669988274574, + 0.0020962872076779604, + 0.0011271571274846792, + 0.000996281043626368, + 0.0007270858041010797, + 0.002104595536366105, + 0.0015702395467087626, + 0.0016129215946421027, + 0.0021019349806010723, + 0.001866063685156405, + 0.0013853125274181366, + 0.0011302554048597813, + 0.0014722300693392754, + 0.0017237067222595215, + 0.0017679191660135984, + 0.0013897822936996818, + 0.0014585839817300439, + 0.0021128898952156305, + 0.000957963231485337, + 0.0012888937490060925, + 0.0007121514063328505, + 0.0014657016145065427, + 0.0015272315358743072, + 0.0012601027265191078, + 0.01465769112110138, + 0.012485310435295105, + 0.013699119910597801, + 0.005743182729929686, + 0.005317844916135073, + 0.016013316810131073, + 0.009031875059008598, + 0.011487262323498726, + 0.0029935254715383053, + 0.0018432666547596455, + 0.0017557013779878616, + 0.0009125375654548407, + 0.0007823288906365633, + 0.0014656769344583154, + 0.0014695466961711645, + 0.0010260103736072779, + 0.004409098066389561, + 0.002572399564087391, + 0.0019889862742275, + 0.0017383679514750838, + 0.0018707585986703634, + 0.0039348481222987175, + 0.005276606418192387, + 0.0035192754585295916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.008660022169351578 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a multi-layered chocolate fudge cake on a white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting rendered in dramatic black and white contrast with warm golden highlights on the chocolate ganache?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure partially visible, leaning toward the cake with eyes gently closed and long lashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's mouth slightly parted with a subtle tremor visible on the lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, maintaining an unwavering frame on the scene without movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018188826739788055 + ], + "opposite_logprob_tokens": [ + -4.018188953399658 + ], + "logprob_reward": -0.018188826739788055, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018188826739788055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6160140633583069, + 0.09630651772022247, + 0.040911074727773666, + 0.10908528417348862, + 0.035088375210762024, + 0.03419818356633186, + 0.009168029762804508, + 0.035177577286958694, + 0.011482753790915012, + 0.012568093836307526 + ], + "credit_map": [ + 0.1420287936925888, + 0.06181628257036209, + 0.09098094701766968, + 0.03915159031748772, + 0.04561951011419296, + 0.11223316937685013, + 0.08100731670856476, + 0.04317649081349373, + 0.01916997693479061, + 0.024348439648747444, + 0.009574628435075283, + 0.006481483578681946, + 0.009079092182219028, + 0.008440999314188957, + 0.011561542749404907, + 0.007650356739759445, + 0.005727056879550219, + 0.010992967523634434, + 0.003312997054308653, + 0.002974075498059392, + 0.003579509910196066, + 0.004256613552570343, + 0.006501607596874237, + 0.003566247643902898, + 0.019859351217746735, + 0.025410396978259087, + 0.010742141865193844, + 0.00949816219508648, + 0.00690787797793746, + 0.009404804557561874, + 0.014367053285241127, + 0.012895493768155575, + 0.005397172644734383, + 0.008865904994308949, + 0.0034276496153324842, + 0.003996267449110746, + 0.002851241733878851, + 0.003394383704289794, + 0.0045430646277964115, + 0.0026126911398023367, + 0.007121659349650145, + 0.009488902054727077, + 0.0024951882660388947, + 0.003143305890262127, + 0.0026022824458777905, + 0.0032973831985145807, + 0.0038368452806025743, + 0.0022126196417957544, + 0.001464382978156209, + 0.00221324828453362, + 0.0010365181369706988, + 0.0007540355436503887, + 0.0006693211616948247, + 0.0009276943746954203, + 0.001351958722807467, + 0.0007508705020882189, + 0.009810259565711021, + 0.012636938132345676, + 0.0023064464330673218, + 0.001370566082186997, + 0.0012622288195416331, + 0.002422860823571682, + 0.003208866575732827, + 0.0021594143472611904, + 0.002531980397179723, + 0.0017521448899060488, + 0.0017679433804005384, + 0.0007176519138738513, + 0.0007483166991733015, + 0.00155899568926543, + 0.0012868510093539953, + 0.001118869986385107, + 0.0016909047262743115, + 0.0031754609663039446, + 0.001491478644311428, + 0.0008480504620820284, + 0.00083022506441921, + 0.0014037054497748613, + 0.0019200444221496582, + 0.0012082249158993363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.018188826739788055 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a multi-layered chocolate fudge cake on a white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting rendered in dramatic black and white contrast with warm golden highlights on the chocolate ganache?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure partially visible, leaning toward the cake with eyes gently closed and long lashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's mouth slightly parted with a subtle tremor visible on the lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, maintaining an unwavering frame on the scene without movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307698830962181 + ], + "opposite_logprob_tokens": [ + -3.168076992034912 + ], + "logprob_reward": -0.04307698830962181, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307698830962181, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46933743357658386, + 0.19747714698314667, + 0.03678848221898079, + 0.09869604557752609, + 0.021132884547114372, + 0.027624821290373802, + 0.01398510206490755, + 0.07282799482345581, + 0.027710378170013428, + 0.03441976383328438 + ], + "credit_map": [ + 0.1069972813129425, + 0.05251162499189377, + 0.07352736592292786, + 0.028416723012924194, + 0.035088930279016495, + 0.08073808997869492, + 0.05880498141050339, + 0.033252421766519547, + 0.03600910305976868, + 0.04929911345243454, + 0.021395258605480194, + 0.016650361940264702, + 0.013776149600744247, + 0.020218374207615852, + 0.027014927938580513, + 0.013113867491483688, + 0.0051083993166685104, + 0.009042193181812763, + 0.0037626472767442465, + 0.0025756503455340862, + 0.0035829159896820784, + 0.005198546219617128, + 0.004938395228236914, + 0.002579732099547982, + 0.022639937698841095, + 0.019554130733013153, + 0.00767778605222702, + 0.010784169659018517, + 0.006201461888849735, + 0.011174660176038742, + 0.01067289337515831, + 0.009991009719669819, + 0.003381371032446623, + 0.004233936779201031, + 0.0018748597940430045, + 0.002719922224059701, + 0.0015523384790867567, + 0.0020292026456445456, + 0.0031865211203694344, + 0.002154733520001173, + 0.005088610574603081, + 0.005349456332623959, + 0.0019461688352748752, + 0.002393800998106599, + 0.002514101332053542, + 0.0033753844909369946, + 0.0043383631855249405, + 0.002618935192003846, + 0.0022113218437880278, + 0.0023237315472215414, + 0.0014170118374750018, + 0.00154306017793715, + 0.000986689468845725, + 0.00173861812800169, + 0.001726592076011002, + 0.002038076752796769, + 0.01815560832619667, + 0.021816274151206017, + 0.007496802136301994, + 0.0032711855601519346, + 0.002815265441313386, + 0.006732357665896416, + 0.0068097361363470554, + 0.005730765406042337, + 0.005352454259991646, + 0.004795022774487734, + 0.0033924782183021307, + 0.0024685466196388006, + 0.001817374606616795, + 0.0030088878702372313, + 0.0035866673570126295, + 0.0032889475114643574, + 0.004754024092108011, + 0.006041705142706633, + 0.0034621392842382193, + 0.002646408975124359, + 0.0024476516991853714, + 0.006027938332408667, + 0.004754500929266214, + 0.004285395611077547 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.04307698830962181 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a multi-layered chocolate fudge cake on a white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting rendered in dramatic black and white contrast with warm golden highlights on the chocolate ganache?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure partially visible, leaning toward the cake with eyes gently closed and long lashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's mouth slightly parted with a subtle tremor visible on the lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, maintaining an unwavering frame on the scene without movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002159208059311 + ], + "opposite_logprob_tokens": [ + -2.695021629333496 + ], + "logprob_reward": -0.07002159208059311, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002159208059311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18645896017551422, + 0.18917514383792877, + 0.07483021914958954, + 0.09824434667825699, + 0.06407996267080307, + 0.04463478550314903, + 0.08048474043607712, + 0.10644517838954926, + 0.0613727942109108, + 0.09427390247583389 + ], + "credit_map": [ + 0.04412216693162918, + 0.025464113801717758, + 0.018431052565574646, + 0.012710358016192913, + 0.017494263127446175, + 0.02969118393957615, + 0.024512622505426407, + 0.014033176936209202, + 0.03428015485405922, + 0.02577534317970276, + 0.01884886808693409, + 0.017291979864239693, + 0.02591337077319622, + 0.025864802300930023, + 0.02451702021062374, + 0.01668359525501728, + 0.014628633856773376, + 0.02087482064962387, + 0.0070536090061068535, + 0.004900597967207432, + 0.005633190274238586, + 0.007264013402163982, + 0.00981694646179676, + 0.004658397752791643, + 0.017019683495163918, + 0.02298067696392536, + 0.00901021622121334, + 0.00946580246090889, + 0.0067534055560827255, + 0.008263285271823406, + 0.013689533807337284, + 0.011061729863286018, + 0.0089645367115736, + 0.01340264081954956, + 0.0044487896375358105, + 0.008663631044328213, + 0.005308606196194887, + 0.008170364424586296, + 0.00925297662615776, + 0.005868409760296345, + 0.007011977955698967, + 0.00850282609462738, + 0.003407401731237769, + 0.004987221676856279, + 0.004180608317255974, + 0.0071645076386630535, + 0.00569734163582325, + 0.003682896262034774, + 0.01715770922601223, + 0.01896020956337452, + 0.008231975138187408, + 0.0075238607823848724, + 0.004946605302393436, + 0.00812666118144989, + 0.00858088955283165, + 0.006956815719604492, + 0.01688465103507042, + 0.01766863279044628, + 0.01295710913836956, + 0.00857709813863039, + 0.011453185230493546, + 0.0118418512865901, + 0.01306786946952343, + 0.013994778506457806, + 0.011609815061092377, + 0.011000852100551128, + 0.006259331479668617, + 0.0036364395637065172, + 0.004897888284176588, + 0.009230121038854122, + 0.0067767733708024025, + 0.00796156469732523, + 0.018663665279746056, + 0.017431369051337242, + 0.008549501188099384, + 0.0070724100805819035, + 0.005624963436275721, + 0.014241731725633144, + 0.012081703171133995, + 0.010608555749058723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.07002159208059311 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot captures a scene of intense sensual anticipation, rendered in dramatic black and white contrast lighting. A decadent, multi-layered chocolate fudge cake rests on a pristine white ceramic plate. Warm golden light from a single, sharp, off-camera source sculpts the dessert's shape, creating shimmering highlights on the smooth surface of the dark chocolate ganache and casting deep, velvety shadows that accentuate its rich texture. A figure, partially visible, slowly and reverently leans toward the cake. Their eyes are gently closed, long lashes resting on their cheeks, their full attention focused on the experience about to begin. A subtle, almost imperceptible tremor crosses their slightly parted lips, a physical expression of pure, unadulterated desire. The camera unwaveringly captures this delicately tense moment, just before the bite, creating an atmosphere of luxurious, solitary indulgence that is remarkably intimate and focused.", + "seed": 5906, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_3/index_2635_seed_5906_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1342651505023241, + "vlm_logprob": -0.1342651505023241, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.1342651505023241, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a multi-layered chocolate fudge cake on a white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting rendered in dramatic black and white contrast with warm golden highlights on the chocolate ganache?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure partially visible, leaning toward the cake with eyes gently closed and long lashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's mouth slightly parted with a subtle tremor visible on the lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, maintaining an unwavering frame on the scene without movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019722314551472664 + ], + "opposite_logprob_tokens": [ + -4.1447224617004395 + ], + "logprob_reward": -0.019722314551472664, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.019722314551472664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1307055652141571, + 0.1254035234451294, + 0.11773069202899933, + 0.09684785455465317, + 0.23859752714633942, + 0.08775114268064499, + 0.08145641535520554, + 0.027997802942991257, + 0.02875058725476265, + 0.06475894153118134 + ], + "credit_map": [ + 0.01517861895263195, + 0.015012724325060844, + 0.013966439291834831, + 0.015314691700041294, + 0.024786507710814476, + 0.018594129011034966, + 0.016046656295657158, + 0.011805785819888115, + 0.014703338034451008, + 0.018562287092208862, + 0.01298341155052185, + 0.01242100354284048, + 0.017968783155083656, + 0.019793476909399033, + 0.01686958782374859, + 0.012101620435714722, + 0.015230940654873848, + 0.013484238646924496, + 0.0145242465659976, + 0.011515213176608086, + 0.015566867776215076, + 0.019951337948441505, + 0.013087349012494087, + 0.014370487071573734, + 0.01650616154074669, + 0.017562078312039375, + 0.007262570783495903, + 0.0069426121190190315, + 0.01110868714749813, + 0.01579165831208229, + 0.014652653597295284, + 0.007021429482847452, + 0.08643821626901627, + 0.010543172247707844, + 0.00886076595634222, + 0.010183018632233143, + 0.01008022390305996, + 0.01031594816595316, + 0.09352808445692062, + 0.00864806491881609, + 0.01569531112909317, + 0.011509890668094158, + 0.008472204208374023, + 0.007608949206769466, + 0.011760475113987923, + 0.009679564274847507, + 0.016317380592226982, + 0.006707359571009874, + 0.011754930950701237, + 0.00974898599088192, + 0.004900256637483835, + 0.008661281317472458, + 0.013189922086894512, + 0.010830583050847054, + 0.011934655718505383, + 0.010435787960886955, + 0.0041655125096440315, + 0.0060541811399161816, + 0.0013886417727917433, + 0.001528546679764986, + 0.002450716681778431, + 0.004351054318249226, + 0.004563167225569487, + 0.003495981218293309, + 0.003919089678674936, + 0.004127244930714369, + 0.0030889390036463737, + 0.0029789679683744907, + 0.004031720571219921, + 0.00448936503380537, + 0.0036553372628986835, + 0.0024599211756139994, + 0.006869208067655563, + 0.008165483362972736, + 0.005599757190793753, + 0.006487572565674782, + 0.009473207406699657, + 0.012984657660126686, + 0.009126520715653896, + 0.006052529904991388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.019722314551472664 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a multi-layered chocolate fudge cake on a white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting rendered in dramatic black and white contrast with warm golden highlights on the chocolate ganache?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure partially visible, leaning toward the cake with eyes gently closed and long lashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's mouth slightly parted with a subtle tremor visible on the lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, maintaining an unwavering frame on the scene without movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01256372407078743 + ], + "opposite_logprob_tokens": [ + -4.387563705444336 + ], + "logprob_reward": -0.01256372407078743, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01256372407078743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10380087047815323, + 0.055981699377298355, + 0.1944112926721573, + 0.08140883594751358, + 0.2672906517982483, + 0.054984696209430695, + 0.032333068549633026, + 0.027476534247398376, + 0.040984638035297394, + 0.14132767915725708 + ], + "credit_map": [ + 0.010959072969853878, + 0.012394099496304989, + 0.006460639648139477, + 0.012038403190672398, + 0.02011597715318203, + 0.01589684560894966, + 0.0183427631855011, + 0.00759308272972703, + 0.007895202375948429, + 0.008372769691050053, + 0.006472340319305658, + 0.004773226100951433, + 0.01014542393386364, + 0.005672917235642672, + 0.008060810156166553, + 0.004589020740240812, + 0.04490023851394653, + 0.023075245320796967, + 0.018674273043870926, + 0.00962820928543806, + 0.022212157025933266, + 0.020715631544589996, + 0.041678592562675476, + 0.013526965864002705, + 0.01734287291765213, + 0.02168997935950756, + 0.005202287342399359, + 0.002763889031484723, + 0.006897418759763241, + 0.009135918691754341, + 0.013619390316307545, + 0.004757087677717209, + 0.09534221142530441, + 0.013538389466702938, + 0.009621922858059406, + 0.010989655740559101, + 0.01322740688920021, + 0.011592189781367779, + 0.10331200063228607, + 0.009666891768574715, + 0.00993411336094141, + 0.005594451446086168, + 0.004361659288406372, + 0.003958943299949169, + 0.006932460702955723, + 0.006306038238108158, + 0.012099216692149639, + 0.005797820631414652, + 0.005117721855640411, + 0.004913151729851961, + 0.0027591094840317965, + 0.0031208782456815243, + 0.004595551639795303, + 0.0037127556279301643, + 0.004153883084654808, + 0.003960020840167999, + 0.0049312058836221695, + 0.004270062781870365, + 0.0020688173826783895, + 0.002122690435498953, + 0.002468943130224943, + 0.0041264016181230545, + 0.004903670400381088, + 0.0025847458746284246, + 0.0075140902772545815, + 0.005573264788836241, + 0.0036362784449011087, + 0.003220653161406517, + 0.0035273495595902205, + 0.005451924167573452, + 0.008401780389249325, + 0.003659304231405258, + 0.030203992500901222, + 0.016423359513282776, + 0.009101791307330132, + 0.009934576228260994, + 0.017426123842597008, + 0.028680481016635895, + 0.01738823764026165, + 0.012169132009148598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 9 + ], + "selected_reward": -0.01256372407078743 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a multi-layered chocolate fudge cake on a white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting rendered in dramatic black and white contrast with warm golden highlights on the chocolate ganache?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure partially visible, leaning toward the cake with eyes gently closed and long lashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's mouth slightly parted with a subtle tremor visible on the lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, maintaining an unwavering frame on the scene without movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014198355376720428 + ], + "opposite_logprob_tokens": [ + -4.264198303222656 + ], + "logprob_reward": -0.014198355376720428, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014198355376720428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06963831186294556, + 0.0842440202832222, + 0.09919323027133942, + 0.08768914639949799, + 0.11257009953260422, + 0.2785264849662781, + 0.10938621312379837, + 0.07691806554794312, + 0.03775372356176376, + 0.044080700725317 + ], + "credit_map": [ + 0.008532855659723282, + 0.009666018187999725, + 0.006663551088422537, + 0.005340472795069218, + 0.011163286864757538, + 0.010173548012971878, + 0.009913403540849686, + 0.00818517617881298, + 0.008954544551670551, + 0.015912294387817383, + 0.006372848525643349, + 0.008654718287289143, + 0.014566021040081978, + 0.011869537644088268, + 0.01129097118973732, + 0.0066230869852006435, + 0.01815265789628029, + 0.010494214482605457, + 0.007573172450065613, + 0.008671115152537823, + 0.008981009013950825, + 0.018147172406315804, + 0.018130412325263023, + 0.00904347188770771, + 0.015718325972557068, + 0.023106535896658897, + 0.00797178503125906, + 0.005561121739447117, + 0.00795689970254898, + 0.010381958447396755, + 0.011282321065664291, + 0.005710205528885126, + 0.02563917636871338, + 0.014784491620957851, + 0.0062293196097016335, + 0.007064683362841606, + 0.01488569751381874, + 0.009702394716441631, + 0.027637241408228874, + 0.006627099122852087, + 0.047695595771074295, + 0.04230397567152977, + 0.02913389541208744, + 0.02405373565852642, + 0.02569381706416607, + 0.03179652616381645, + 0.055259086191654205, + 0.022589869797229767, + 0.015891902148723602, + 0.0183663722127676, + 0.009899678640067577, + 0.007487907540053129, + 0.009678308852016926, + 0.014398082159459591, + 0.01657777465879917, + 0.017086181789636612, + 0.012078287079930305, + 0.017342226579785347, + 0.008482861332595348, + 0.007082309108227491, + 0.007413710467517376, + 0.008457314223051071, + 0.009534199722111225, + 0.006527155637741089, + 0.006358997896313667, + 0.006010693497955799, + 0.0026879175566136837, + 0.002223888412117958, + 0.0050052679143846035, + 0.005430837161839008, + 0.007197318598628044, + 0.0028388029895722866, + 0.009985155425965786, + 0.0044374060817062855, + 0.00343671184964478, + 0.002858650404959917, + 0.00521122757345438, + 0.008021396584808826, + 0.006530868820846081, + 0.0035992832854390144 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.014198355376720428 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a multi-layered chocolate fudge cake on a white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting rendered in dramatic black and white contrast with warm golden highlights on the chocolate ganache?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure partially visible, leaning toward the cake with eyes gently closed and long lashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's mouth slightly parted with a subtle tremor visible on the lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, maintaining an unwavering frame on the scene without movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866227135062218 + ], + "opposite_logprob_tokens": [ + -3.0486621856689453 + ], + "logprob_reward": -0.04866227135062218, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866227135062218, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09818190336227417, + 0.04139059782028198, + 0.05237298831343651, + 0.13261063396930695, + 0.23308832943439484, + 0.2969732880592346, + 0.05843411386013031, + 0.05058160796761513, + 0.01847490854561329, + 0.01789163611829281 + ], + "credit_map": [ + 0.010333703830838203, + 0.013071865774691105, + 0.008053844794631004, + 0.009102170355618, + 0.01741672307252884, + 0.016755327582359314, + 0.014117635786533356, + 0.0093306303024292, + 0.004903852939605713, + 0.007836909033358097, + 0.004219315946102142, + 0.003019820200279355, + 0.008324532769620419, + 0.005385599099099636, + 0.004675114061683416, + 0.003025454469025135, + 0.007809320464730263, + 0.005910787265747786, + 0.006142535246908665, + 0.00527550233528018, + 0.007168280426412821, + 0.008684281259775162, + 0.005699296481907368, + 0.005682986695319414, + 0.02831031195819378, + 0.029698267579078674, + 0.00875691045075655, + 0.007047442719340324, + 0.013015703298151493, + 0.015709562227129936, + 0.02333959750831127, + 0.0067328400909900665, + 0.08431492000818253, + 0.010662024840712547, + 0.008473936468362808, + 0.009584939107298851, + 0.011189237236976624, + 0.009390424937009811, + 0.09085129201412201, + 0.008621550165116787, + 0.06412294507026672, + 0.04324270039796829, + 0.021734165027737617, + 0.020591450855135918, + 0.022540435194969177, + 0.024665240198373795, + 0.07440755516290665, + 0.02566877193748951, + 0.009047193452715874, + 0.008937239646911621, + 0.005260832607746124, + 0.005225084722042084, + 0.00652138190343976, + 0.0079458337277174, + 0.009978746064007282, + 0.005517803598195314, + 0.009472266770899296, + 0.007857203483581543, + 0.004035136196762323, + 0.003901110030710697, + 0.003910193685442209, + 0.005697996821254492, + 0.011373202316462994, + 0.004334497265517712, + 0.003496675519272685, + 0.002408689819276333, + 0.001845753868110478, + 0.0013006319059059024, + 0.00228264182806015, + 0.0029048568103462458, + 0.0028808717615902424, + 0.0013547880807891488, + 0.0027481038123369217, + 0.0019324433524161577, + 0.0010288364719599485, + 0.0009650916326791048, + 0.002371998503804207, + 0.0031317302491515875, + 0.0019755768589675426, + 0.0037378559354692698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.04866227135062218 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a multi-layered chocolate fudge cake on a white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting rendered in dramatic black and white contrast with warm golden highlights on the chocolate ganache?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure partially visible, leaning toward the cake with eyes gently closed and long lashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's mouth slightly parted with a subtle tremor visible on the lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, maintaining an unwavering frame on the scene without movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761790871620178 + ], + "opposite_logprob_tokens": [ + -0.8261790871620178 + ], + "logprob_reward": -0.5761790871620178, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761790871620178, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08628746122121811, + 0.05587928742170334, + 0.05843857303261757, + 0.10906213521957397, + 0.259050190448761, + 0.2526426911354065, + 0.04216529056429863, + 0.0767526775598526, + 0.038961563259363174, + 0.020760122686624527 + ], + "credit_map": [ + 0.01235868874937296, + 0.014035402797162533, + 0.005092025268822908, + 0.006286725867539644, + 0.015191616490483284, + 0.014657045714557171, + 0.010948645882308483, + 0.007717313710600138, + 0.0065678805112838745, + 0.0070990948006510735, + 0.005511627998203039, + 0.0049384357407689095, + 0.012279601767659187, + 0.008218863047659397, + 0.007202145177870989, + 0.004061638843268156, + 0.012451449409127235, + 0.00794103741645813, + 0.004933282732963562, + 0.0034755277447402477, + 0.004548177123069763, + 0.008041350170969963, + 0.012761106714606285, + 0.004286645911633968, + 0.02145664393901825, + 0.026062972843647003, + 0.008861086331307888, + 0.005768911447376013, + 0.010449663735926151, + 0.013360545039176941, + 0.01731927879154682, + 0.00578303774818778, + 0.08098813146352768, + 0.018712801858782768, + 0.010364962741732597, + 0.012181614525616169, + 0.023441264405846596, + 0.015456612221896648, + 0.08812391757965088, + 0.009780884720385075, + 0.05195551738142967, + 0.044489990919828415, + 0.019685840234160423, + 0.017522886395454407, + 0.023476572707295418, + 0.021321920678019524, + 0.050478495657444, + 0.023711467161774635, + 0.005476659629493952, + 0.005699144676327705, + 0.00272753625176847, + 0.0038497101049870253, + 0.006626774091273546, + 0.005638863891363144, + 0.006770031061023474, + 0.005376569926738739, + 0.012958649545907974, + 0.006378316320478916, + 0.004172086250036955, + 0.005530073773115873, + 0.012435270473361015, + 0.011080468073487282, + 0.015315772034227848, + 0.008882040157914162, + 0.005249057896435261, + 0.004447543527930975, + 0.004278793931007385, + 0.0035205124877393246, + 0.006659990176558495, + 0.006596301682293415, + 0.005157297011464834, + 0.0030520670115947723, + 0.0026591967325657606, + 0.0022619052324444056, + 0.001241573365405202, + 0.0015413298970088363, + 0.0028449806850403547, + 0.00476220203563571, + 0.0030068408232182264, + 0.0024420940317213535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.5761790871620178 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot captures a scene of intense sensual anticipation, rendered in dramatic black and white contrast lighting. A decadent, multi-layered chocolate fudge cake rests on a pristine white ceramic plate. Warm golden light from a single, sharp, off-camera source sculpts the dessert's shape, creating shimmering highlights on the smooth surface of the dark chocolate ganache and casting deep, velvety shadows that accentuate its rich texture. A figure, partially visible, slowly and reverently leans toward the cake. Their eyes are gently closed, long lashes resting on their cheeks, their full attention focused on the experience about to begin. A subtle, almost imperceptible tremor crosses their slightly parted lips, a physical expression of pure, unadulterated desire. The camera unwaveringly captures this delicately tense moment, just before the bite, creating an atmosphere of luxurious, solitary indulgence that is remarkably intimate and focused.", + "seed": 5907, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_3/index_2635_seed_5907_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0503332482650876, + "vlm_logprob": -0.0503332482650876, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.0503332482650876, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a multi-layered chocolate fudge cake on a white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting rendered in dramatic black and white contrast with warm golden highlights on the chocolate ganache?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure partially visible, leaning toward the cake with eyes gently closed and long lashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's mouth slightly parted with a subtle tremor visible on the lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, maintaining an unwavering frame on the scene without movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05623314157128334 + ], + "opposite_logprob_tokens": [ + -3.0562331676483154 + ], + "logprob_reward": -0.05623314157128334, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05623314157128334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36138078570365906, + 0.16485537588596344, + 0.21916572749614716, + 0.06432642042636871, + 0.061758436262607574, + 0.0350128710269928, + 0.01529107615351677, + 0.026424527168273926, + 0.019118793308734894, + 0.03266600891947746 + ], + "credit_map": [ + 0.06052234768867493, + 0.03512662276625633, + 0.03590882569551468, + 0.05606205388903618, + 0.05514932796359062, + 0.060519762337207794, + 0.025092996656894684, + 0.032998815178871155, + 0.019917476922273636, + 0.01857830211520195, + 0.01736169122159481, + 0.02067398466169834, + 0.02000974304974079, + 0.022403733804821968, + 0.023285580798983574, + 0.022624844685196877, + 0.029014788568019867, + 0.017595073208212852, + 0.014681686647236347, + 0.015947207808494568, + 0.02735510841012001, + 0.030524246394634247, + 0.05460968613624573, + 0.029437897726893425, + 0.009207761846482754, + 0.005335829686373472, + 0.0059526595287024975, + 0.004965886473655701, + 0.007429435849189758, + 0.008824152871966362, + 0.012486917898058891, + 0.01012376882135868, + 0.007998877204954624, + 0.006110864225775003, + 0.006009336095303297, + 0.006611792836338282, + 0.008342214860022068, + 0.008848937228322029, + 0.010018753819167614, + 0.007817653939127922, + 0.004726480692625046, + 0.0034987679682672024, + 0.004966185428202152, + 0.007292725145816803, + 0.003866826416924596, + 0.005474486388266087, + 0.002584044123068452, + 0.002603350905701518, + 0.0016521408688277006, + 0.0016969186253845692, + 0.0021595617290586233, + 0.001895090565085411, + 0.002111489651724696, + 0.002205077325925231, + 0.0013107926351949573, + 0.002260003937408328, + 0.002751493128016591, + 0.002636243123561144, + 0.0044270846992731094, + 0.0033251505810767412, + 0.0041285413317382336, + 0.004568354692310095, + 0.0018756008939817548, + 0.002712053479626775, + 0.002137301955372095, + 0.002136488910764456, + 0.0018339769449084997, + 0.002504173666238785, + 0.0038123521953821182, + 0.0033581829629838467, + 0.0013650001492351294, + 0.0019713151268661022, + 0.004862396512180567, + 0.004109219182282686, + 0.004816518165171146, + 0.0048795947805047035, + 0.005250499118119478, + 0.004751923028379679, + 0.001886028447188437, + 0.0021098267752677202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.05623314157128334 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a multi-layered chocolate fudge cake on a white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting rendered in dramatic black and white contrast with warm golden highlights on the chocolate ganache?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure partially visible, leaning toward the cake with eyes gently closed and long lashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's mouth slightly parted with a subtle tremor visible on the lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, maintaining an unwavering frame on the scene without movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007652136497199535 + ], + "opposite_logprob_tokens": [ + -4.882652282714844 + ], + "logprob_reward": -0.007652136497199535, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007652136497199535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0936410054564476, + 0.1639881283044815, + 0.19631627202033997, + 0.1349717080593109, + 0.08947689086198807, + 0.039049286395311356, + 0.13127702474594116, + 0.09173671156167984, + 0.03729044646024704, + 0.02225254848599434 + ], + "credit_map": [ + 0.01415121927857399, + 0.012863660231232643, + 0.010937568731606007, + 0.009168846532702446, + 0.015515194274485111, + 0.013164859265089035, + 0.00824060756713152, + 0.009599047712981701, + 0.01870780438184738, + 0.018939530476927757, + 0.012464740313589573, + 0.019318684935569763, + 0.020378602668642998, + 0.0210647564381361, + 0.026210499927401543, + 0.026903493329882622, + 0.02483174204826355, + 0.01737021654844284, + 0.017703915014863014, + 0.015936316922307014, + 0.027313383296132088, + 0.02810375764966011, + 0.042657505720853806, + 0.022399403154850006, + 0.02203183062374592, + 0.01822640188038349, + 0.015577662736177444, + 0.012480887584388256, + 0.020613882690668106, + 0.019647469744086266, + 0.013700490817427635, + 0.012693078257143497, + 0.009710508398711681, + 0.008761381730437279, + 0.006661982741206884, + 0.00798210222274065, + 0.011440638452768326, + 0.014674711041152477, + 0.016972100362181664, + 0.01327346358448267, + 0.004267609212547541, + 0.003497342113405466, + 0.003855448681861162, + 0.005696286913007498, + 0.004865408409386873, + 0.006261602509766817, + 0.004890741780400276, + 0.005714843515306711, + 0.015900788828730583, + 0.013367434963583946, + 0.013623666949570179, + 0.011948774568736553, + 0.02103784680366516, + 0.021044360473752022, + 0.014503400772809982, + 0.019850727170705795, + 0.009152124635875225, + 0.010207015089690685, + 0.010091383941471577, + 0.013120697811245918, + 0.016275689005851746, + 0.015786362811923027, + 0.009952339343726635, + 0.007151094265282154, + 0.004152553156018257, + 0.004016183316707611, + 0.004186499863862991, + 0.006346018519252539, + 0.004379099700599909, + 0.005866541061550379, + 0.003652887651696801, + 0.004690659232437611, + 0.002801804570481181, + 0.002390769310295582, + 0.0033657781314104795, + 0.0030051590874791145, + 0.00406225398182869, + 0.0034521606285125017, + 0.0014553507789969444, + 0.0017192711820825934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.007652136497199535 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a multi-layered chocolate fudge cake on a white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting rendered in dramatic black and white contrast with warm golden highlights on the chocolate ganache?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure partially visible, leaning toward the cake with eyes gently closed and long lashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's mouth slightly parted with a subtle tremor visible on the lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, maintaining an unwavering frame on the scene without movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980075217783451 + ], + "opposite_logprob_tokens": [ + -3.5298006534576416 + ], + "logprob_reward": -0.02980075217783451, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980075217783451, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38936102390289307, + 0.09691332280635834, + 0.10651463270187378, + 0.03109530359506607, + 0.0665101706981659, + 0.07860244065523148, + 0.09185253828763962, + 0.10067020356655121, + 0.025446735322475433, + 0.013033620081841946 + ], + "credit_map": [ + 0.05799679458141327, + 0.04386264458298683, + 0.050629980862140656, + 0.06522636115550995, + 0.05925054848194122, + 0.05323118343949318, + 0.02582419663667679, + 0.03333931788802147, + 0.011801062151789665, + 0.007911634631454945, + 0.006672424729913473, + 0.010206960141658783, + 0.013504894450306892, + 0.01143601443618536, + 0.020116835832595825, + 0.015263502486050129, + 0.01596594788134098, + 0.01592097617685795, + 0.012044882401823997, + 0.012605450116097927, + 0.01837676763534546, + 0.016234859824180603, + 0.008605623617768288, + 0.006760117597877979, + 0.004814298823475838, + 0.00359228253364563, + 0.003450200194492936, + 0.0031077007297426462, + 0.004448354709893465, + 0.004672168288379908, + 0.003880353644490242, + 0.00312994746491313, + 0.00793373305350542, + 0.00874677486717701, + 0.008741667494177818, + 0.009321335703134537, + 0.010231707245111465, + 0.009793286211788654, + 0.005085700657218695, + 0.006655971519649029, + 0.009658681228756905, + 0.012610133737325668, + 0.007737987209111452, + 0.009068929590284824, + 0.01381165161728859, + 0.011222134344279766, + 0.006161427590996027, + 0.008331495337188244, + 0.011697700247168541, + 0.009104138240218163, + 0.008729425258934498, + 0.008545521646738052, + 0.014200426638126373, + 0.014732527546584606, + 0.011216231621801853, + 0.013626569882035255, + 0.009497474879026413, + 0.011109434068202972, + 0.01215376053005457, + 0.01341716293245554, + 0.0175961684435606, + 0.017590098083019257, + 0.010678073391318321, + 0.008628028444945812, + 0.0025315633974969387, + 0.002997444011271, + 0.002921863691881299, + 0.003625433426350355, + 0.004081747028976679, + 0.004400461446493864, + 0.0020749717950820923, + 0.0028132502920925617, + 0.001408879877999425, + 0.0025276404339820147, + 0.0019967544358223677, + 0.0016992696328088641, + 0.0016531419241800904, + 0.0019172742031514645, + 0.00077130610588938, + 0.0010593527695164084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.02980075217783451 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a multi-layered chocolate fudge cake on a white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting rendered in dramatic black and white contrast with warm golden highlights on the chocolate ganache?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure partially visible, leaning toward the cake with eyes gently closed and long lashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's mouth slightly parted with a subtle tremor visible on the lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, maintaining an unwavering frame on the scene without movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899969071149826 + ], + "opposite_logprob_tokens": [ + -2.5789997577667236 + ], + "logprob_reward": -0.07899969071149826, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899969071149826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38096317648887634, + 0.16192445158958435, + 0.10764636099338531, + 0.04564914107322693, + 0.0437360517680645, + 0.04590368643403053, + 0.058557700365781784, + 0.06096309795975685, + 0.053991418331861496, + 0.04066489264369011 + ], + "credit_map": [ + 0.052245546132326126, + 0.04906424134969711, + 0.047788359224796295, + 0.06774324178695679, + 0.05864768102765083, + 0.05127668380737305, + 0.022492721676826477, + 0.03170473501086235, + 0.020561646670103073, + 0.010958384722471237, + 0.011980208568274975, + 0.01996704749763012, + 0.01680467091500759, + 0.019990278407931328, + 0.03536291420459747, + 0.02629932574927807, + 0.014927140437066555, + 0.009150420315563679, + 0.009105963632464409, + 0.009502614848315716, + 0.013925529085099697, + 0.015273374505341053, + 0.02352912724018097, + 0.012232201173901558, + 0.006995782721787691, + 0.004932379350066185, + 0.0043061235919594765, + 0.0048493570648133755, + 0.005701180547475815, + 0.0059899212792515755, + 0.005951772443950176, + 0.006922627333551645, + 0.006949745584279299, + 0.006410727743059397, + 0.004146798048168421, + 0.0034605246037244797, + 0.005700543988496065, + 0.007122258190065622, + 0.004333384335041046, + 0.0056120725348591805, + 0.00679390225559473, + 0.008117590099573135, + 0.00422395346686244, + 0.004928259178996086, + 0.007231013383716345, + 0.0068701789714396, + 0.0038036310579627752, + 0.0039351629093289375, + 0.007414114195853472, + 0.006512010935693979, + 0.006093097850680351, + 0.0052780890837311745, + 0.009555384516716003, + 0.009501410648226738, + 0.006706088315695524, + 0.007497510407119989, + 0.006661583203822374, + 0.00695998128503561, + 0.004390972666442394, + 0.006213862914592028, + 0.010418913327157497, + 0.009873208589851856, + 0.008577888831496239, + 0.00786669086664915, + 0.006227741949260235, + 0.005532220005989075, + 0.0048859743401408195, + 0.007634805981069803, + 0.005884310230612755, + 0.008010546676814556, + 0.007532490883022547, + 0.008283331990242004, + 0.005826342385262251, + 0.004776036832481623, + 0.006042706314474344, + 0.0059198918752372265, + 0.005434515420347452, + 0.005565387662500143, + 0.002428660402074456, + 0.004671357572078705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.07899969071149826 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a multi-layered chocolate fudge cake on a white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting rendered in dramatic black and white contrast with warm golden highlights on the chocolate ganache?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure partially visible, leaning toward the cake with eyes gently closed and long lashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's mouth slightly parted with a subtle tremor visible on the lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, maintaining an unwavering frame on the scene without movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898052036762238 + ], + "opposite_logprob_tokens": [ + -2.5789804458618164 + ], + "logprob_reward": -0.07898052036762238, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898052036762238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2848823070526123, + 0.19235087931156158, + 0.17583471536636353, + 0.07744386047124863, + 0.0373942032456398, + 0.041515156626701355, + 0.04766932502388954, + 0.0515715517103672, + 0.03312138095498085, + 0.058216605335474014 + ], + "credit_map": [ + 0.04046782851219177, + 0.04159601777791977, + 0.025540359318256378, + 0.04021953046321869, + 0.05375847592949867, + 0.03876820206642151, + 0.019552156329154968, + 0.024979734793305397, + 0.026780594140291214, + 0.024022134020924568, + 0.018502263352274895, + 0.021338757127523422, + 0.03070257417857647, + 0.02822238579392433, + 0.017681175842881203, + 0.02510100044310093, + 0.019838761538267136, + 0.010625715367496014, + 0.011037079617381096, + 0.016055192798376083, + 0.019600890576839447, + 0.026819676160812378, + 0.04255751892924309, + 0.02929988130927086, + 0.011962729506194592, + 0.009734819643199444, + 0.007722950074821711, + 0.009140204638242722, + 0.009852618910372257, + 0.011658436618745327, + 0.008610900491476059, + 0.00876119825989008, + 0.004761823453009129, + 0.004499298986047506, + 0.004760523792356253, + 0.0048948656767606735, + 0.005875925067812204, + 0.005450185853987932, + 0.0028871686663478613, + 0.004264412447810173, + 0.0050539555959403515, + 0.004097647499293089, + 0.0053781080059707165, + 0.00549037242308259, + 0.004876696038991213, + 0.005337303038686514, + 0.00427045114338398, + 0.007010622415691614, + 0.005484865512698889, + 0.0043123275972902775, + 0.007438055705279112, + 0.0075011528097093105, + 0.006778930313885212, + 0.0069232783280313015, + 0.003185243345797062, + 0.006045473273843527, + 0.005979620385915041, + 0.008211642503738403, + 0.005232046823948622, + 0.007050527725368738, + 0.008175704628229141, + 0.008452439680695534, + 0.004016336519271135, + 0.004453234374523163, + 0.0037223517429083586, + 0.0029498778749257326, + 0.0038466716650873423, + 0.005250528920441866, + 0.004944340791553259, + 0.005183086264878511, + 0.003713632235303521, + 0.0035108940210193396, + 0.006444273050874472, + 0.007309250067919493, + 0.010822958312928677, + 0.008888619020581245, + 0.006316313985735178, + 0.008623464033007622, + 0.004948883317410946, + 0.004862843081355095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.07898052036762238 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_4/index_tensor([5548, 5548])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_4/index_tensor([5548, 5548])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7717e463d525ae6a4b04ae0ed0047f36eb613487 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_4/index_tensor([5548, 5548])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, static wide shot reveals the majestic Himeji Castle, its pristine white walls and elegantly curved, multi-tiered roofs set against an ominous, bruised purple and dark gray sky, thick with dark clouds. Instead of a traditional moat, the castle is surrounded by a terrifying torrent of black, viscous mud, violently churning and carrying with it broken trees, rocks, and unrecognizable debris. This relentless torrent cascades down the hillside with unstoppable force, crashing against the castle's massive stone foundations with powerful waves. Thick, earthy sludge splashes, staining the pristine white plaster of the lower walls. The storm's dim, oppressive light casts deep shadows, accentuating the stark, terrifying contrast between the castle's tranquil architectural beauty and the raw, destructive force of the mudslide, creating a deeply unsettling and surreal atmosphere of impending doom.", + "seed": 6025, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_4/index_5548_seed_6025_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.044698886945843695, + "vlm_logprob": -0.044698886945843695, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.044698886945843695, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle's architecture white with elegantly curved, multi-tiered roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the castle surrounded by a torrent of black, viscous mud instead of a traditional moat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky ominous, bruised purple and dark gray with thick dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09895601123571396 + ], + "opposite_logprob_tokens": [ + -2.5989561080932617 + ], + "logprob_reward": -0.09895601123571396, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.09895601123571396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6333837509155273, + 0.0770145133137703, + 0.026893161237239838, + 0.049416590481996536, + 0.0212177075445652, + 0.020597366616129875, + 0.08157867193222046, + 0.04087935760617256, + 0.03904077410697937, + 0.009978164918720722 + ], + "credit_map": [ + 0.2380073368549347, + 0.027050403878092766, + 0.031243622303009033, + 0.022540917620062828, + 0.029965201392769814, + 0.24652643501758575, + 0.01728963851928711, + 0.020760174840688705, + 0.010349227115511894, + 0.009383458644151688, + 0.00970915425568819, + 0.0075744809582829475, + 0.012494681403040886, + 0.00959039106965065, + 0.009482741355895996, + 0.008430368266999722, + 0.003248210297897458, + 0.00374444923363626, + 0.0040590898133814335, + 0.0028637198265641928, + 0.005182737018913031, + 0.002573929261416197, + 0.003004285506904125, + 0.002216739347204566, + 0.011970356106758118, + 0.0040925005450844765, + 0.004966591019183397, + 0.002947436412796378, + 0.005574256181716919, + 0.014301246963441372, + 0.0026635213289409876, + 0.0029006802942603827, + 0.004153605084866285, + 0.0024489418137818575, + 0.002390496199950576, + 0.0018525822088122368, + 0.0020666578784585, + 0.004875936545431614, + 0.001996653852984309, + 0.0014328351244330406, + 0.003523273626342416, + 0.003082775743678212, + 0.002695936243981123, + 0.0020639279391616583, + 0.002060703234747052, + 0.003995876293629408, + 0.001618653885088861, + 0.0015562194166705012, + 0.02346671000123024, + 0.004277013707906008, + 0.0056589823216199875, + 0.005831608548760414, + 0.007482524961233139, + 0.023754671216011047, + 0.006293377839028835, + 0.004813780542463064, + 0.010644160211086273, + 0.004245472606271505, + 0.0027663635555654764, + 0.002210671314969659, + 0.003145744325593114, + 0.01350910123437643, + 0.002281010616570711, + 0.0020768342074006796, + 0.007603972684592009, + 0.0037414326798170805, + 0.0038446676917374134, + 0.003668304067105055, + 0.005943300202488899, + 0.007472523022443056, + 0.004061262123286724, + 0.002705311169847846, + 0.0015594619326293468, + 0.0011474357452243567, + 0.0011856259079650044, + 0.0010533009190112352, + 0.0015662872465327382, + 0.001470415503717959, + 0.0009653908200562, + 0.0010302466107532382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.09895601123571396 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle's architecture white with elegantly curved, multi-tiered roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the castle surrounded by a torrent of black, viscous mud instead of a traditional moat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky ominous, bruised purple and dark gray with thick dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0182375218719244 + ], + "opposite_logprob_tokens": [ + -4.018237590789795 + ], + "logprob_reward": -0.0182375218719244, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0182375218719244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38855284452438354, + 0.08158678561449051, + 0.03957207500934601, + 0.05199899151921272, + 0.032076407223939896, + 0.03020578622817993, + 0.12136057764291763, + 0.08576777577400208, + 0.08147595077753067, + 0.08740285038948059 + ], + "credit_map": [ + 0.1369343400001526, + 0.01815658062696457, + 0.020323557779192924, + 0.0172395259141922, + 0.0226628165692091, + 0.14121714234352112, + 0.016410943120718002, + 0.015607893466949463, + 0.009396275505423546, + 0.008596925064921379, + 0.008469358086585999, + 0.010162396356463432, + 0.0150224007666111, + 0.010190900415182114, + 0.010428361594676971, + 0.009320165030658245, + 0.004430065862834454, + 0.0043679517693817616, + 0.005587396677583456, + 0.005330342333763838, + 0.005385510623455048, + 0.0055345757864415646, + 0.0050258394330739975, + 0.003910391591489315, + 0.014699000865221024, + 0.0036479011178016663, + 0.002761638257652521, + 0.0026870553847402334, + 0.0042150490917265415, + 0.018203672021627426, + 0.003328870516270399, + 0.0024558003060519695, + 0.004907555878162384, + 0.00322475447319448, + 0.004164103418588638, + 0.003396145533770323, + 0.0028933740686625242, + 0.007371420040726662, + 0.003724802052602172, + 0.0023942487314343452, + 0.005367106292396784, + 0.0032506603747606277, + 0.005199577193707228, + 0.0031589819118380547, + 0.0035968660376966, + 0.0048212045803666115, + 0.0023377351462841034, + 0.0024736537598073483, + 0.03676998242735863, + 0.006067498121410608, + 0.008848556317389011, + 0.0073758759535849094, + 0.009407414123415947, + 0.03878118097782135, + 0.007687320467084646, + 0.006422743666917086, + 0.011104952543973923, + 0.010119645856320858, + 0.011466745287179947, + 0.008955743163824081, + 0.014277936890721321, + 0.013961435295641422, + 0.008821197785437107, + 0.007060113362967968, + 0.017405446618795395, + 0.00616395054385066, + 0.007332343142479658, + 0.006967972964048386, + 0.011534243822097778, + 0.017988786101341248, + 0.008493348956108093, + 0.005589853972196579, + 0.011773664504289627, + 0.00792074203491211, + 0.012131153605878353, + 0.010429038666188717, + 0.01345906499773264, + 0.010287415236234665, + 0.0132265230640769, + 0.008175245486199856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.0182375218719244 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle's architecture white with elegantly curved, multi-tiered roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the castle surrounded by a torrent of black, viscous mud instead of a traditional moat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky ominous, bruised purple and dark gray with thick dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904535323381424 + ], + "opposite_logprob_tokens": [ + -2.464045286178589 + ], + "logprob_reward": -0.08904535323381424, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904535323381424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6751628518104553, + 0.07030000537633896, + 0.03615216538310051, + 0.01198627334088087, + 0.02695288509130478, + 0.03926079720258713, + 0.0854395180940628, + 0.02226482890546322, + 0.022132810205221176, + 0.010347832925617695 + ], + "credit_map": [ + 0.25422507524490356, + 0.02915332093834877, + 0.03227246180176735, + 0.02418500930070877, + 0.03137238323688507, + 0.2632831037044525, + 0.016332918778061867, + 0.024338629096746445, + 0.008854742161929607, + 0.008853073231875896, + 0.010497476905584335, + 0.007640511728823185, + 0.009693519212305546, + 0.009731153026223183, + 0.007288713473826647, + 0.007740816567093134, + 0.005063255783170462, + 0.003831291338428855, + 0.0059449328109622, + 0.003531314432621002, + 0.007226068060845137, + 0.0042856959626078606, + 0.003095519496127963, + 0.003174089826643467, + 0.0014740136684849858, + 0.001721707172691822, + 0.0019425464561209083, + 0.0017673231195658445, + 0.001346781151369214, + 0.0012850826606154442, + 0.0013535999460145831, + 0.001095220330171287, + 0.008015139028429985, + 0.0020335260778665543, + 0.001991448923945427, + 0.0018081746529787779, + 0.0018342867260798812, + 0.008781538344919682, + 0.0011970172636210918, + 0.001291757682338357, + 0.004506884608417749, + 0.0038003234658390284, + 0.005179254803806543, + 0.004639697726815939, + 0.006264473311603069, + 0.004299597814679146, + 0.006266464479267597, + 0.00430410448461771, + 0.017499255016446114, + 0.005219846498221159, + 0.011886199936270714, + 0.00845527183264494, + 0.008927733637392521, + 0.01656261458992958, + 0.008938843384385109, + 0.00794975832104683, + 0.003814822295680642, + 0.0021336853969842196, + 0.0024119652807712555, + 0.002023515524342656, + 0.0027745189145207405, + 0.0051094647496938705, + 0.0025474356953054667, + 0.0014494233764708042, + 0.004648618400096893, + 0.002248401753604412, + 0.0020856584887951612, + 0.0018321368843317032, + 0.0032236177939921618, + 0.004528839606791735, + 0.0021011868957430124, + 0.0014643521280959249, + 0.0011977690737694502, + 0.0014110434567555785, + 0.0016325420001521707, + 0.0011820077197626233, + 0.0015395289519801736, + 0.0013752547092735767, + 0.000802155933342874, + 0.0012075314298272133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.08904535323381424 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle's architecture white with elegantly curved, multi-tiered roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the castle surrounded by a torrent of black, viscous mud instead of a traditional moat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky ominous, bruised purple and dark gray with thick dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011848341673612595 + ], + "opposite_logprob_tokens": [ + -6.751184940338135 + ], + "logprob_reward": -0.0011848341673612595, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011848341673612595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44038453698158264, + 0.0911220908164978, + 0.09413101524114609, + 0.026291407644748688, + 0.05567168816924095, + 0.04218835383653641, + 0.09869387745857239, + 0.0640605241060257, + 0.046502962708473206, + 0.040953490883111954 + ], + "credit_map": [ + 0.15804287791252136, + 0.031241925433278084, + 0.020648064091801643, + 0.018963376060128212, + 0.022793874144554138, + 0.1587352752685547, + 0.0147407790645957, + 0.015218382701277733, + 0.011406878009438515, + 0.010964520275592804, + 0.01609182357788086, + 0.009972858242690563, + 0.011797684244811535, + 0.01357409916818142, + 0.008641381748020649, + 0.008672845549881458, + 0.025495249778032303, + 0.007627185899764299, + 0.008321813307702541, + 0.005141489673405886, + 0.008196369744837284, + 0.030047744512557983, + 0.004449189640581608, + 0.004851972684264183, + 0.0060146041214466095, + 0.0024881914723664522, + 0.003117725718766451, + 0.0017501054098829627, + 0.0021236175671219826, + 0.006553750019520521, + 0.002031337469816208, + 0.002212075050920248, + 0.014619076624512672, + 0.004598298575729132, + 0.005140192341059446, + 0.004057941492646933, + 0.0042013912461698055, + 0.016920650377869606, + 0.003732516663148999, + 0.0024016215465962887, + 0.0064532943069934845, + 0.005628284532576799, + 0.005608397535979748, + 0.0042449720203876495, + 0.006391718052327633, + 0.00671934150159359, + 0.003693803446367383, + 0.003448543371632695, + 0.030485864728689194, + 0.006375924684107304, + 0.005843483842909336, + 0.004974937532097101, + 0.006903120782226324, + 0.03372539207339287, + 0.005325686186552048, + 0.005059465765953064, + 0.014236334711313248, + 0.004458781331777573, + 0.006115437019616365, + 0.00513498205691576, + 0.006745416671037674, + 0.016200589016079903, + 0.007085397373884916, + 0.004083586856722832, + 0.00855403020977974, + 0.006822387222200632, + 0.005263401195406914, + 0.004632849711924791, + 0.007551359478384256, + 0.006469103507697582, + 0.004081538878381252, + 0.0031282915733754635, + 0.00560448644682765, + 0.0049814810045063496, + 0.00493603153154254, + 0.00473072100430727, + 0.005391987971961498, + 0.007203614804893732, + 0.0043624634854495525, + 0.0037427048664540052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.0011848341673612595 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle's architecture white with elegantly curved, multi-tiered roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the castle surrounded by a torrent of black, viscous mud instead of a traditional moat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky ominous, bruised purple and dark gray with thick dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016070714220404625 + ], + "opposite_logprob_tokens": [ + -4.14107084274292 + ], + "logprob_reward": -0.016070714220404625, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016070714220404625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5750741958618164, + 0.04166075959801674, + 0.029413865879178047, + 0.018420593813061714, + 0.01761678420007229, + 0.03494836390018463, + 0.15745434165000916, + 0.04150786250829697, + 0.06549321860074997, + 0.018410054966807365 + ], + "credit_map": [ + 0.21384449303150177, + 0.025356223806738853, + 0.027867058292031288, + 0.02044828236103058, + 0.026534665375947952, + 0.22569939494132996, + 0.014768369495868683, + 0.020555736497044563, + 0.004050226882100105, + 0.003928066231310368, + 0.0052742306143045425, + 0.004190618172287941, + 0.0080401711165905, + 0.004784541670233011, + 0.0056536453776061535, + 0.005739262793213129, + 0.004180749878287315, + 0.0029036267660558224, + 0.0047891875728964806, + 0.002306521637365222, + 0.005054947454482317, + 0.004607679788023233, + 0.002809123368933797, + 0.0027620289474725723, + 0.005022761411964893, + 0.0017712042899802327, + 0.001475850585848093, + 0.0011384659446775913, + 0.0014266816433519125, + 0.005749482195824385, + 0.0009009947534650564, + 0.0009351513581350446, + 0.0027660478372126818, + 0.0017215072875842452, + 0.0025001137983053923, + 0.0021653936710208654, + 0.0015535826096311212, + 0.003016554517671466, + 0.002547615673393011, + 0.0013459700858220458, + 0.006027962081134319, + 0.004271887708455324, + 0.005123250652104616, + 0.003161296248435974, + 0.004471173509955406, + 0.006266227923333645, + 0.0027312294114381075, + 0.0028953358996659517, + 0.043249163776636124, + 0.008598976768553257, + 0.01183382049202919, + 0.00974768865853548, + 0.013577859848737717, + 0.049485765397548676, + 0.01167959626764059, + 0.009281462989747524, + 0.005222783889621496, + 0.0062931980937719345, + 0.005474066361784935, + 0.004550048150122166, + 0.006575056351721287, + 0.004581338260322809, + 0.005337792448699474, + 0.0034735819790512323, + 0.012811577878892422, + 0.008411217480897903, + 0.006781095638871193, + 0.006388779729604721, + 0.010433412156999111, + 0.009933008812367916, + 0.006228309124708176, + 0.004505820572376251, + 0.002620382932946086, + 0.0015159148024395108, + 0.0024236030876636505, + 0.0020794731099158525, + 0.0028199271764606237, + 0.0028545826207846403, + 0.0024787276051938534, + 0.0016174427000805736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.016070714220404625 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static wide shot reveals the majestic Himeji Castle, its pristine white walls and elegantly curved, multi-tiered roofs set against an ominous, bruised purple and dark gray sky, thick with dark clouds. Instead of a traditional moat, the castle is surrounded by a terrifying torrent of black, viscous mud, violently churning and carrying with it broken trees, rocks, and unrecognizable debris. This relentless torrent cascades down the hillside with unstoppable force, crashing against the castle's massive stone foundations with powerful waves. Thick, earthy sludge splashes, staining the pristine white plaster of the lower walls. The storm's dim, oppressive light casts deep shadows, accentuating the stark, terrifying contrast between the castle's tranquil architectural beauty and the raw, destructive force of the mudslide, creating a deeply unsettling and surreal atmosphere of impending doom.", + "seed": 6026, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_4/index_5548_seed_6026_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07096859831362963, + "vlm_logprob": -0.07096859831362963, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.07096859831362963, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle's architecture white with elegantly curved, multi-tiered roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the castle surrounded by a torrent of black, viscous mud instead of a traditional moat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky ominous, bruised purple and dark gray with thick dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2253858894109726 + ], + "opposite_logprob_tokens": [ + -1.7253859043121338 + ], + "logprob_reward": -0.2253858894109726, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2253858894109726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23611612617969513, + 0.09692336618900299, + 0.09387357532978058, + 0.17636491358280182, + 0.09251219779253006, + 0.07910525798797607, + 0.07822807133197784, + 0.0386865958571434, + 0.09504806250333786, + 0.013141808100044727 + ], + "credit_map": [ + 0.03219690918922424, + 0.030725618824362755, + 0.024723362177610397, + 0.023638801649212837, + 0.028034910559654236, + 0.034693267196416855, + 0.03298744186758995, + 0.029115816578269005, + 0.011099365539848804, + 0.01582206040620804, + 0.009588297456502914, + 0.009990229271352291, + 0.010292043909430504, + 0.017622582614421844, + 0.011810350231826305, + 0.010698440484702587, + 0.009744251146912575, + 0.010718866251409054, + 0.01236657239496708, + 0.010453976690769196, + 0.013523731380701065, + 0.01639631576836109, + 0.009022136218845844, + 0.011647725477814674, + 0.014649405144155025, + 0.016789060086011887, + 0.024489082396030426, + 0.021810179576277733, + 0.02063179947435856, + 0.03172234073281288, + 0.02441411465406418, + 0.02185892127454281, + 0.009882266633212566, + 0.010847830213606358, + 0.012566078454256058, + 0.00890796072781086, + 0.015505059622228146, + 0.014396400190889835, + 0.011507614515721798, + 0.008898984640836716, + 0.010134240612387657, + 0.009423965588212013, + 0.014375897124409676, + 0.010645742528140545, + 0.007692962419241667, + 0.010226604528725147, + 0.005990410689264536, + 0.010615433566272259, + 0.010330354794859886, + 0.008067609742283821, + 0.01497182622551918, + 0.009098077192902565, + 0.008959411643445492, + 0.010376815684139729, + 0.007012443151324987, + 0.00941153708845377, + 0.0050979116931557655, + 0.00467463955283165, + 0.004511179402470589, + 0.005694160703569651, + 0.005620377138257027, + 0.00538121722638607, + 0.003525956068187952, + 0.004181151278316975, + 0.017342599108815193, + 0.012940611690282822, + 0.011998455040156841, + 0.013134908862411976, + 0.009950106032192707, + 0.01193494163453579, + 0.0072388737462460995, + 0.01050756499171257, + 0.0015131885884329677, + 0.0017560920678079128, + 0.0015022649895399809, + 0.0017233751714229584, + 0.0018909100908786058, + 0.001843799720518291, + 0.0013099894858896732, + 0.0016021870542317629 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.2253858894109726 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle's architecture white with elegantly curved, multi-tiered roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the castle surrounded by a torrent of black, viscous mud instead of a traditional moat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky ominous, bruised purple and dark gray with thick dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0487087108194828 + ], + "opposite_logprob_tokens": [ + -3.04870867729187 + ], + "logprob_reward": -0.0487087108194828, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0487087108194828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23472340404987335, + 0.08881282061338425, + 0.07132799178361893, + 0.09431114792823792, + 0.2232053130865097, + 0.05158938094973564, + 0.03674710541963577, + 0.07035918533802032, + 0.033604130148887634, + 0.09531954675912857 + ], + "credit_map": [ + 0.02560046687722206, + 0.02463683858513832, + 0.02534513548016548, + 0.02296515554189682, + 0.0534309446811676, + 0.028346071019768715, + 0.02207445539534092, + 0.03232431784272194, + 0.008776046335697174, + 0.009222003631293774, + 0.011682721786201, + 0.00875109527260065, + 0.0185234434902668, + 0.010753261856734753, + 0.008964037522673607, + 0.012140205129981041, + 0.007601628545671701, + 0.008879204280674458, + 0.010820885188877583, + 0.00811056699603796, + 0.012181977741420269, + 0.009735791012644768, + 0.0064931525848805904, + 0.007504784502089024, + 0.007078076712787151, + 0.00926138088107109, + 0.010687995702028275, + 0.010088637471199036, + 0.019186507910490036, + 0.015500246547162533, + 0.010245841927826405, + 0.01226245891302824, + 0.017989465966820717, + 0.022280339151620865, + 0.024299517273902893, + 0.019778460264205933, + 0.042415887117385864, + 0.03554176911711693, + 0.03148958459496498, + 0.029410282149910927, + 0.004123717080801725, + 0.004589005373418331, + 0.005961155518889427, + 0.006259536370635033, + 0.00784460548311472, + 0.00709202466532588, + 0.0063257827423512936, + 0.009393549524247646, + 0.0034188725985586643, + 0.003865391481667757, + 0.003946828190237284, + 0.00538427010178566, + 0.0063133565708994865, + 0.005752675700932741, + 0.0033293606247752905, + 0.0047363475896418095, + 0.0060593183152377605, + 0.007144827861338854, + 0.005888276267796755, + 0.006481699179857969, + 0.016402406617999077, + 0.013232109136879444, + 0.008004291914403439, + 0.007146256510168314, + 0.004762409254908562, + 0.003711468307301402, + 0.003486651461571455, + 0.004953633062541485, + 0.005803478416055441, + 0.004056766629219055, + 0.0029044041875749826, + 0.0039253197610378265, + 0.00737520307302475, + 0.009579787962138653, + 0.010597859509289265, + 0.007944456301629543, + 0.018650216981768608, + 0.01552010327577591, + 0.010651757009327412, + 0.015000156126916409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.0487087108194828 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle's architecture white with elegantly curved, multi-tiered roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the castle surrounded by a torrent of black, viscous mud instead of a traditional moat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky ominous, bruised purple and dark gray with thick dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496400222182274 + ], + "opposite_logprob_tokens": [ + -2.929964065551758 + ], + "logprob_reward": -0.05496400222182274, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496400222182274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23569168150424957, + 0.22184482216835022, + 0.12765498459339142, + 0.05051880702376366, + 0.05083877593278885, + 0.16511854529380798, + 0.03944404795765877, + 0.03383902087807655, + 0.04287390410900116, + 0.032175462692976 + ], + "credit_map": [ + 0.03416954725980759, + 0.028285017237067223, + 0.0386379174888134, + 0.027425585314631462, + 0.024333059787750244, + 0.03238062933087349, + 0.018730176612734795, + 0.03172973915934563, + 0.028026970103383064, + 0.02916323021054268, + 0.03328976407647133, + 0.035010553896427155, + 0.01692383922636509, + 0.03162098675966263, + 0.018700754269957542, + 0.029108703136444092, + 0.015699921175837517, + 0.01661716401576996, + 0.02039908803999424, + 0.016756096854805946, + 0.012744201347231865, + 0.016633016988635063, + 0.010476835072040558, + 0.018328657373785973, + 0.006563729140907526, + 0.006090337410569191, + 0.009166011586785316, + 0.007207927759736776, + 0.006067271344363689, + 0.005726987961679697, + 0.004150340333580971, + 0.005546197760850191, + 0.0062591396272182465, + 0.005351779516786337, + 0.007244723383337259, + 0.006645397748798132, + 0.007038844283670187, + 0.00681248540058732, + 0.00544244609773159, + 0.006043955683708191, + 0.02046581543982029, + 0.020075781270861626, + 0.03365406394004822, + 0.022146953269839287, + 0.011841226369142532, + 0.024529313668608665, + 0.013136318884789944, + 0.019269052892923355, + 0.004446242470294237, + 0.0062951878644526005, + 0.004414761904627085, + 0.0036704286467283964, + 0.005798008292913437, + 0.007022934965789318, + 0.0040786475874483585, + 0.003717834362760186, + 0.0046177939511835575, + 0.00402551656588912, + 0.003388222772628069, + 0.006180477794259787, + 0.003345372388139367, + 0.004351609852164984, + 0.003765442641451955, + 0.0041645788587629795, + 0.006852938327938318, + 0.006814371328800917, + 0.004820515867322683, + 0.00487768929451704, + 0.005404877010732889, + 0.005746349226683378, + 0.004613877274096012, + 0.0037432806566357613, + 0.004155575763434172, + 0.005138413514941931, + 0.0035258070565760136, + 0.003865218721330166, + 0.004471939988434315, + 0.005414011422544718, + 0.0028288604225963354, + 0.0027756309136748314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.05496400222182274 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle's architecture white with elegantly curved, multi-tiered roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the castle surrounded by a torrent of black, viscous mud instead of a traditional moat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky ominous, bruised purple and dark gray with thick dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002495034597814083 + ], + "opposite_logprob_tokens": [ + -6.002494812011719 + ], + "logprob_reward": -0.002495034597814083, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002495034597814083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42342424392700195, + 0.18987396359443665, + 0.06347259134054184, + 0.060447804629802704, + 0.06172121688723564, + 0.048225730657577515, + 0.0413598008453846, + 0.027751179412007332, + 0.04784266650676727, + 0.035880811512470245 + ], + "credit_map": [ + 0.0852338895201683, + 0.08225465565919876, + 0.05846480280160904, + 0.04165053367614746, + 0.036748696118593216, + 0.04739770665764809, + 0.03505055606365204, + 0.03662342205643654, + 0.03239823877811432, + 0.03201199695467949, + 0.030247898772358894, + 0.02205684408545494, + 0.016891473904252052, + 0.024652959778904915, + 0.012919231317937374, + 0.018695324659347534, + 0.00785598810762167, + 0.009060855023562908, + 0.010810067877173424, + 0.006618504878133535, + 0.006111488211899996, + 0.01140817441046238, + 0.00454408535733819, + 0.007063431665301323, + 0.01250772550702095, + 0.00786035880446434, + 0.00721650430932641, + 0.007257592398673296, + 0.005393964238464832, + 0.007829715497791767, + 0.0063285306096076965, + 0.006053417455404997, + 0.009792587719857693, + 0.010899835266172886, + 0.00787767767906189, + 0.0059223473072052, + 0.005844308994710445, + 0.010044815950095654, + 0.004766379948705435, + 0.006573262624442577, + 0.0056914989836514, + 0.00704721175134182, + 0.007836837321519852, + 0.006555282510817051, + 0.004295422695577145, + 0.006381375715136528, + 0.0036437322851270437, + 0.006774369161576033, + 0.0070677949115633965, + 0.005193586461246014, + 0.005818482954055071, + 0.0065501052886247635, + 0.003701645415276289, + 0.004615291953086853, + 0.0033250912092626095, + 0.005087802652269602, + 0.0036934835370630026, + 0.0034358601551502943, + 0.003442671149969101, + 0.004808112047612667, + 0.002686029067263007, + 0.003244766965508461, + 0.0028342576697468758, + 0.0036059985868632793, + 0.007125916425138712, + 0.007937782444059849, + 0.005418590269982815, + 0.006105718202888966, + 0.004267836455255747, + 0.006228580605238676, + 0.004999706521630287, + 0.005758537445217371, + 0.004659105557948351, + 0.005465882830321789, + 0.004491955041885376, + 0.005083031952381134, + 0.0034210700541734695, + 0.005068403203040361, + 0.0033532045781612396, + 0.004338154103606939 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.002495034597814083 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle's architecture white with elegantly curved, multi-tiered roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the castle surrounded by a torrent of black, viscous mud instead of a traditional moat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky ominous, bruised purple and dark gray with thick dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023289354518055916 + ], + "opposite_logprob_tokens": [ + -3.773289442062378 + ], + "logprob_reward": -0.023289354518055916, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023289354518055916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1928921341896057, + 0.0778561532497406, + 0.06906968355178833, + 0.09590042382478714, + 0.1301916092634201, + 0.12339015305042267, + 0.2035631388425827, + 0.04914276674389839, + 0.036850783973932266, + 0.021143095567822456 + ], + "credit_map": [ + 0.03122684732079506, + 0.03156924992799759, + 0.03414769470691681, + 0.020158367231488228, + 0.01663559302687645, + 0.024034714326262474, + 0.014764414168894291, + 0.02035524696111679, + 0.00865966733545065, + 0.010410415939986706, + 0.013669720850884914, + 0.009847261011600494, + 0.007550041191279888, + 0.00922316499054432, + 0.008418231271207333, + 0.010077648796141148, + 0.008500570431351662, + 0.007886124774813652, + 0.00944498274475336, + 0.011152767576277256, + 0.005906332284212112, + 0.00970187783241272, + 0.0055580963380634785, + 0.010918928310275078, + 0.01852111518383026, + 0.013907432556152344, + 0.00981072522699833, + 0.009768621996045113, + 0.011163750663399696, + 0.011299827136099339, + 0.012600663118064404, + 0.008828289806842804, + 0.018472185358405113, + 0.017428284510970116, + 0.014885541051626205, + 0.013838567771017551, + 0.011147967539727688, + 0.018403002992272377, + 0.01575901173055172, + 0.02025705575942993, + 0.013304113410413265, + 0.013400369323790073, + 0.023253871127963066, + 0.017456457018852234, + 0.016738133504986763, + 0.014867085963487625, + 0.010838904418051243, + 0.013531221076846123, + 0.022011253982782364, + 0.021871080622076988, + 0.04461134597659111, + 0.023645060136914253, + 0.019730692729353905, + 0.028664784505963326, + 0.01711737923324108, + 0.02591153234243393, + 0.004565120674669743, + 0.005614545661956072, + 0.0037274726200848818, + 0.004279673099517822, + 0.009000924415886402, + 0.008924837224185467, + 0.006021498702466488, + 0.007008696440607309, + 0.004846672527492046, + 0.004504224751144648, + 0.004056742414832115, + 0.003110471647232771, + 0.0056683081202209, + 0.006412137299776077, + 0.004792287014424801, + 0.0034599383361637592, + 0.0028571225702762604, + 0.0022303853183984756, + 0.002822099719196558, + 0.002757040550932288, + 0.003469686722382903, + 0.0030927255284041166, + 0.0015683748060837388, + 0.002345658838748932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.023289354518055916 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static wide shot reveals the majestic Himeji Castle, its pristine white walls and elegantly curved, multi-tiered roofs set against an ominous, bruised purple and dark gray sky, thick with dark clouds. Instead of a traditional moat, the castle is surrounded by a terrifying torrent of black, viscous mud, violently churning and carrying with it broken trees, rocks, and unrecognizable debris. This relentless torrent cascades down the hillside with unstoppable force, crashing against the castle's massive stone foundations with powerful waves. Thick, earthy sludge splashes, staining the pristine white plaster of the lower walls. The storm's dim, oppressive light casts deep shadows, accentuating the stark, terrifying contrast between the castle's tranquil architectural beauty and the raw, destructive force of the mudslide, creating a deeply unsettling and surreal atmosphere of impending doom.", + "seed": 6027, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_4/index_5548_seed_6027_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04378992966376245, + "vlm_logprob": -0.04378992966376245, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.04378992966376245, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle's architecture white with elegantly curved, multi-tiered roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the castle surrounded by a torrent of black, viscous mud instead of a traditional moat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky ominous, bruised purple and dark gray with thick dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1264146864414215 + ], + "opposite_logprob_tokens": [ + -2.2514147758483887 + ], + "logprob_reward": -0.1264146864414215, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1264146864414215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16174878180027008, + 0.06773263961076736, + 0.04277237504720688, + 0.21812765300273895, + 0.22334535419940948, + 0.031416893005371094, + 0.12296892702579498, + 0.06013983115553856, + 0.05171949788928032, + 0.020028064027428627 + ], + "credit_map": [ + 0.019833030179142952, + 0.021022457629442215, + 0.01810152269899845, + 0.021531404927372932, + 0.021774809807538986, + 0.016820503398776054, + 0.022134849801659584, + 0.020530201494693756, + 0.008242742158472538, + 0.00825680885463953, + 0.0076314834877848625, + 0.009606205858290195, + 0.009393349289894104, + 0.007433888502418995, + 0.008035172708332539, + 0.009132983162999153, + 0.004845190793275833, + 0.0058682747185230255, + 0.0032938725780695677, + 0.005559655372053385, + 0.0064670671708881855, + 0.0049041020683944225, + 0.005407629534602165, + 0.006426579784601927, + 0.02979469858109951, + 0.021587926894426346, + 0.01901932805776596, + 0.02807869389653206, + 0.018377648666501045, + 0.01776989921927452, + 0.03929798677563667, + 0.044201478362083435, + 0.02617359161376953, + 0.022030126303434372, + 0.019804619252681732, + 0.028138767927885056, + 0.0201105996966362, + 0.017693568021059036, + 0.044221315532922745, + 0.045172758400440216, + 0.003306832630187273, + 0.0041010635904967785, + 0.004173372872173786, + 0.0026046193670481443, + 0.0033317951019853354, + 0.004502392373979092, + 0.0037706438452005386, + 0.005626175086945295, + 0.013752524740993977, + 0.015878351405262947, + 0.0161838810890913, + 0.020018354058265686, + 0.01666361093521118, + 0.010446150787174702, + 0.01684417389333248, + 0.013181880116462708, + 0.006962551269680262, + 0.0068695805966854095, + 0.004474637098610401, + 0.008989901281893253, + 0.01131744310259819, + 0.00619141012430191, + 0.0077988822013139725, + 0.00753542548045516, + 0.006556242238730192, + 0.006888895761221647, + 0.00400528172031045, + 0.0054665859788656235, + 0.009882726706564426, + 0.005343243945389986, + 0.005866142455488443, + 0.007710383739322424, + 0.002192319370806217, + 0.0022568893618881702, + 0.0015656311297789216, + 0.003284949343651533, + 0.003614055458456278, + 0.0020853488240391016, + 0.003006983082741499, + 0.0020218882709741592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.1264146864414215 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle's architecture white with elegantly curved, multi-tiered roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the castle surrounded by a torrent of black, viscous mud instead of a traditional moat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky ominous, bruised purple and dark gray with thick dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026390083134174347 + ], + "opposite_logprob_tokens": [ + -3.6513900756835938 + ], + "logprob_reward": -0.026390083134174347, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026390083134174347, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20814195275306702, + 0.05558803677558899, + 0.10999215394258499, + 0.21088971197605133, + 0.14334140717983246, + 0.06469710171222687, + 0.026325251907110214, + 0.0715738832950592, + 0.05436309427022934, + 0.05508741736412048 + ], + "credit_map": [ + 0.029477640986442566, + 0.025273239240050316, + 0.022386012598872185, + 0.030664116144180298, + 0.023643074557185173, + 0.02044137753546238, + 0.02869030088186264, + 0.02756619267165661, + 0.006144162267446518, + 0.0062909601256251335, + 0.005110712721943855, + 0.00767028471454978, + 0.010553468018770218, + 0.006856138352304697, + 0.007405965123325586, + 0.005556347314268351, + 0.013609404675662518, + 0.012020586989820004, + 0.010064449161291122, + 0.01093831192702055, + 0.0237052533775568, + 0.010782141238451004, + 0.013910220004618168, + 0.014961792156100273, + 0.028051581233739853, + 0.02108577825129032, + 0.020646367222070694, + 0.025033196434378624, + 0.027299420908093452, + 0.017939411103725433, + 0.03385930135846138, + 0.03697466105222702, + 0.014591039158403873, + 0.01592070236802101, + 0.01282643061131239, + 0.019377168267965317, + 0.022651396691799164, + 0.014805227518081665, + 0.022942936047911644, + 0.020226500928401947, + 0.007511884905397892, + 0.00704943249002099, + 0.006029240787029266, + 0.010803665965795517, + 0.008976347744464874, + 0.005538912955671549, + 0.009736457839608192, + 0.009051159024238586, + 0.003229835070669651, + 0.0038006159011274576, + 0.001959152054041624, + 0.0031518866308033466, + 0.00409857090562582, + 0.003138221800327301, + 0.0038969586603343487, + 0.0030500106513500214, + 0.00680116843432188, + 0.0073255496099591255, + 0.007159517612308264, + 0.015280475839972496, + 0.008684530854225159, + 0.006407614331692457, + 0.011144501157104969, + 0.008770531043410301, + 0.006139041390269995, + 0.007237771525979042, + 0.005142301321029663, + 0.0076309191063046455, + 0.010996800847351551, + 0.005125692579895258, + 0.007458316627889872, + 0.004632251802831888, + 0.005954003892838955, + 0.006381168495863676, + 0.0047793276607990265, + 0.010265464894473553, + 0.00974581204354763, + 0.00519386725500226, + 0.008045203052461147, + 0.004722566343843937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.026390083134174347 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle's architecture white with elegantly curved, multi-tiered roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the castle surrounded by a torrent of black, viscous mud instead of a traditional moat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky ominous, bruised purple and dark gray with thick dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306979477405548 + ], + "opposite_logprob_tokens": [ + -3.168069839477539 + ], + "logprob_reward": -0.04306979477405548, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306979477405548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34136343002319336, + 0.287641316652298, + 0.05622883141040802, + 0.04003217816352844, + 0.10038578510284424, + 0.03261398524045944, + 0.0406016930937767, + 0.0196381863206625, + 0.05468286946415901, + 0.026811743155121803 + ], + "credit_map": [ + 0.05110640451312065, + 0.04952685534954071, + 0.03316688910126686, + 0.04544399678707123, + 0.03255554661154747, + 0.0422908216714859, + 0.04249459132552147, + 0.04477832838892937, + 0.02916029281914234, + 0.036591317504644394, + 0.03612084686756134, + 0.04463275521993637, + 0.03375212103128433, + 0.035379402339458466, + 0.035712242126464844, + 0.03629235550761223, + 0.005838742014020681, + 0.007694814819842577, + 0.005111184902489185, + 0.0058921948075294495, + 0.010073736310005188, + 0.006430296692997217, + 0.0072072334587574005, + 0.007980630733072758, + 0.004027791321277618, + 0.004817686975002289, + 0.0038179359398782253, + 0.005776393227279186, + 0.004316675942391157, + 0.005028229206800461, + 0.006224751006811857, + 0.00602271594107151, + 0.012168528512120247, + 0.010961013846099377, + 0.009578732773661613, + 0.012042996473610401, + 0.009608122520148754, + 0.007968650199472904, + 0.018655234947800636, + 0.019402511417865753, + 0.00330043095164001, + 0.004182068631052971, + 0.004481562413275242, + 0.0035355035215616226, + 0.0040077404119074345, + 0.0051273261196911335, + 0.0030337844509631395, + 0.004945568740367889, + 0.004371061455458403, + 0.005278244614601135, + 0.0030982072930783033, + 0.006032672245055437, + 0.006122366059571505, + 0.004172641783952713, + 0.006162128411233425, + 0.005364370532333851, + 0.002117560477927327, + 0.002371092326939106, + 0.002914208685979247, + 0.003037702990695834, + 0.0031620326917618513, + 0.002007960109040141, + 0.002062636660411954, + 0.0019649937748908997, + 0.005884433165192604, + 0.006294273305684328, + 0.005222267936915159, + 0.010819945484399796, + 0.009586983360350132, + 0.00518529349938035, + 0.007519707083702087, + 0.004169972147792578, + 0.002920751925557852, + 0.0028023244813084602, + 0.0023996655363589525, + 0.004327402450144291, + 0.005157863721251488, + 0.0025440629106014967, + 0.004043858963996172, + 0.0026158164255321026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.04306979477405548 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle's architecture white with elegantly curved, multi-tiered roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the castle surrounded by a torrent of black, viscous mud instead of a traditional moat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky ominous, bruised purple and dark gray with thick dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002492894185706973 + ], + "opposite_logprob_tokens": [ + -6.002492904663086 + ], + "logprob_reward": -0.002492894185706973, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002492894185706973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4603881239891052, + 0.10718588531017303, + 0.06839567422866821, + 0.0376436710357666, + 0.11278901994228363, + 0.05477943643927574, + 0.03545796871185303, + 0.044227126985788345, + 0.06053139269351959, + 0.018601683899760246 + ], + "credit_map": [ + 0.06711973994970322, + 0.06633859872817993, + 0.055836282670497894, + 0.06536795198917389, + 0.043211061507463455, + 0.07283858954906464, + 0.044014547020196915, + 0.04566136747598648, + 0.013547150418162346, + 0.017647696658968925, + 0.009201465174555779, + 0.01330226007848978, + 0.013592069037258625, + 0.01875864341855049, + 0.010816473513841629, + 0.010320133529603481, + 0.00786267127841711, + 0.010196053422987461, + 0.0054817101918160915, + 0.007890576496720314, + 0.008341731503605843, + 0.012974292039871216, + 0.0073854257352650166, + 0.00826321355998516, + 0.005975032225251198, + 0.0070417458191514015, + 0.002728100400418043, + 0.004111703485250473, + 0.004050810355693102, + 0.006858526729047298, + 0.0031646001152694225, + 0.0037131502758711576, + 0.017560072243213654, + 0.01964462921023369, + 0.007382987067103386, + 0.012758632190525532, + 0.011768314987421036, + 0.015704965218901634, + 0.013552344404160976, + 0.014417069964110851, + 0.005349501967430115, + 0.006496478337794542, + 0.005586065351963043, + 0.00781662855297327, + 0.00886300764977932, + 0.007260502316057682, + 0.007023471873253584, + 0.0063837794587016106, + 0.004222902934998274, + 0.004192148335278034, + 0.002753995358943939, + 0.00485078664496541, + 0.006336777471005917, + 0.004049085080623627, + 0.0050004590302705765, + 0.004051811993122101, + 0.004191390238702297, + 0.005383883137255907, + 0.005269299726933241, + 0.0073278965428471565, + 0.007540833204984665, + 0.004652558360248804, + 0.0060912542976439, + 0.0037700084503740072, + 0.006925669964402914, + 0.007158844731748104, + 0.005035135429352522, + 0.010686557739973068, + 0.01022517029196024, + 0.005739152431488037, + 0.009397025220096111, + 0.005363840609788895, + 0.0019798593129962683, + 0.0026750313118100166, + 0.0019750201608985662, + 0.0025958139449357986, + 0.0026213068049401045, + 0.0022414452396333218, + 0.0023845264222472906, + 0.0021286807022988796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.002492894185706973 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle's architecture white with elegantly curved, multi-tiered roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the castle surrounded by a torrent of black, viscous mud instead of a traditional moat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky ominous, bruised purple and dark gray with thick dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058218978345394 + ], + "opposite_logprob_tokens": [ + -3.8955821990966797 + ], + "logprob_reward": -0.02058218978345394, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058218978345394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38236305117607117, + 0.0871226042509079, + 0.06817793846130371, + 0.07181518524885178, + 0.10033760219812393, + 0.06423822045326233, + 0.06728759407997131, + 0.05840066075325012, + 0.05595482513308525, + 0.04430234059691429 + ], + "credit_map": [ + 0.03418610990047455, + 0.03901036083698273, + 0.034172192215919495, + 0.06315568834543228, + 0.04964768514037132, + 0.040641333907842636, + 0.05974573642015457, + 0.06180393323302269, + 0.00969500932842493, + 0.010727018117904663, + 0.009389884769916534, + 0.014961429871618748, + 0.008842692710459232, + 0.012266334146261215, + 0.011716067790985107, + 0.009524160996079445, + 0.007033444941043854, + 0.008553235791623592, + 0.006559954956173897, + 0.008834661915898323, + 0.013401186093688011, + 0.006307629868388176, + 0.008262029848992825, + 0.009225788526237011, + 0.00902264378964901, + 0.009225479327142239, + 0.006208761595189571, + 0.007612191140651703, + 0.014598753303289413, + 0.007290505338460207, + 0.010365265421569347, + 0.007491578347980976, + 0.009883958846330643, + 0.010012446902692318, + 0.009194206446409225, + 0.01405386719852686, + 0.011115223169326782, + 0.011235604993999004, + 0.018598539754748344, + 0.01624373160302639, + 0.007592794485390186, + 0.0074833836406469345, + 0.007178028579801321, + 0.009733307175338268, + 0.009668417274951935, + 0.00601158058270812, + 0.009276362136006355, + 0.007294336799532175, + 0.006194337736815214, + 0.006981357466429472, + 0.007582408841699362, + 0.012239924632012844, + 0.008125899359583855, + 0.007685504388064146, + 0.009436015971004963, + 0.009042133577167988, + 0.006587023381143808, + 0.007023755926638842, + 0.005165064707398415, + 0.007934769615530968, + 0.012481555342674255, + 0.006190643645823002, + 0.006533025298267603, + 0.006484817247837782, + 0.006009525619447231, + 0.007164078298956156, + 0.005484071560204029, + 0.008260438218712807, + 0.009670408442616463, + 0.0055621834471821785, + 0.007527745328843594, + 0.00627636443823576, + 0.004687005188316107, + 0.006250265520066023, + 0.004186347126960754, + 0.007695995271205902, + 0.006884763482958078, + 0.004633681382983923, + 0.0063194818794727325, + 0.00364479492418468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.02058218978345394 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static wide shot reveals the majestic Himeji Castle, its pristine white walls and elegantly curved, multi-tiered roofs set against an ominous, bruised purple and dark gray sky, thick with dark clouds. Instead of a traditional moat, the castle is surrounded by a terrifying torrent of black, viscous mud, violently churning and carrying with it broken trees, rocks, and unrecognizable debris. This relentless torrent cascades down the hillside with unstoppable force, crashing against the castle's massive stone foundations with powerful waves. Thick, earthy sludge splashes, staining the pristine white plaster of the lower walls. The storm's dim, oppressive light casts deep shadows, accentuating the stark, terrifying contrast between the castle's tranquil architectural beauty and the raw, destructive force of the mudslide, creating a deeply unsettling and surreal atmosphere of impending doom.", + "seed": 6028, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_4/index_5548_seed_6028_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06415132489055395, + "vlm_logprob": -0.06415132489055395, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.06415132489055395, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle's architecture white with elegantly curved, multi-tiered roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the castle surrounded by a torrent of black, viscous mud instead of a traditional moat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky ominous, bruised purple and dark gray with thick dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09675386548042297 + ], + "opposite_logprob_tokens": [ + -2.5967538356781006 + ], + "logprob_reward": -0.09675386548042297, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.09675386548042297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20169341564178467, + 0.17607346177101135, + 0.09437796473503113, + 0.11665491759777069, + 0.10248397290706635, + 0.02867075242102146, + 0.08589809387922287, + 0.02988349087536335, + 0.11994814872741699, + 0.04431577026844025 + ], + "credit_map": [ + 0.0289311520755291, + 0.028169913217425346, + 0.022344907745718956, + 0.04127330332994461, + 0.023504246026277542, + 0.018793189898133278, + 0.02741626836359501, + 0.011260440573096275, + 0.022841550409793854, + 0.027680790051817894, + 0.018537480384111404, + 0.03148407116532326, + 0.026594508439302444, + 0.016293134540319443, + 0.021392526105046272, + 0.011249399743974209, + 0.012439267709851265, + 0.013314601965248585, + 0.01256426889449358, + 0.014497464522719383, + 0.013211190700531006, + 0.010526478290557861, + 0.009305099956691265, + 0.008519592694938183, + 0.020096557214856148, + 0.015560382045805454, + 0.008871141821146011, + 0.015169045887887478, + 0.02490946277976036, + 0.012523609213531017, + 0.011682866141200066, + 0.007841846905648708, + 0.015241734683513641, + 0.015247123315930367, + 0.0076563069596886635, + 0.014089739881455898, + 0.018384383991360664, + 0.01003956701606512, + 0.015438294038176537, + 0.006386806722730398, + 0.0038810474798083305, + 0.0033398650120943785, + 0.0034028347581624985, + 0.005706580355763435, + 0.0032557654194533825, + 0.0032959836535155773, + 0.0027965044137090445, + 0.0029921724926680326, + 0.012986121699213982, + 0.010577321983873844, + 0.013427283614873886, + 0.016298316419124603, + 0.007461346685886383, + 0.008655312471091747, + 0.008698350749909878, + 0.007794043514877558, + 0.005557303782552481, + 0.004310061223804951, + 0.0023575620725750923, + 0.004490376450121403, + 0.0036370791494846344, + 0.003511173417791724, + 0.0032381194178014994, + 0.002781816991046071, + 0.01660163141787052, + 0.02377145178616047, + 0.011848547495901585, + 0.015394099988043308, + 0.014587310142815113, + 0.013376710005104542, + 0.015774697065353394, + 0.008593698032200336, + 0.006867764052003622, + 0.008650777861475945, + 0.004439737647771835, + 0.007281404919922352, + 0.0052660126239061356, + 0.004387142136693001, + 0.004244531504809856, + 0.0031783985905349255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.09675386548042297 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle's architecture white with elegantly curved, multi-tiered roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the castle surrounded by a torrent of black, viscous mud instead of a traditional moat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky ominous, bruised purple and dark gray with thick dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03814342990517616 + ], + "opposite_logprob_tokens": [ + -3.2881433963775635 + ], + "logprob_reward": -0.03814342990517616, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03814342990517616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4467504620552063, + 0.10556021332740784, + 0.041559793055057526, + 0.2959214746952057, + 0.020752299576997757, + 0.021090900525450706, + 0.02186424471437931, + 0.013072884641587734, + 0.02684672549366951, + 0.006580960005521774 + ], + "credit_map": [ + 0.06796642392873764, + 0.07166020572185516, + 0.04705410450696945, + 0.07031464576721191, + 0.08274301141500473, + 0.04010565206408501, + 0.044707585126161575, + 0.02219885215163231, + 0.01431814394891262, + 0.017183968797326088, + 0.012405729852616787, + 0.01733151264488697, + 0.015481731854379177, + 0.009791475720703602, + 0.012289581820368767, + 0.006758080795407295, + 0.0054735783487558365, + 0.007013375870883465, + 0.004866837989538908, + 0.006860724184662104, + 0.005258532240986824, + 0.0036223172210156918, + 0.004878766369074583, + 0.0035856664180755615, + 0.04613325744867325, + 0.049548469483852386, + 0.026976650580763817, + 0.05133146792650223, + 0.040686141699552536, + 0.023810438811779022, + 0.03621724992990494, + 0.021217811852693558, + 0.004087132401764393, + 0.0028879791498184204, + 0.001974885817617178, + 0.003442605957388878, + 0.002428738633170724, + 0.001733010751195252, + 0.0027465359307825565, + 0.001451413962058723, + 0.0039176479913294315, + 0.0035512743052095175, + 0.001513599418103695, + 0.0029945930000394583, + 0.003166097914800048, + 0.0017732464475557208, + 0.002469113562256098, + 0.001705329748801887, + 0.003030578140169382, + 0.0032248625066131353, + 0.002607843140140176, + 0.0033863121643662453, + 0.0034448024816811085, + 0.0024062676820904016, + 0.0019413511035963893, + 0.001822228659875691, + 0.002226912183687091, + 0.0021133103873580694, + 0.0011215012054890394, + 0.0021702649537473917, + 0.002097473945468664, + 0.0010407628724351525, + 0.0014117701211944222, + 0.0008908906602300704, + 0.004201458767056465, + 0.004092920105904341, + 0.003319465322420001, + 0.004447754472494125, + 0.0035878510680049658, + 0.0023319569882005453, + 0.002593005308881402, + 0.0022723162546753883, + 0.0011204921174794436, + 0.0011116780806332827, + 0.0006849854835309088, + 0.0007536823977716267, + 0.0006331541808322072, + 0.0009203535155393183, + 0.0007465663948096335, + 0.0006100484752096236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.03814342990517616 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle's architecture white with elegantly curved, multi-tiered roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the castle surrounded by a torrent of black, viscous mud instead of a traditional moat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky ominous, bruised purple and dark gray with thick dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16036881506443024 + ], + "opposite_logprob_tokens": [ + -1.910368800163269 + ], + "logprob_reward": -0.16036881506443024, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16036881506443024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4815690219402313, + 0.04272834211587906, + 0.07498179376125336, + 0.18011239171028137, + 0.037516999989748, + 0.05856891721487045, + 0.04077170416712761, + 0.021523533388972282, + 0.027785878628492355, + 0.03444143012166023 + ], + "credit_map": [ + 0.08510100096464157, + 0.07075610011816025, + 0.04786358401179314, + 0.06962667405605316, + 0.08544614911079407, + 0.05466976761817932, + 0.04098605737090111, + 0.02711966261267662, + 0.006691502872854471, + 0.006181754171848297, + 0.003956277389079332, + 0.00611708452925086, + 0.00637318380177021, + 0.00542245851829648, + 0.0043161408975720406, + 0.003669938538223505, + 0.010661431588232517, + 0.013736920431256294, + 0.00788423977792263, + 0.013439484871923923, + 0.007655715569853783, + 0.008756743744015694, + 0.006369101349264383, + 0.00647816201671958, + 0.031534966081380844, + 0.026818882673978806, + 0.01625688001513481, + 0.029027245938777924, + 0.02505422569811344, + 0.017068393528461456, + 0.020765600726008415, + 0.013586205430328846, + 0.008013777434825897, + 0.0066688465885818005, + 0.001717928098514676, + 0.003951042424887419, + 0.007082739844918251, + 0.0031835369300097227, + 0.004113513045012951, + 0.002785617485642433, + 0.010951802134513855, + 0.008894728496670723, + 0.005792752839624882, + 0.010652998462319374, + 0.007576643023639917, + 0.004704207181930542, + 0.005449633579701185, + 0.004546153359115124, + 0.006010503973811865, + 0.0061228531412780285, + 0.004954449832439423, + 0.006657004822045565, + 0.005107598844915628, + 0.004436643328517675, + 0.004084030166268349, + 0.003398620057851076, + 0.0034285529982298613, + 0.0031972366850823164, + 0.0023503119591623545, + 0.004025959875434637, + 0.002160885604098439, + 0.0022878411691635847, + 0.0021622939966619015, + 0.0019104512175545096, + 0.004560668952763081, + 0.004181270953267813, + 0.002963320119306445, + 0.004428465384989977, + 0.003665926633402705, + 0.0027954818215221167, + 0.003025274956598878, + 0.00216546724550426, + 0.005000985227525234, + 0.005838650278747082, + 0.0036926455795764923, + 0.006447385996580124, + 0.0037295911461114883, + 0.003417723812162876, + 0.0035850340500473976, + 0.002729410771280527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.16036881506443024 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle's architecture white with elegantly curved, multi-tiered roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the castle surrounded by a torrent of black, viscous mud instead of a traditional moat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky ominous, bruised purple and dark gray with thick dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022011594846844673 + ], + "opposite_logprob_tokens": [ + -6.127201080322266 + ], + "logprob_reward": -0.0022011594846844673, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022011594846844673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5084682703018188, + 0.13712048530578613, + 0.06434440612792969, + 0.06742580980062485, + 0.05140868201851845, + 0.04184829443693161, + 0.05186581239104271, + 0.032767072319984436, + 0.027015967294573784, + 0.017735198140144348 + ], + "credit_map": [ + 0.09371507167816162, + 0.06618720293045044, + 0.055044155567884445, + 0.07452036440372467, + 0.08077031373977661, + 0.06571855396032333, + 0.04351424798369408, + 0.02899835631251335, + 0.018899304792284966, + 0.019522344693541527, + 0.013985896483063698, + 0.020203113555908203, + 0.02316712960600853, + 0.017885679379105568, + 0.016103526577353477, + 0.0073534902185201645, + 0.009320137090981007, + 0.008317570202052593, + 0.006222594529390335, + 0.008152213878929615, + 0.0106773367151618, + 0.010363117791712284, + 0.006889536045491695, + 0.004401896148920059, + 0.013009713031351566, + 0.00728059746325016, + 0.006597870495170355, + 0.009779645130038261, + 0.0070702568627893925, + 0.010107768699526787, + 0.007438527885824442, + 0.006141426507383585, + 0.009585314430296421, + 0.007164721842855215, + 0.003811680246144533, + 0.008099963888525963, + 0.006287652999162674, + 0.006442568730562925, + 0.0059405043721199036, + 0.0040762764401733875, + 0.0078187957406044, + 0.003921446856111288, + 0.00414384575560689, + 0.006645479239523411, + 0.005336352623999119, + 0.006743201054632664, + 0.003596862079575658, + 0.003642312018200755, + 0.0075834887102246284, + 0.007753373589366674, + 0.006806833203881979, + 0.008886288851499557, + 0.0073821935802698135, + 0.0036891705822199583, + 0.005181333050131798, + 0.004583130124956369, + 0.005050663836300373, + 0.004673986695706844, + 0.0027324832044541836, + 0.0034447021316736937, + 0.005510029848664999, + 0.005142123904079199, + 0.0037778650876134634, + 0.0024352173786610365, + 0.0042793904431164265, + 0.0035526305437088013, + 0.003459599334746599, + 0.0035629526246339083, + 0.003788986010476947, + 0.003100080881267786, + 0.0026488667353987694, + 0.0026234614197164774, + 0.002850694814696908, + 0.0022975800093263388, + 0.0016523591475561261, + 0.0028225379064679146, + 0.002117685740813613, + 0.0022415684070438147, + 0.002108298009261489, + 0.0016444741049781442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.0022011594846844673 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle's architecture white with elegantly curved, multi-tiered roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the castle surrounded by a torrent of black, viscous mud instead of a traditional moat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky ominous, bruised purple and dark gray with thick dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023289354518055916 + ], + "opposite_logprob_tokens": [ + -3.773289442062378 + ], + "logprob_reward": -0.023289354518055916, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023289354518055916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08137054741382599, + 0.07828545570373535, + 0.12986014783382416, + 0.1762656718492508, + 0.12056642025709152, + 0.0922444686293602, + 0.04466290399432182, + 0.06773561239242554, + 0.19315144419670105, + 0.015857243910431862 + ], + "credit_map": [ + 0.011730463244020939, + 0.0166480652987957, + 0.007887003012001514, + 0.010138014331459999, + 0.012440592050552368, + 0.00788932666182518, + 0.010783931240439415, + 0.003853149712085724, + 0.012186110019683838, + 0.013378655537962914, + 0.008025098592042923, + 0.012887813150882721, + 0.01225601602345705, + 0.004673159681260586, + 0.009728997945785522, + 0.005149607080966234, + 0.018672442063689232, + 0.02531975507736206, + 0.012696580961346626, + 0.021296413615345955, + 0.019013702869415283, + 0.01058527547866106, + 0.013479483313858509, + 0.008796509355306625, + 0.028004858642816544, + 0.026827409863471985, + 0.01673070155084133, + 0.02893701009452343, + 0.02765006199479103, + 0.015458773821592331, + 0.0201956108212471, + 0.012461249716579914, + 0.021514462307095528, + 0.019193219020962715, + 0.00991808995604515, + 0.017803430557250977, + 0.01986905001103878, + 0.010183926671743393, + 0.013771002180874348, + 0.008313244208693504, + 0.01643083058297634, + 0.013952835462987423, + 0.009786978363990784, + 0.015215245075523853, + 0.012286694720387459, + 0.008616099134087563, + 0.008952122181653976, + 0.0070036654360592365, + 0.006065591238439083, + 0.006373546086251736, + 0.005888459272682667, + 0.0054463488049805164, + 0.006164971273392439, + 0.004854215309023857, + 0.0058606998063623905, + 0.004009074065834284, + 0.009227745234966278, + 0.013702997006475925, + 0.005918263457715511, + 0.00874614343047142, + 0.009350364096462727, + 0.007297935895621777, + 0.008380602113902569, + 0.0051115695387125015, + 0.029116787016391754, + 0.03356958553195, + 0.021398933604359627, + 0.03282080218195915, + 0.019272640347480774, + 0.019196750596165657, + 0.021482273936271667, + 0.01629367657005787, + 0.0023677649442106485, + 0.0024848999455571175, + 0.0016780371079221368, + 0.0025225270073860884, + 0.002303477143868804, + 0.0015601429622620344, + 0.0016256138915196061, + 0.001314780325628817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 2 + ], + "selected_reward": -0.023289354518055916 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_5/index_tensor([9255, 9255])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_5/index_tensor([9255, 9255])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..338a1f7ea9a5868b00200b2cde505b02b5854c95 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_5/index_tensor([9255, 9255])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "striking low-angle, close-up tracking shot follows a majestic, black-smooth, muscular horse, galloping with unstoppable force across the ancient Great Wall. The camera moves in sync with the horse's rapid pace, never losing focus on its swift posture and tumbling hooves, its long mane and tail dancing wildly in the wind. The horse is adorned with a pristine bronze harness and a dark red carved saddle. Each hoofbeat treads powerfully against the weathered, rough-textured gray bricks, a steady and rhythmic movement that stirs up a fine dust cloud, as if resonating with the echoes of history. The backdrop is the golden afterglow of the setting sun, tinting the battlements and beacon towers winding into the distance with a warm hue and framing the horse's silhouette with a dazzling golden edge. The entire scene exudes a rich, classical Chinese martial arts style, with powerful and majestic tones, imbued with a sense of loneliness, tragedy, and unstoppable power, and an epic atmosphere.", + "seed": 32215, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_5/index_9255_seed_32215_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03602800266817212, + "vlm_logprob": -0.03602800266817212, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.03602800266817212, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black, smooth, and muscular horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient Great Wall with weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse adorned with a pristine bronze harness and a dark red carved saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by the golden afterglow of a setting sun with a warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up tracking shot that moves in sync with the horse's rapid pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015032579191029072 + ], + "opposite_logprob_tokens": [ + -4.640032768249512 + ], + "logprob_reward": -0.015032579191029072, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.015032579191029072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16899003088474274, + 0.10836894065141678, + 0.032424747943878174, + 0.08397967368364334, + 0.061031218618154526, + 0.24962885677814484, + 0.06168198212981224, + 0.0824178159236908, + 0.08488388359546661, + 0.06659282743930817 + ], + "credit_map": [ + 0.01566757820546627, + 0.01994893327355385, + 0.016377000138163567, + 0.021525518968701363, + 0.023473506793379784, + 0.02564173750579357, + 0.02316337265074253, + 0.0231923945248127, + 0.00768438633531332, + 0.01475007925182581, + 0.016765456646680832, + 0.014906509779393673, + 0.01204816997051239, + 0.013166101649403572, + 0.014646364375948906, + 0.014401879161596298, + 0.003979107365012169, + 0.0029508201405406, + 0.0045996461994946, + 0.004744087345898151, + 0.0022840320598334074, + 0.0044142561964690685, + 0.004281166009604931, + 0.0051716347225010395, + 0.011626620776951313, + 0.010503149591386318, + 0.00601372541859746, + 0.008559729903936386, + 0.004145248327404261, + 0.01414623111486435, + 0.006128944922238588, + 0.022856028750538826, + 0.005726405419409275, + 0.014031391590833664, + 0.008202971890568733, + 0.005978438537567854, + 0.01012473739683628, + 0.005022365599870682, + 0.005430500488728285, + 0.006514410022646189, + 0.05322510376572609, + 0.0296921506524086, + 0.02639673836529255, + 0.05003584921360016, + 0.02502220869064331, + 0.020364904776215553, + 0.014420758001506329, + 0.030471159145236015, + 0.0081245182082057, + 0.00713338190689683, + 0.005460911430418491, + 0.007961624301970005, + 0.008699914440512657, + 0.007778477389365435, + 0.009523921646177769, + 0.006999235600233078, + 0.007822584360837936, + 0.007091844454407692, + 0.007513965480029583, + 0.01315182726830244, + 0.010371183045208454, + 0.00973428413271904, + 0.014718816615641117, + 0.012013319879770279, + 0.010888400487601757, + 0.012685424648225307, + 0.010576965287327766, + 0.012944206595420837, + 0.009435422718524933, + 0.010069121606647968, + 0.009048436768352985, + 0.009235900826752186, + 0.011406137607991695, + 0.01010991632938385, + 0.014003638178110123, + 0.009301260113716125, + 0.006040734238922596, + 0.0045209042727947235, + 0.00616095494478941, + 0.0050492859445512295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.015032579191029072 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black, smooth, and muscular horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient Great Wall with weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse adorned with a pristine bronze harness and a dark red carved saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by the golden afterglow of a setting sun with a warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up tracking shot that moves in sync with the horse's rapid pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635931596159935 + ], + "opposite_logprob_tokens": [ + -3.6513593196868896 + ], + "logprob_reward": -0.02635931596159935, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635931596159935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15227225422859192, + 0.1278543323278427, + 0.07884124666452408, + 0.09955009818077087, + 0.0846569612622261, + 0.06632030010223389, + 0.04934436082839966, + 0.11936245113611221, + 0.17841200530529022, + 0.04338600113987923 + ], + "credit_map": [ + 0.01638752967119217, + 0.020179038867354393, + 0.012124106287956238, + 0.018705276772379875, + 0.019545070827007294, + 0.01985577493906021, + 0.023004168644547462, + 0.02247127890586853, + 0.011406553909182549, + 0.0253180880099535, + 0.013178194873034954, + 0.01578010432422161, + 0.012903579510748386, + 0.008841408416628838, + 0.01818368397653103, + 0.022242726758122444, + 0.006847151089459658, + 0.00576019287109375, + 0.0068323081359267235, + 0.010749144479632378, + 0.005742750130593777, + 0.017050135880708694, + 0.012469070963561535, + 0.013390492647886276, + 0.01277703046798706, + 0.011082110926508904, + 0.007065319921821356, + 0.010541816242039204, + 0.005145198665559292, + 0.018363283947110176, + 0.00662416685372591, + 0.027951175346970558, + 0.011047215200960636, + 0.013263902626931667, + 0.006462418474256992, + 0.008100136183202267, + 0.009030359797179699, + 0.0145101398229599, + 0.0052250283770263195, + 0.017017759382724762, + 0.010235029272735119, + 0.009928899817168713, + 0.005876017268747091, + 0.0059685781598091125, + 0.008797865360975266, + 0.011070536449551582, + 0.007040821947157383, + 0.007402552291750908, + 0.004883387591689825, + 0.0061933486722409725, + 0.004778704605996609, + 0.004638464655727148, + 0.010159028694033623, + 0.005296221002936363, + 0.006721052806824446, + 0.006674155592918396, + 0.01093930285423994, + 0.010146597400307655, + 0.009013748727738857, + 0.016144337132573128, + 0.018668096512556076, + 0.01374184899032116, + 0.023548506200313568, + 0.017160022631287575, + 0.023229733109474182, + 0.02009705826640129, + 0.024981429800391197, + 0.028483789414167404, + 0.01870903931558132, + 0.018444089218974113, + 0.019205065444111824, + 0.025261806324124336, + 0.006559748202562332, + 0.0057118427939713, + 0.007902700453996658, + 0.00566865224391222, + 0.004509727004915476, + 0.003917995374649763, + 0.004422382917255163, + 0.004692952148616314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.02635931596159935 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black, smooth, and muscular horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient Great Wall with weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse adorned with a pristine bronze harness and a dark red carved saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by the golden afterglow of a setting sun with a warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up tracking shot that moves in sync with the horse's rapid pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703566253185272 + ], + "opposite_logprob_tokens": [ + -2.127035617828369 + ], + "logprob_reward": -0.12703566253185272, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703566253185272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21712784469127655, + 0.06134069710969925, + 0.0507703460752964, + 0.09452744573354721, + 0.15177375078201294, + 0.23422472178936005, + 0.0556294247508049, + 0.051824409514665604, + 0.04631298780441284, + 0.03646835684776306 + ], + "credit_map": [ + 0.02176804468035698, + 0.025115719065070152, + 0.022189030423760414, + 0.027690134942531586, + 0.029477931559085846, + 0.030987633392214775, + 0.026401877403259277, + 0.03349747881293297, + 0.005651467014104128, + 0.00813433900475502, + 0.005010885186493397, + 0.0062770722433924675, + 0.005629415158182383, + 0.012426696717739105, + 0.010198824107646942, + 0.008011993020772934, + 0.004425440914928913, + 0.007431945763528347, + 0.005128365475684404, + 0.007840815931558609, + 0.003669626545161009, + 0.00881909765303135, + 0.005900187883526087, + 0.007554869167506695, + 0.013103626668453217, + 0.01935950480401516, + 0.007095978129655123, + 0.011528979055583477, + 0.006002711597830057, + 0.013233774341642857, + 0.008253824897110462, + 0.01594904251396656, + 0.014633391983807087, + 0.03029223531484604, + 0.02070551924407482, + 0.014311227016150951, + 0.0266127847135067, + 0.012211905792355537, + 0.01761282980442047, + 0.015393858775496483, + 0.024292398244142532, + 0.05337153747677803, + 0.021362071856856346, + 0.028909124433994293, + 0.03612200915813446, + 0.027696950361132622, + 0.018707361072301865, + 0.023763267323374748, + 0.00585648650303483, + 0.0061011966317892075, + 0.005132345017045736, + 0.0062521533109247684, + 0.0109282610937953, + 0.006628529168665409, + 0.00852848682552576, + 0.006201967131346464, + 0.005474488250911236, + 0.0058980667963624, + 0.0030527482740581036, + 0.00614197039976716, + 0.013026244007050991, + 0.0051036435179412365, + 0.008533400483429432, + 0.004593850113451481, + 0.006093157455325127, + 0.004687709733843803, + 0.005530631635338068, + 0.005253848619759083, + 0.005498844664543867, + 0.00500332610681653, + 0.006112048402428627, + 0.0081334188580513, + 0.0057852971367537975, + 0.0036039003171026707, + 0.006045200861990452, + 0.006470118183642626, + 0.003143651643767953, + 0.002426982158794999, + 0.003779398510232568, + 0.005213807336986065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.12703566253185272 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black, smooth, and muscular horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient Great Wall with weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse adorned with a pristine bronze harness and a dark red carved saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by the golden afterglow of a setting sun with a warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up tracking shot that moves in sync with the horse's rapid pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019378946162760258 + ], + "opposite_logprob_tokens": [ + -6.2519378662109375 + ], + "logprob_reward": -0.0019378946162760258, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019378946162760258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.156138077378273, + 0.07617507129907608, + 0.21631304919719696, + 0.0666671171784401, + 0.035014085471630096, + 0.09773506969213486, + 0.04846329614520073, + 0.20685726404190063, + 0.039941903203725815, + 0.056695129722356796 + ], + "credit_map": [ + 0.015362132340669632, + 0.016171496361494064, + 0.015512874349951744, + 0.020213274285197258, + 0.02078590728342533, + 0.0240483358502388, + 0.022014832124114037, + 0.0220292117446661, + 0.005430641118437052, + 0.01216813363134861, + 0.005957189481705427, + 0.009327935054898262, + 0.008078196085989475, + 0.010799749754369259, + 0.011923871003091335, + 0.012489342130720615, + 0.015700986608862877, + 0.02149808406829834, + 0.03156103566288948, + 0.03011731617152691, + 0.01719231903553009, + 0.038888975977897644, + 0.02504199557006359, + 0.03631232678890228, + 0.007064371835440397, + 0.007383907679468393, + 0.004947342909872532, + 0.006081798113882542, + 0.004796098917722702, + 0.014327678829431534, + 0.004986809100955725, + 0.017079109326004982, + 0.003148044226691127, + 0.008492168970406055, + 0.003971712198108435, + 0.003645532298833132, + 0.003709248034283519, + 0.0036914460361003876, + 0.003173138713464141, + 0.005182791035622358, + 0.019121697172522545, + 0.012953558936715126, + 0.00946143176406622, + 0.013288706541061401, + 0.01089321356266737, + 0.011341831646859646, + 0.008017108775675297, + 0.012657512910664082, + 0.0046433620154857635, + 0.005489904433488846, + 0.004519365727901459, + 0.008088621310889721, + 0.008535942994058132, + 0.004170711152255535, + 0.008513754233717918, + 0.004501629155129194, + 0.03301648050546646, + 0.02542351558804512, + 0.010687172412872314, + 0.010737542994320393, + 0.07246570289134979, + 0.02098163589835167, + 0.02274523675441742, + 0.010799947194755077, + 0.004787318408489227, + 0.00500165531411767, + 0.005868940148502588, + 0.004846587777137756, + 0.004487087484449148, + 0.004817532375454903, + 0.004353307653218508, + 0.005779467523097992, + 0.010860146023333073, + 0.004423136357218027, + 0.014871610328555107, + 0.010759549215435982, + 0.0033955026883631945, + 0.0034604233223944902, + 0.004492281936109066, + 0.004432477056980133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.0019378946162760258 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black, smooth, and muscular horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient Great Wall with weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse adorned with a pristine bronze harness and a dark red carved saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by the golden afterglow of a setting sun with a warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up tracking shot that moves in sync with the horse's rapid pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009774561040103436 + ], + "opposite_logprob_tokens": [ + -4.634774684906006 + ], + "logprob_reward": -0.009774561040103436, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009774561040103436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1841031014919281, + 0.06028663367033005, + 0.03339667618274689, + 0.06737811118364334, + 0.059333737939596176, + 0.11083365231752396, + 0.04644197225570679, + 0.1402810513973236, + 0.2584741711616516, + 0.03947091102600098 + ], + "credit_map": [ + 0.01653917320072651, + 0.029756715521216393, + 0.016049956902861595, + 0.02436567097902298, + 0.028858333826065063, + 0.020986931398510933, + 0.024273265153169632, + 0.02327306754887104, + 0.005693037994205952, + 0.011133214458823204, + 0.005180694162845612, + 0.006120635662227869, + 0.0065234024077653885, + 0.007072400767356157, + 0.010018723085522652, + 0.008544528856873512, + 0.0027064832393079996, + 0.003404945367947221, + 0.003578110132366419, + 0.00397996511310339, + 0.002613045508041978, + 0.006403112318366766, + 0.0054758470505476, + 0.005235165823251009, + 0.007308549713343382, + 0.01077057234942913, + 0.004789533093571663, + 0.006646295543760061, + 0.0046539013274014, + 0.015836549922823906, + 0.006584756541997194, + 0.010787956416606903, + 0.005772666539996862, + 0.008271699771285057, + 0.00854437705129385, + 0.006963523104786873, + 0.008110995404422283, + 0.006680294405668974, + 0.006904628127813339, + 0.008085552603006363, + 0.011516649276018143, + 0.011280392296612263, + 0.012025827541947365, + 0.010660292580723763, + 0.011280936188995838, + 0.021652350202202797, + 0.013171115890145302, + 0.019246090203523636, + 0.005231942515820265, + 0.004649467766284943, + 0.004037599544972181, + 0.004881402477622032, + 0.007774447090923786, + 0.006852220743894577, + 0.008242961019277573, + 0.004771927837282419, + 0.02067844569683075, + 0.016729895025491714, + 0.007871675305068493, + 0.014150184579193592, + 0.041116807609796524, + 0.015415197238326073, + 0.01619500108063221, + 0.008123844861984253, + 0.03319477662444115, + 0.03227579593658447, + 0.03520992025732994, + 0.03491703048348427, + 0.03449545055627823, + 0.02940960042178631, + 0.02572430670261383, + 0.03324728459119797, + 0.0065222978591918945, + 0.006102916784584522, + 0.007534841075539589, + 0.005217281635850668, + 0.003499601734802127, + 0.00290346029214561, + 0.003424472641199827, + 0.004266040865331888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.009774561040103436 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking low-angle, close-up tracking shot follows a majestic, black-smooth, muscular horse, galloping with unstoppable force across the ancient Great Wall. The camera moves in sync with the horse's rapid pace, never losing focus on its swift posture and tumbling hooves, its long mane and tail dancing wildly in the wind. The horse is adorned with a pristine bronze harness and a dark red carved saddle. Each hoofbeat treads powerfully against the weathered, rough-textured gray bricks, a steady and rhythmic movement that stirs up a fine dust cloud, as if resonating with the echoes of history. The backdrop is the golden afterglow of the setting sun, tinting the battlements and beacon towers winding into the distance with a warm hue and framing the horse's silhouette with a dazzling golden edge. The entire scene exudes a rich, classical Chinese martial arts style, with powerful and majestic tones, imbued with a sense of loneliness, tragedy, and unstoppable power, and an epic atmosphere.", + "seed": 32216, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_5/index_9255_seed_32216_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05468648606911301, + "vlm_logprob": -0.05468648606911301, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.05468648606911301, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black, smooth, and muscular horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient Great Wall with weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse adorned with a pristine bronze harness and a dark red carved saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by the golden afterglow of a setting sun with a warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up tracking shot that moves in sync with the horse's rapid pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04767346754670143 + ], + "opposite_logprob_tokens": [ + -3.172673463821411 + ], + "logprob_reward": -0.04767346754670143, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04767346754670143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07451529055833817, + 0.050616271793842316, + 0.07459985464811325, + 0.09402668476104736, + 0.1992364078760147, + 0.228644460439682, + 0.1410519927740097, + 0.05772547423839569, + 0.02598913200199604, + 0.0535944364964962 + ], + "credit_map": [ + 0.010400217957794666, + 0.010512488894164562, + 0.008310950361192226, + 0.011799401603639126, + 0.010068262927234173, + 0.008787676692008972, + 0.007701062131673098, + 0.006935217883437872, + 0.008631736971437931, + 0.0057553816586732864, + 0.009073499590158463, + 0.011288770474493504, + 0.0034079551696777344, + 0.003831815905869007, + 0.004998707212507725, + 0.0036284024827182293, + 0.006863800343126059, + 0.010905515402555466, + 0.00930961687117815, + 0.012168879620730877, + 0.010394325479865074, + 0.008268973790109158, + 0.006347236689180136, + 0.010341503657400608, + 0.010967550799250603, + 0.013251696713268757, + 0.009526497684419155, + 0.008907516486942768, + 0.022636841982603073, + 0.012311282567679882, + 0.009439622983336449, + 0.006985670421272516, + 0.029320096597075462, + 0.03426315635442734, + 0.033747028559446335, + 0.0250039491802454, + 0.02031935378909111, + 0.026400655508041382, + 0.013199605979025364, + 0.016982542350888252, + 0.03040885366499424, + 0.020482907071709633, + 0.04129433259367943, + 0.019040023908019066, + 0.04564410820603371, + 0.02507045306265354, + 0.021778658032417297, + 0.024925120174884796, + 0.01435789279639721, + 0.013853516429662704, + 0.02257579006254673, + 0.010746353305876255, + 0.027703149244189262, + 0.012442504987120628, + 0.009629612788558006, + 0.029743166640400887, + 0.006901943124830723, + 0.009993690066039562, + 0.005366071593016386, + 0.007552511990070343, + 0.00633598305284977, + 0.004972376395016909, + 0.005647833924740553, + 0.010955063626170158, + 0.0025617468636482954, + 0.004712941590696573, + 0.001942143077030778, + 0.0055552017875015736, + 0.002679963130503893, + 0.0023139487020671368, + 0.00278767803683877, + 0.0034355083480477333, + 0.0055718752555549145, + 0.008454115130007267, + 0.005504510831087828, + 0.00855831429362297, + 0.004916419740766287, + 0.006097126752138138, + 0.007041035685688257, + 0.007451036013662815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.04767346754670143 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black, smooth, and muscular horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient Great Wall with weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse adorned with a pristine bronze harness and a dark red carved saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by the golden afterglow of a setting sun with a warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up tracking shot that moves in sync with the horse's rapid pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004549354314804 + ], + "opposite_logprob_tokens": [ + -2.6950454711914062 + ], + "logprob_reward": -0.07004549354314804, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004549354314804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15125802159309387, + 0.11688348650932312, + 0.08690369874238968, + 0.08793601393699646, + 0.08992180228233337, + 0.10170800983905792, + 0.11030962318181992, + 0.025907184928655624, + 0.05084345489740372, + 0.1783287674188614 + ], + "credit_map": [ + 0.01972692459821701, + 0.016158193349838257, + 0.01850925199687481, + 0.027478385716676712, + 0.016110271215438843, + 0.01729396916925907, + 0.014226267114281654, + 0.021754756569862366, + 0.01414003036916256, + 0.014276333153247833, + 0.016656218096613884, + 0.031568847596645355, + 0.012423131614923477, + 0.009386557154357433, + 0.009898032061755657, + 0.008534333668649197, + 0.009008102118968964, + 0.015161683782935143, + 0.010670396499335766, + 0.014875887893140316, + 0.011521748267114162, + 0.006719959434121847, + 0.006613550707697868, + 0.012332363054156303, + 0.008018961176276207, + 0.011248580180108547, + 0.010982046835124493, + 0.01223292201757431, + 0.011112025007605553, + 0.01677902415394783, + 0.009009633213281631, + 0.008552816696465015, + 0.007705207914113998, + 0.011418596841394901, + 0.009228608570992947, + 0.012336955405771732, + 0.013595041818916798, + 0.011766576208174229, + 0.011253244243562222, + 0.012617573142051697, + 0.013183227740228176, + 0.009052314795553684, + 0.02184065245091915, + 0.009190557524561882, + 0.015259739011526108, + 0.010981306433677673, + 0.009347372688353062, + 0.012852835468947887, + 0.009621982462704182, + 0.009990150108933449, + 0.022632524371147156, + 0.010327803902328014, + 0.012123423628509045, + 0.008294477127492428, + 0.008657852187752724, + 0.028661413118243217, + 0.0019412323599681258, + 0.003954393323510885, + 0.0034881355240941048, + 0.005135222338140011, + 0.003317334456369281, + 0.002274878090247512, + 0.0027773098554462194, + 0.003018679330125451, + 0.00635079899802804, + 0.01170383207499981, + 0.005012137815356255, + 0.00718064745888114, + 0.005035093519836664, + 0.005201115272939205, + 0.005015942268073559, + 0.005343891214579344, + 0.019137490540742874, + 0.016952836886048317, + 0.017063893377780914, + 0.02517513558268547, + 0.013861395418643951, + 0.038419030606746674, + 0.02131832204759121, + 0.02640065923333168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.07004549354314804 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black, smooth, and muscular horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient Great Wall with weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse adorned with a pristine bronze harness and a dark red carved saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by the golden afterglow of a setting sun with a warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up tracking shot that moves in sync with the horse's rapid pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905255049467087 + ], + "opposite_logprob_tokens": [ + -2.464052438735962 + ], + "logprob_reward": -0.08905255049467087, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905255049467087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10529740154743195, + 0.05500838905572891, + 0.03608616814017296, + 0.09836680442094803, + 0.13388215005397797, + 0.2836378514766693, + 0.05279236659407616, + 0.09618698805570602, + 0.07519133388996124, + 0.06355062127113342 + ], + "credit_map": [ + 0.013039708137512207, + 0.014355737715959549, + 0.011167091317474842, + 0.011692346073687077, + 0.011098806746304035, + 0.018936993554234505, + 0.008725612424314022, + 0.01628109999001026, + 0.00456309923902154, + 0.00632037129253149, + 0.010775823146104813, + 0.013576652854681015, + 0.004860608372837305, + 0.005188929848372936, + 0.005591133143752813, + 0.004131765570491552, + 0.003910481929779053, + 0.004010593984276056, + 0.004670232068747282, + 0.0056834169663488865, + 0.006179930176585913, + 0.0030066585168242455, + 0.004173613153398037, + 0.004451241809874773, + 0.009011338464915752, + 0.010067535564303398, + 0.010663559660315514, + 0.016742181032896042, + 0.024241914972662926, + 0.010078782215714455, + 0.010580785572528839, + 0.006980708800256252, + 0.01868080534040928, + 0.019508177414536476, + 0.025522267445921898, + 0.014389933086931705, + 0.01544115412980318, + 0.01763645000755787, + 0.011342031881213188, + 0.011361346580088139, + 0.037814050912857056, + 0.023294560611248016, + 0.046326085925102234, + 0.0274362750351429, + 0.059650905430316925, + 0.022185510024428368, + 0.03798605129122734, + 0.02894439548254013, + 0.006373459938913584, + 0.00713322265073657, + 0.013630612753331661, + 0.005732221994549036, + 0.005108810495585203, + 0.004399881232529879, + 0.0040466696955263615, + 0.0063674901612102985, + 0.007337580900639296, + 0.017363736405968666, + 0.010560505092144012, + 0.011220895685255527, + 0.0059436410665512085, + 0.008130104281008244, + 0.012835409492254257, + 0.022795114666223526, + 0.013077856972813606, + 0.009785663336515427, + 0.0061408220790326595, + 0.019176024943590164, + 0.0061418465338647366, + 0.007312871981412172, + 0.007875537499785423, + 0.005680708214640617, + 0.007716713473200798, + 0.010876616463065147, + 0.0049700248055160046, + 0.010553760454058647, + 0.005792588926851749, + 0.007199964951723814, + 0.0068488041870296, + 0.009592147544026375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.08905255049467087 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black, smooth, and muscular horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient Great Wall with weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse adorned with a pristine bronze harness and a dark red carved saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by the golden afterglow of a setting sun with a warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up tracking shot that moves in sync with the horse's rapid pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636608529835939 + ], + "opposite_logprob_tokens": [ + -5.379636764526367 + ], + "logprob_reward": -0.004636608529835939, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636608529835939, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21849466860294342, + 0.15279124677181244, + 0.05084242299199104, + 0.12595944106578827, + 0.12471339106559753, + 0.13290387392044067, + 0.03329497575759888, + 0.0527694933116436, + 0.04061222821474075, + 0.067618228495121 + ], + "credit_map": [ + 0.024800531566143036, + 0.023548657074570656, + 0.038877926766872406, + 0.04310170188546181, + 0.01646650955080986, + 0.024649206548929214, + 0.019466446712613106, + 0.027583690360188484, + 0.025169644504785538, + 0.01999683305621147, + 0.018495643511414528, + 0.0318862721323967, + 0.013985127210617065, + 0.018535945564508438, + 0.010401325300335884, + 0.014320466667413712, + 0.004962721839547157, + 0.005531801376491785, + 0.0063133458606898785, + 0.011047476902604103, + 0.007723808288574219, + 0.004881076980382204, + 0.004401087295264006, + 0.005981109105050564, + 0.01709422655403614, + 0.00947768148034811, + 0.0193658247590065, + 0.027633462101221085, + 0.012105824425816536, + 0.012754065915942192, + 0.01624416746199131, + 0.011284194886684418, + 0.017575759440660477, + 0.009950539097189903, + 0.023389948531985283, + 0.013666338287293911, + 0.01748534105718136, + 0.017498690634965897, + 0.010483827441930771, + 0.014662952162325382, + 0.01144946925342083, + 0.013493332080543041, + 0.03228364884853363, + 0.010405974462628365, + 0.01607043854892254, + 0.013521884568035603, + 0.021630890667438507, + 0.014048246666789055, + 0.0028657319489866495, + 0.004424882587045431, + 0.007516985293477774, + 0.0024990271776914597, + 0.006446214858442545, + 0.002271045232191682, + 0.002428329549729824, + 0.004842760041356087, + 0.0039833784103393555, + 0.012177139520645142, + 0.005047486629337072, + 0.006803648080676794, + 0.006444141734391451, + 0.004001174122095108, + 0.005279778502881527, + 0.009032750502228737, + 0.009461983107030392, + 0.004957398399710655, + 0.005444019567221403, + 0.006177221424877644, + 0.00280980346724391, + 0.003740894142538309, + 0.0032634600065648556, + 0.004757450427860022, + 0.008123122155666351, + 0.009628094732761383, + 0.008834034204483032, + 0.008444524370133877, + 0.005918079987168312, + 0.011843209154903889, + 0.005643668118864298, + 0.009183499030768871 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.004636608529835939 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black, smooth, and muscular horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient Great Wall with weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse adorned with a pristine bronze harness and a dark red carved saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by the golden afterglow of a setting sun with a warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up tracking shot that moves in sync with the horse's rapid pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0620243102312088 + ], + "opposite_logprob_tokens": [ + -2.8120243549346924 + ], + "logprob_reward": -0.0620243102312088, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620243102312088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06736744195222855, + 0.059379883110523224, + 0.12144859880208969, + 0.1561640352010727, + 0.06706102192401886, + 0.1328275501728058, + 0.03877091407775879, + 0.26135995984077454, + 0.04314073175191879, + 0.05247984826564789 + ], + "credit_map": [ + 0.007454532664269209, + 0.007926602847874165, + 0.008398685604333878, + 0.01127967145293951, + 0.008353276178240776, + 0.00950563233345747, + 0.006318304687738419, + 0.008130738511681557, + 0.005998391192406416, + 0.007666667457669973, + 0.00959078874439001, + 0.012810914777219296, + 0.006409838795661926, + 0.0068458858877420425, + 0.005652151070535183, + 0.004405244719237089, + 0.014162522740662098, + 0.01694783940911293, + 0.014738068915903568, + 0.019634811207652092, + 0.017188776284456253, + 0.009881718084216118, + 0.009679857641458511, + 0.01921500824391842, + 0.01536979153752327, + 0.017935939133167267, + 0.016590775921940804, + 0.017230521887540817, + 0.033625125885009766, + 0.02457546442747116, + 0.015802571550011635, + 0.015033837407827377, + 0.007991617545485497, + 0.011177572421729565, + 0.009206555783748627, + 0.00880281999707222, + 0.009083380922675133, + 0.008037489838898182, + 0.005416262894868851, + 0.007345318328589201, + 0.01713331788778305, + 0.012802067212760448, + 0.020033158361911774, + 0.01336595043540001, + 0.019161848351359367, + 0.015765124931931496, + 0.01964305341243744, + 0.01492302492260933, + 0.004090262111276388, + 0.0035374825820326805, + 0.006314177066087723, + 0.0038577336817979813, + 0.0077162571251392365, + 0.0031691580079495907, + 0.003957981243729591, + 0.006127864122390747, + 0.027201475575566292, + 0.047776538878679276, + 0.025023043155670166, + 0.03314590826630592, + 0.023314377292990685, + 0.026197070255875587, + 0.021644318476319313, + 0.057057227939367294, + 0.004973831120878458, + 0.008818347007036209, + 0.004936044104397297, + 0.005076541565358639, + 0.0045302025973796844, + 0.004398769699037075, + 0.004986879415810108, + 0.00542011484503746, + 0.004347203299403191, + 0.0074228569865226746, + 0.005189806222915649, + 0.011533607728779316, + 0.004288884345442057, + 0.004432201851159334, + 0.009013672359287739, + 0.006251620128750801 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 5 + ], + "selected_reward": -0.0620243102312088 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking low-angle, close-up tracking shot follows a majestic, black-smooth, muscular horse, galloping with unstoppable force across the ancient Great Wall. The camera moves in sync with the horse's rapid pace, never losing focus on its swift posture and tumbling hooves, its long mane and tail dancing wildly in the wind. The horse is adorned with a pristine bronze harness and a dark red carved saddle. Each hoofbeat treads powerfully against the weathered, rough-textured gray bricks, a steady and rhythmic movement that stirs up a fine dust cloud, as if resonating with the echoes of history. The backdrop is the golden afterglow of the setting sun, tinting the battlements and beacon towers winding into the distance with a warm hue and framing the horse's silhouette with a dazzling golden edge. The entire scene exudes a rich, classical Chinese martial arts style, with powerful and majestic tones, imbued with a sense of loneliness, tragedy, and unstoppable power, and an epic atmosphere.", + "seed": 32217, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_5/index_9255_seed_32217_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02853717766702175, + "vlm_logprob": -0.02853717766702175, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.02853717766702175, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black, smooth, and muscular horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient Great Wall with weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse adorned with a pristine bronze harness and a dark red carved saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by the golden afterglow of a setting sun with a warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up tracking shot that moves in sync with the horse's rapid pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018878893926739693 + ], + "opposite_logprob_tokens": [ + -4.143878936767578 + ], + "logprob_reward": -0.018878893926739693, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.018878893926739693, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17599482834339142, + 0.16728593409061432, + 0.04819900169968605, + 0.07922207564115524, + 0.09170383960008621, + 0.08152057230472565, + 0.23057006299495697, + 0.02627847157418728, + 0.06022900342941284, + 0.038996268063783646 + ], + "credit_map": [ + 0.024966726079583168, + 0.02807106077671051, + 0.023889390751719475, + 0.014389574527740479, + 0.02616095542907715, + 0.015642395243048668, + 0.018526362255215645, + 0.02434837445616722, + 0.012900875881314278, + 0.016262955963611603, + 0.008194453082978725, + 0.006329097785055637, + 0.050178006291389465, + 0.009499987587332726, + 0.01116105355322361, + 0.05275950953364372, + 0.006873588543385267, + 0.007448227144777775, + 0.003922855015844107, + 0.002265965100377798, + 0.00845640990883112, + 0.004965677857398987, + 0.0039388141594827175, + 0.010327461175620556, + 0.007619161158800125, + 0.01125654112547636, + 0.008023031055927277, + 0.008237341418862343, + 0.011859011836349964, + 0.006319568958133459, + 0.007931418716907501, + 0.017975999042391777, + 0.010324477218091488, + 0.018127620220184326, + 0.011691401712596416, + 0.007016288116574287, + 0.013285509310662746, + 0.01241306122392416, + 0.005291505716741085, + 0.013553975149989128, + 0.005950496532022953, + 0.017236830666661263, + 0.0070070670917630196, + 0.005513486452400684, + 0.020577561110258102, + 0.01104690320789814, + 0.006828673183917999, + 0.007359554525464773, + 0.015179101377725601, + 0.029866039752960205, + 0.009187258780002594, + 0.010877802036702633, + 0.06616935133934021, + 0.021050164476037025, + 0.026783321052789688, + 0.051457032561302185, + 0.0035688753705471754, + 0.002916239434853196, + 0.0040681674145162106, + 0.0037432813551276922, + 0.0024033328518271446, + 0.0018916316330432892, + 0.002885820809751749, + 0.004801120609045029, + 0.008038844913244247, + 0.008813714608550072, + 0.009393916465342045, + 0.006570618599653244, + 0.0047737183049321175, + 0.004062809981405735, + 0.00801903661340475, + 0.01055634580552578, + 0.004853043705224991, + 0.004948396235704422, + 0.006681835278868675, + 0.004338754806667566, + 0.0025394170079380274, + 0.0024059407878667116, + 0.008324901573359966, + 0.004903977736830711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.018878893926739693 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black, smooth, and muscular horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient Great Wall with weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse adorned with a pristine bronze harness and a dark red carved saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by the golden afterglow of a setting sun with a warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up tracking shot that moves in sync with the horse's rapid pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014204937033355236 + ], + "opposite_logprob_tokens": [ + -4.264204978942871 + ], + "logprob_reward": -0.014204937033355236, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014204937033355236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0961165800690651, + 0.12632879614830017, + 0.037759341299533844, + 0.1767815798521042, + 0.0766444280743599, + 0.1589900702238083, + 0.11315158754587173, + 0.029954075813293457, + 0.09629473090171814, + 0.08797885477542877 + ], + "credit_map": [ + 0.010683717206120491, + 0.0206871647387743, + 0.007341311778873205, + 0.009216828271746635, + 0.013754853047430515, + 0.009101497940719128, + 0.010667783208191395, + 0.014663427136838436, + 0.011490206234157085, + 0.014292135834693909, + 0.0154163409024477, + 0.015065084211528301, + 0.01645616441965103, + 0.012237103655934334, + 0.01963846944272518, + 0.021733291447162628, + 0.004396498203277588, + 0.006178243085741997, + 0.002499970607459545, + 0.00317208981141448, + 0.006714634597301483, + 0.0032919871155172586, + 0.006158860400319099, + 0.005347057245671749, + 0.013993297703564167, + 0.02210877276957035, + 0.016495661810040474, + 0.019057240337133408, + 0.028378726914525032, + 0.013177135959267616, + 0.014526481740176678, + 0.04904424771666527, + 0.008423272520303726, + 0.013973992317914963, + 0.00959298387169838, + 0.0062395669519901276, + 0.011542001739144325, + 0.008948447182774544, + 0.006060429848730564, + 0.011863737367093563, + 0.011614171788096428, + 0.05095737800002098, + 0.008619190193712711, + 0.008958003483712673, + 0.04127751663327217, + 0.012189177796244621, + 0.012169771827757359, + 0.013204865157604218, + 0.00904921256005764, + 0.013979541137814522, + 0.00594446761533618, + 0.007055595982819796, + 0.029559440910816193, + 0.011308436281979084, + 0.012388892471790314, + 0.02386600710451603, + 0.003968746401369572, + 0.005269181914627552, + 0.0037309820763766766, + 0.003907281905412674, + 0.0028502459172159433, + 0.001865274622105062, + 0.0036921624559909105, + 0.004670200403779745, + 0.012466276995837688, + 0.010021878406405449, + 0.01922985166311264, + 0.014065941795706749, + 0.008745439350605011, + 0.006435828749090433, + 0.01236170157790184, + 0.012967810966074467, + 0.011394721455872059, + 0.010814367793500423, + 0.013629324734210968, + 0.011274266988039017, + 0.0052820732817053795, + 0.004277103580534458, + 0.01984163001179695, + 0.01146536972373724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 1 + ], + "selected_reward": -0.014204937033355236 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black, smooth, and muscular horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient Great Wall with weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse adorned with a pristine bronze harness and a dark red carved saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by the golden afterglow of a setting sun with a warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up tracking shot that moves in sync with the horse's rapid pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003182172775269 + ], + "opposite_logprob_tokens": [ + -2.6950318813323975 + ], + "logprob_reward": -0.07003182172775269, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003182172775269, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1236715093255043, + 0.1508435159921646, + 0.2326328158378601, + 0.03642762452363968, + 0.033067286014556885, + 0.14283600449562073, + 0.05574943125247955, + 0.038279805332422256, + 0.0967784896492958, + 0.08971355855464935 + ], + "credit_map": [ + 0.013985203579068184, + 0.03439673036336899, + 0.012999434024095535, + 0.009073170833289623, + 0.017904803156852722, + 0.008661334402859211, + 0.012985875830054283, + 0.013664948754012585, + 0.01651356741786003, + 0.015797093510627747, + 0.02105676755309105, + 0.017764175310730934, + 0.019273146986961365, + 0.017634151503443718, + 0.02424023300409317, + 0.018564360216259956, + 0.03512587398290634, + 0.053394924849271774, + 0.013705162331461906, + 0.009863504208624363, + 0.039763398468494415, + 0.015879778191447258, + 0.016865070909261703, + 0.048035092651844025, + 0.005185400135815144, + 0.006340662948787212, + 0.004460879601538181, + 0.002584664849564433, + 0.004980976693332195, + 0.0030634882859885693, + 0.0038987339939922094, + 0.005912814289331436, + 0.003058010945096612, + 0.0034333043731749058, + 0.003124792827293277, + 0.0031888624653220177, + 0.005603540688753128, + 0.004018484614789486, + 0.003169534495100379, + 0.007470752112567425, + 0.009112230502068996, + 0.030116641893982887, + 0.00919139664620161, + 0.012541218660771847, + 0.0327182300388813, + 0.020496025681495667, + 0.013876620680093765, + 0.014783625490963459, + 0.005163552705198526, + 0.007681138813495636, + 0.003403716254979372, + 0.002989480970427394, + 0.01492592878639698, + 0.0055892132222652435, + 0.006686350796371698, + 0.009310043416917324, + 0.005541025195270777, + 0.0038135445211082697, + 0.006615218240767717, + 0.005116533953696489, + 0.004096095450222492, + 0.0026064699050039053, + 0.004175978247076273, + 0.006314934231340885, + 0.012936946004629135, + 0.014526312239468098, + 0.013314802199602127, + 0.009842927567660809, + 0.008276870474219322, + 0.008504175581037998, + 0.012029653415083885, + 0.01734679564833641, + 0.010725243017077446, + 0.009402133524417877, + 0.015633484348654747, + 0.011075634509325027, + 0.006359736435115337, + 0.007059283554553986, + 0.013677827082574368, + 0.015780212357640266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 5 + ], + "selected_reward": -0.07003182172775269 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black, smooth, and muscular horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient Great Wall with weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse adorned with a pristine bronze harness and a dark red carved saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by the golden afterglow of a setting sun with a warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up tracking shot that moves in sync with the horse's rapid pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029794966802001 + ], + "opposite_logprob_tokens": [ + -3.529794931411743 + ], + "logprob_reward": -0.029794966802001, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029794966802001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.038944680243730545, + 0.3064461052417755, + 0.035351719707250595, + 0.02601437456905842, + 0.022164631634950638, + 0.36207881569862366, + 0.03426989167928696, + 0.01880914345383644, + 0.030573770403862, + 0.12534688413143158 + ], + "credit_map": [ + 0.005194879602640867, + 0.005694122519344091, + 0.005360465496778488, + 0.0037575331516563892, + 0.0049130674451589584, + 0.0034939872566610575, + 0.005120600573718548, + 0.0054100253619253635, + 0.030835356563329697, + 0.029660828411579132, + 0.012809483334422112, + 0.0093252407386899, + 0.10555990785360336, + 0.015851402655243874, + 0.016644958406686783, + 0.08575890213251114, + 0.004665777087211609, + 0.0050855777226388454, + 0.0024609381798654795, + 0.003458739258348942, + 0.00560997286811471, + 0.0029789756517857313, + 0.006868608295917511, + 0.004223129712045193, + 0.00294787366874516, + 0.0033463158179074526, + 0.0022080526687204838, + 0.002561780624091625, + 0.0037272137124091387, + 0.0031758605036884546, + 0.002553625963628292, + 0.005493652541190386, + 0.0019318960839882493, + 0.00330788129940629, + 0.001957143424078822, + 0.0021668453700840473, + 0.0034489461686462164, + 0.004131178837269545, + 0.0019019241444766521, + 0.003318814793601632, + 0.021807989105582237, + 0.0943337008357048, + 0.018242087215185165, + 0.020238416269421577, + 0.090315081179142, + 0.05243835970759392, + 0.031078577041625977, + 0.033624645322561264, + 0.002141006290912628, + 0.004507044795900583, + 0.0023221420124173164, + 0.002797758672386408, + 0.004545826930552721, + 0.003227249253541231, + 0.007025253493338823, + 0.007703608833253384, + 0.002623135456815362, + 0.0022370785009115934, + 0.0032098942901939154, + 0.00233464571647346, + 0.002014701021835208, + 0.0012629693374037743, + 0.001978195272386074, + 0.0031485245563089848, + 0.004195377696305513, + 0.005801567807793617, + 0.004170072730630636, + 0.0036250425036996603, + 0.0022376966662704945, + 0.0020447629503905773, + 0.0033954507671296597, + 0.005103799980133772, + 0.017750367522239685, + 0.01586497388780117, + 0.023297352716326714, + 0.015160915441811085, + 0.00986369140446186, + 0.008372691459953785, + 0.017991697415709496, + 0.017045192420482635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 9 + ], + "selected_reward": -0.029794966802001 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black, smooth, and muscular horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient Great Wall with weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse adorned with a pristine bronze harness and a dark red carved saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by the golden afterglow of a setting sun with a warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up tracking shot that moves in sync with the horse's rapid pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009775268845260143 + ], + "opposite_logprob_tokens": [ + -4.634775161743164 + ], + "logprob_reward": -0.009775268845260143, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009775268845260143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09352100640535355, + 0.07447482645511627, + 0.11875372380018234, + 0.1717482954263687, + 0.05573667585849762, + 0.1745222508907318, + 0.035252876579761505, + 0.122545026242733, + 0.08071108907461166, + 0.07273421436548233 + ], + "credit_map": [ + 0.00865846686065197, + 0.015084680169820786, + 0.013755733147263527, + 0.0073457746766507626, + 0.012875353917479515, + 0.007395798806101084, + 0.013071849942207336, + 0.015333352610468864, + 0.009323464706540108, + 0.010079196654260159, + 0.007114036474376917, + 0.004863928072154522, + 0.02116215042769909, + 0.005666031036525965, + 0.005826884415000677, + 0.010439134202897549, + 0.01652895286679268, + 0.015724340453743935, + 0.007313329726457596, + 0.006554374936968088, + 0.01717238314449787, + 0.007030979730188847, + 0.0102054663002491, + 0.03822389245033264, + 0.01512786652892828, + 0.023961300030350685, + 0.021341616287827492, + 0.018803514540195465, + 0.026968277990818024, + 0.012347700074315071, + 0.015159177593886852, + 0.03803884610533714, + 0.006426477339118719, + 0.00845179334282875, + 0.005488760303705931, + 0.0043991743586957455, + 0.00876616220921278, + 0.007272318471223116, + 0.005258099641650915, + 0.0096738925203681, + 0.014430677518248558, + 0.04831406846642494, + 0.011096621863543987, + 0.009589320980012417, + 0.04440302029252052, + 0.01811220683157444, + 0.013623795472085476, + 0.014952528290450573, + 0.0048371851444244385, + 0.00658036395907402, + 0.003996334504336119, + 0.003737442661076784, + 0.004985510371625423, + 0.0026189046911895275, + 0.004997494164854288, + 0.0034996424801647663, + 0.0158946942538023, + 0.015320084057748318, + 0.02290952205657959, + 0.016553279012441635, + 0.0107231130823493, + 0.009189045056700706, + 0.01323302648961544, + 0.018722258508205414, + 0.01023760624229908, + 0.014390475116670132, + 0.00902070663869381, + 0.009678622707724571, + 0.007450494449585676, + 0.0058556729927659035, + 0.01157673541456461, + 0.012500769458711147, + 0.010807023383677006, + 0.00882254634052515, + 0.011850152164697647, + 0.010515637695789337, + 0.006550345569849014, + 0.004899817984551191, + 0.009326249361038208, + 0.009962438605725765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 7 + ], + "selected_reward": -0.009775268845260143 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking low-angle, close-up tracking shot follows a majestic, black-smooth, muscular horse, galloping with unstoppable force across the ancient Great Wall. The camera moves in sync with the horse's rapid pace, never losing focus on its swift posture and tumbling hooves, its long mane and tail dancing wildly in the wind. The horse is adorned with a pristine bronze harness and a dark red carved saddle. Each hoofbeat treads powerfully against the weathered, rough-textured gray bricks, a steady and rhythmic movement that stirs up a fine dust cloud, as if resonating with the echoes of history. The backdrop is the golden afterglow of the setting sun, tinting the battlements and beacon towers winding into the distance with a warm hue and framing the horse's silhouette with a dazzling golden edge. The entire scene exudes a rich, classical Chinese martial arts style, with powerful and majestic tones, imbued with a sense of loneliness, tragedy, and unstoppable power, and an epic atmosphere.", + "seed": 32218, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_5/index_9255_seed_32218_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02393229566514492, + "vlm_logprob": -0.02393229566514492, + "vlm_yes_no_margin": 4.049999952316284, + "vlm_token_credit": -0.02393229566514492, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black, smooth, and muscular horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient Great Wall with weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse adorned with a pristine bronze harness and a dark red carved saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by the golden afterglow of a setting sun with a warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up tracking shot that moves in sync with the horse's rapid pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015723302960395813 + ], + "opposite_logprob_tokens": [ + -4.51572322845459 + ], + "logprob_reward": -0.015723302960395813, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.015723302960395813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.192175030708313, + 0.05026701092720032, + 0.09500160068273544, + 0.03641337901353836, + 0.04255308210849762, + 0.10975116491317749, + 0.20926597714424133, + 0.09027869254350662, + 0.1097947433590889, + 0.06449928134679794 + ], + "credit_map": [ + 0.02261570654809475, + 0.030073869973421097, + 0.027233310043811798, + 0.02105427160859108, + 0.016475437209010124, + 0.02505471557378769, + 0.027577243745326996, + 0.02209046110510826, + 0.00443734135478735, + 0.004475919064134359, + 0.003028975101187825, + 0.0036260574124753475, + 0.013237973675131798, + 0.004918640945106745, + 0.005950635299086571, + 0.010591469705104828, + 0.011621332727372646, + 0.01667889952659607, + 0.00996267981827259, + 0.009699461981654167, + 0.01434825174510479, + 0.010758675634860992, + 0.009707500226795673, + 0.012224795296788216, + 0.0031836428679525852, + 0.004995300900191069, + 0.004375319927930832, + 0.005343212280422449, + 0.005014415364712477, + 0.0036014365032315254, + 0.004469763953238726, + 0.005430287215858698, + 0.00641622394323349, + 0.007824103347957134, + 0.005426538176834583, + 0.004643273074179888, + 0.004438429605215788, + 0.003703956725075841, + 0.005198079627007246, + 0.004902481101453304, + 0.0147943627089262, + 0.015887310728430748, + 0.015095934271812439, + 0.008614971302449703, + 0.019946761429309845, + 0.008117897436022758, + 0.015189907513558865, + 0.012104016728699207, + 0.018386362120509148, + 0.041989535093307495, + 0.021451367065310478, + 0.021523620933294296, + 0.03574462607502937, + 0.016370072960853577, + 0.030828706920146942, + 0.022971680387854576, + 0.009699949994683266, + 0.014545617625117302, + 0.008264997974038124, + 0.011608735658228397, + 0.01327148824930191, + 0.008129126392304897, + 0.011064634658396244, + 0.01369413547217846, + 0.01864541321992874, + 0.008588743396103382, + 0.014797360636293888, + 0.011530834250152111, + 0.01371827907860279, + 0.021765973418951035, + 0.007727723568677902, + 0.01302042230963707, + 0.017717020586133003, + 0.0044625140726566315, + 0.003835348878055811, + 0.00330874091014266, + 0.005905785132199526, + 0.01984170451760292, + 0.003586790757253766, + 0.005841375794261694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.015723302960395813 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black, smooth, and muscular horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient Great Wall with weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse adorned with a pristine bronze harness and a dark red carved saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by the golden afterglow of a setting sun with a warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up tracking shot that moves in sync with the horse's rapid pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016081036999821663 + ], + "opposite_logprob_tokens": [ + -4.141080856323242 + ], + "logprob_reward": -0.016081036999821663, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016081036999821663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32798081636428833, + 0.04090181738138199, + 0.11009418219327927, + 0.020341571420431137, + 0.030235247686505318, + 0.09227076917886734, + 0.07498127222061157, + 0.04861251264810562, + 0.05553814768791199, + 0.1990436464548111 + ], + "credit_map": [ + 0.038131002336740494, + 0.04990054666996002, + 0.029853742569684982, + 0.023976830765604973, + 0.05531568452715874, + 0.05326893553137779, + 0.030194688588380814, + 0.047339413315057755, + 0.003933570347726345, + 0.006650413852185011, + 0.0034533643629401922, + 0.0038448278792202473, + 0.004488848615437746, + 0.0061150649562478065, + 0.004578120075166225, + 0.00783760566264391, + 0.011481991037726402, + 0.017564525827765465, + 0.011735863983631134, + 0.01082780584692955, + 0.015468782745301723, + 0.015109008178114891, + 0.00980597548186779, + 0.018100233748555183, + 0.002098838333040476, + 0.0027745221741497517, + 0.002396727679297328, + 0.002846953459084034, + 0.002225259318947792, + 0.002310671377927065, + 0.0018518575234338641, + 0.0038367414381355047, + 0.004258684813976288, + 0.00440513901412487, + 0.0038642422296106815, + 0.003297592978924513, + 0.0033115127589553595, + 0.003912266343832016, + 0.0040045492351055145, + 0.0031812619417905807, + 0.010092825628817081, + 0.009100722149014473, + 0.009884778410196304, + 0.011748623102903366, + 0.011729284189641476, + 0.013806341215968132, + 0.007372505497187376, + 0.018535688519477844, + 0.008143341168761253, + 0.018871035426855087, + 0.005294891074299812, + 0.011219099164009094, + 0.0089634470641613, + 0.00522683747112751, + 0.007944878190755844, + 0.009317743591964245, + 0.004561424721032381, + 0.006573780439794064, + 0.004163264762610197, + 0.0053827790543437, + 0.00675479369238019, + 0.005662562791258097, + 0.005551076028496027, + 0.009962831623852253, + 0.007445009425282478, + 0.00438606645911932, + 0.006112259346991777, + 0.007116612978279591, + 0.004511823412030935, + 0.012842412106692791, + 0.004209883511066437, + 0.008914082311093807, + 0.05915125086903572, + 0.012997155077755451, + 0.009036130271852016, + 0.0074792164377868176, + 0.009425200521945953, + 0.06885562837123871, + 0.009220799431204796, + 0.022878272458910942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.016081036999821663 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black, smooth, and muscular horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient Great Wall with weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse adorned with a pristine bronze harness and a dark red carved saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by the golden afterglow of a setting sun with a warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up tracking shot that moves in sync with the horse's rapid pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0620555654168129 + ], + "opposite_logprob_tokens": [ + -2.8120555877685547 + ], + "logprob_reward": -0.0620555654168129, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620555654168129, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12253355979919434, + 0.06702182441949844, + 0.07933306694030762, + 0.08890751749277115, + 0.10682127624750137, + 0.22117114067077637, + 0.19698190689086914, + 0.03845280781388283, + 0.05164913833141327, + 0.027127724140882492 + ], + "credit_map": [ + 0.017539767548441887, + 0.01733107678592205, + 0.011140302754938602, + 0.010961636900901794, + 0.013948214240372181, + 0.016930170357227325, + 0.014415700919926167, + 0.0202667023986578, + 0.008006474003195763, + 0.008170598186552525, + 0.006046213675290346, + 0.005295957904309034, + 0.010147830471396446, + 0.009019793942570686, + 0.0106084905564785, + 0.009726470336318016, + 0.01012782659381628, + 0.015839969739317894, + 0.008884736336767673, + 0.010516325943171978, + 0.013872669078409672, + 0.007107310928404331, + 0.006888578180223703, + 0.00609565107151866, + 0.010480709373950958, + 0.021146710962057114, + 0.009760065004229546, + 0.008231861516833305, + 0.01068401150405407, + 0.005872913170605898, + 0.015276942402124405, + 0.00745431100949645, + 0.017012452706694603, + 0.026982851326465607, + 0.01615634560585022, + 0.010330971330404282, + 0.009307211264967918, + 0.008116237819194794, + 0.008383347652852535, + 0.010531867854297161, + 0.026400312781333923, + 0.045936550945043564, + 0.023127494379878044, + 0.01856343448162079, + 0.025920087471604347, + 0.025666793808341026, + 0.021281255409121513, + 0.03427523374557495, + 0.021061426028609276, + 0.05094587430357933, + 0.0177864208817482, + 0.022827530279755592, + 0.021330660209059715, + 0.012663849629461765, + 0.02389420196413994, + 0.02647196687757969, + 0.004002953879535198, + 0.006652057636529207, + 0.002676241798326373, + 0.004034220241010189, + 0.005517345387488604, + 0.0038586461450904608, + 0.006027017720043659, + 0.005684324540197849, + 0.005009152460843325, + 0.005709111224859953, + 0.005261790473014116, + 0.005635365843772888, + 0.005712348502129316, + 0.012896088883280754, + 0.003687353804707527, + 0.0077379303984344006, + 0.006175791844725609, + 0.0014492298942059278, + 0.0016088498523458838, + 0.0019031462725251913, + 0.002238593762740493, + 0.0067191035486757755, + 0.001984315225854516, + 0.005048695486038923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.0620555654168129 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black, smooth, and muscular horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient Great Wall with weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse adorned with a pristine bronze harness and a dark red carved saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by the golden afterglow of a setting sun with a warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up tracking shot that moves in sync with the horse's rapid pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007624335587024689 + ], + "opposite_logprob_tokens": [ + -4.88262414932251 + ], + "logprob_reward": -0.007624335587024689, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007624335587024689, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10486318171024323, + 0.16990335285663605, + 0.05475698411464691, + 0.02109554596245289, + 0.03448466211557388, + 0.08171133697032928, + 0.060510169714689255, + 0.047474827617406845, + 0.3637824058532715, + 0.06141757592558861 + ], + "credit_map": [ + 0.016169147565960884, + 0.01623441092669964, + 0.011053967289626598, + 0.010337344370782375, + 0.01294027455151081, + 0.012794164009392262, + 0.010767542757093906, + 0.014566333964467049, + 0.013312102295458317, + 0.02151891216635704, + 0.009929624386131763, + 0.01715853065252304, + 0.018918262794613838, + 0.019351312890648842, + 0.0363951213657856, + 0.03331948071718216, + 0.007611207198351622, + 0.011286386288702488, + 0.006589478813111782, + 0.007126624695956707, + 0.008864876814186573, + 0.0036298634950071573, + 0.005016743671149015, + 0.004631803371012211, + 0.0017770968843251467, + 0.003236703807488084, + 0.0022735039237886667, + 0.002578240353614092, + 0.0028891772963106632, + 0.001759802340529859, + 0.0029568837489932775, + 0.0036241384223103523, + 0.004056617617607117, + 0.00461914436891675, + 0.004339376464486122, + 0.007460037712007761, + 0.0029792592395097017, + 0.004222120624035597, + 0.002669442445039749, + 0.004138665739446878, + 0.008537237532436848, + 0.007757335901260376, + 0.010523471981287003, + 0.010888984426856041, + 0.010470625944435596, + 0.008321531116962433, + 0.01008077897131443, + 0.015131376683712006, + 0.004801163915544748, + 0.014677152037620544, + 0.00518618943169713, + 0.010850407183170319, + 0.006976580712944269, + 0.004112184979021549, + 0.005708859767764807, + 0.008197633549571037, + 0.004645261913537979, + 0.005402445327490568, + 0.005140212830156088, + 0.0048956056125462055, + 0.008349070325493813, + 0.0046388255432248116, + 0.00611504353582859, + 0.008288361132144928, + 0.03887558728456497, + 0.020461848005652428, + 0.033903591334819794, + 0.043892621994018555, + 0.0396849662065506, + 0.1212654709815979, + 0.025963013991713524, + 0.03973530977964401, + 0.012638397514820099, + 0.005518762860447168, + 0.0036444170400500298, + 0.0027329260483384132, + 0.004642477259039879, + 0.019639860838651657, + 0.0037190609145909548, + 0.008881675079464912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.007624335587024689 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black, smooth, and muscular horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient Great Wall with weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse adorned with a pristine bronze harness and a dark red carved saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by the golden afterglow of a setting sun with a warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up tracking shot that moves in sync with the horse's rapid pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01817723736166954 + ], + "opposite_logprob_tokens": [ + -4.018177032470703 + ], + "logprob_reward": -0.01817723736166954, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01817723736166954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12077287584543228, + 0.055277954787015915, + 0.09938875585794449, + 0.031031744554638863, + 0.09942621737718582, + 0.11388219892978668, + 0.11900190263986588, + 0.13682708144187927, + 0.07654763013124466, + 0.14784368872642517 + ], + "credit_map": [ + 0.017791803926229477, + 0.024843547493219376, + 0.012528115883469582, + 0.012855369597673416, + 0.010516776703298092, + 0.01499260775744915, + 0.010778035968542099, + 0.01646661013364792, + 0.005235878750681877, + 0.004622112028300762, + 0.005512966308742762, + 0.006340009160339832, + 0.009351560845971107, + 0.00548422709107399, + 0.004905145615339279, + 0.01382604893296957, + 0.01355685479938984, + 0.01531520951539278, + 0.011350726708769798, + 0.014074504375457764, + 0.015216339379549026, + 0.011827879585325718, + 0.00972835160791874, + 0.008318870328366756, + 0.004616633988916874, + 0.004962034523487091, + 0.004003386944532394, + 0.003483480541035533, + 0.0034609735012054443, + 0.002369960770010948, + 0.0037565859965980053, + 0.004378683399409056, + 0.013786490075290203, + 0.015102027915418148, + 0.013706603087484837, + 0.0094259362667799, + 0.011137169785797596, + 0.01407814770936966, + 0.008199972100555897, + 0.013989861123263836, + 0.012365823611617088, + 0.01348885241895914, + 0.00938776507973671, + 0.012746014632284641, + 0.008701495826244354, + 0.022969523444771767, + 0.009290570393204689, + 0.024932142347097397, + 0.010327327065169811, + 0.022838594391942024, + 0.012855431996285915, + 0.019724072888493538, + 0.013705721125006676, + 0.009490298107266426, + 0.013653687201440334, + 0.016406754031777382, + 0.013250479474663734, + 0.022151239216327667, + 0.013159972615540028, + 0.014145571738481522, + 0.02333458513021469, + 0.012738239020109177, + 0.014955364167690277, + 0.023091629147529602, + 0.008854984305799007, + 0.007492997217923403, + 0.009913587011396885, + 0.011224164627492428, + 0.009526103734970093, + 0.012383857741951942, + 0.0065215155482292175, + 0.010630415752530098, + 0.025048697367310524, + 0.01849634386599064, + 0.012297459878027439, + 0.012465695850551128, + 0.01812858134508133, + 0.027084143832325935, + 0.011847523972392082, + 0.0224752314388752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.01817723736166954 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_6/index_tensor([8250, 8250])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_6/index_tensor([8250, 8250])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0b2c287d5766bc44d7b75a0cef4b94650b85535a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_6/index_tensor([8250, 8250])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An initially static, high-angle, extreme close-up focuses on the dark, glossy surface of steaming black coffee in a plain, white ceramic cup. Distorted within the liquid's surface is the distorted, fish-eye reflection of a weary office worker in his 30s, his slightly disheveled brown hair and gaunt features illuminated by the cool, sterile glow of the overhead fluorescent lights. As he gazes into the cup, his hand, revealing the faint lines of a long day's work, enters the frame, purposefully wrapped around the warm ceramic. Then, the camera begins a slow, steady upward tilt, perfectly synchronized with the cup's movement as he raises it to his face. As the cup is raised, the reflection disappears, and the man himself comes into focus in an eye-level shot; he wears a crisp but slightly wrinkled white dress shirt, the top button undone, his navy blue tie loosened. He closes his eyes, and in a moment of weary anticipation, his head tilted back slightly as he takes a long, deep sip of the drink, his Adam's apple and the slight bob that marks his swallow. The background remains a soft-focus blur of cubicle walls and the glow of computer monitors, creating a palpable atmosphere of quiet solitude and exhaustion, interrupted only by this small, restorative ritual.", + "seed": 24056, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_6/index_8250_seed_24056_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05939491270110011, + "vlm_logprob": -0.05939491270110011, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.05939491270110011, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up of dark, glossy, steaming black coffee in a white ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a fish-eye reflection of a weary office worker in his 30s with disheveled brown hair and gaunt features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, steady upward tilt synchronized with the cup being raised to the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man shown wearing a white dress shirt with the top button undone and a loosened navy blue tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt his head back slightly to take a long sip, showing his Adam's apple and swallow, against a soft-focus background of cubicle walls and computer monitors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0396902970969677 + ], + "opposite_logprob_tokens": [ + -3.5396902561187744 + ], + "logprob_reward": -0.0396902970969677, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0396902970969677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06637763231992722, + 0.05341910570859909, + 0.06764597445726395, + 0.21954359114170074, + 0.19399072229862213, + 0.2168532758951187, + 0.0780215635895729, + 0.03846915066242218, + 0.036809343844652176, + 0.0288696251809597 + ], + "credit_map": [ + 0.008673127740621567, + 0.007236478850245476, + 0.006257687695324421, + 0.008494294248521328, + 0.010054724290966988, + 0.01165675837546587, + 0.006008364725857973, + 0.007996199652552605, + 0.007244391366839409, + 0.005798890721052885, + 0.0057115028612315655, + 0.006112693343311548, + 0.006255010608583689, + 0.010553915984928608, + 0.006124136969447136, + 0.005618564318865538, + 0.0078878840431571, + 0.008122263476252556, + 0.004089968279004097, + 0.007765499874949455, + 0.01011585257947445, + 0.010943948291242123, + 0.013094710186123848, + 0.005625852383673191, + 0.016949545592069626, + 0.04001101851463318, + 0.026904867962002754, + 0.018081575632095337, + 0.04537750408053398, + 0.021328415721654892, + 0.020704234018921852, + 0.030186425894498825, + 0.03375108912587166, + 0.015849897637963295, + 0.02425248920917511, + 0.025333333760499954, + 0.01668059267103672, + 0.027461692690849304, + 0.018873628228902817, + 0.03178801015019417, + 0.03903546929359436, + 0.022474290803074837, + 0.023519322276115417, + 0.01231011189520359, + 0.038700610399246216, + 0.03433797135949135, + 0.020880617201328278, + 0.025594884529709816, + 0.005476867780089378, + 0.007581695448607206, + 0.00524534285068512, + 0.007199438754469156, + 0.015453414991497993, + 0.012239385396242142, + 0.013183300383388996, + 0.01164211705327034, + 0.004958398640155792, + 0.003511911490932107, + 0.006423642858862877, + 0.00358355650678277, + 0.005325974430888891, + 0.006260193884372711, + 0.004419088829308748, + 0.00398638192564249, + 0.004946226719766855, + 0.0056854975409805775, + 0.0038391745183616877, + 0.003974018152803183, + 0.004307012539356947, + 0.006634980905801058, + 0.003985084593296051, + 0.003437352366745472, + 0.0038630757480859756, + 0.005050439853221178, + 0.0031820086296647787, + 0.0027660324703902006, + 0.0035581961274147034, + 0.0037269059102982283, + 0.00385609595105052, + 0.002866871189326048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.0396902970969677 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up of dark, glossy, steaming black coffee in a white ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a fish-eye reflection of a weary office worker in his 30s with disheveled brown hair and gaunt features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, steady upward tilt synchronized with the cup being raised to the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man shown wearing a white dress shirt with the top button undone and a loosened navy blue tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt his head back slightly to take a long sip, showing his Adam's apple and swallow, against a soft-focus background of cubicle walls and computer monitors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22577784955501556 + ], + "opposite_logprob_tokens": [ + -1.6007778644561768 + ], + "logprob_reward": -0.22577784955501556, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22577784955501556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0700240433216095, + 0.017187662422657013, + 0.03927486762404442, + 0.16151028871536255, + 0.2108522206544876, + 0.19602292776107788, + 0.10546162724494934, + 0.02730967290699482, + 0.12983383238315582, + 0.042522821575403214 + ], + "credit_map": [ + 0.02064189501106739, + 0.006019454915076494, + 0.005167005583643913, + 0.004716811235994101, + 0.014575532637536526, + 0.007409833837300539, + 0.004969884641468525, + 0.006523624528199434, + 0.0026657411362975836, + 0.002024594461545348, + 0.0019330700160935521, + 0.002087425207719207, + 0.0022525042295455933, + 0.003176846308633685, + 0.0012162172934040427, + 0.0018312644679099321, + 0.007026331033557653, + 0.00470103370025754, + 0.0030363763216882944, + 0.003636027220636606, + 0.006331987213343382, + 0.004287875257432461, + 0.0063298549503088, + 0.003925382159650326, + 0.01234489120543003, + 0.03275243565440178, + 0.017650213092565536, + 0.0114051578566432, + 0.03436525538563728, + 0.0178937129676342, + 0.013644207268953323, + 0.02145441435277462, + 0.040777530521154404, + 0.015906453132629395, + 0.03071579709649086, + 0.03521620109677315, + 0.01708851009607315, + 0.02418270707130432, + 0.021543018519878387, + 0.025421999394893646, + 0.03404298797249794, + 0.018948011100292206, + 0.019265811890363693, + 0.011895212344825268, + 0.03529345989227295, + 0.032734595239162445, + 0.0197894424200058, + 0.024053392931818962, + 0.010014197789132595, + 0.007768189534544945, + 0.011930204927921295, + 0.0075424970127642155, + 0.018504776060581207, + 0.01898059993982315, + 0.013862471096217632, + 0.016858696937561035, + 0.0033689343836158514, + 0.003313240595161915, + 0.003056399757042527, + 0.0031338653061538935, + 0.003661841619759798, + 0.004170315805822611, + 0.003280771430581808, + 0.0033243054058402777, + 0.021546058356761932, + 0.014025145210325718, + 0.011310114525258541, + 0.008434589020907879, + 0.013576663099229336, + 0.03483154997229576, + 0.012272650375962257, + 0.01383706834167242, + 0.005025398917496204, + 0.004053876269608736, + 0.0019258726388216019, + 0.0015767752192914486, + 0.014790983870625496, + 0.004301882814615965, + 0.007196096703410149, + 0.0036519342102110386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.22577784955501556 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up of dark, glossy, steaming black coffee in a white ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a fish-eye reflection of a weary office worker in his 30s with disheveled brown hair and gaunt features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, steady upward tilt synchronized with the cup being raised to the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man shown wearing a white dress shirt with the top button undone and a loosened navy blue tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt his head back slightly to take a long sip, showing his Adam's apple and swallow, against a soft-focus background of cubicle walls and computer monitors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046631875447928905 + ], + "opposite_logprob_tokens": [ + -5.379662990570068 + ], + "logprob_reward": -0.0046631875447928905, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046631875447928905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03567162901163101, + 0.028454286977648735, + 0.028591861948370934, + 0.07883293926715851, + 0.12627603113651276, + 0.13088704645633698, + 0.24990399181842804, + 0.07773977518081665, + 0.13805098831653595, + 0.10559146106243134 + ], + "credit_map": [ + 0.004700343124568462, + 0.0036211255937814713, + 0.002778880065307021, + 0.00677473796531558, + 0.004802373703569174, + 0.005655380897223949, + 0.0033612544648349285, + 0.003977533895522356, + 0.0034374347887933254, + 0.0039046236779540777, + 0.003178667277097702, + 0.0027007514145225286, + 0.0038211338687688112, + 0.005815901327878237, + 0.001956951105967164, + 0.0036388251464813948, + 0.004933516960591078, + 0.0032251623924821615, + 0.001953745260834694, + 0.004182880744338036, + 0.004387864377349615, + 0.0024648639373481274, + 0.005180706270039082, + 0.002263122471049428, + 0.0075819166377186775, + 0.012310069985687733, + 0.011698281392455101, + 0.005463275592774153, + 0.01351383700966835, + 0.009433111175894737, + 0.008779884316027164, + 0.010052558034658432, + 0.012945999391376972, + 0.01810038648545742, + 0.011989576742053032, + 0.014001497067511082, + 0.018905872479081154, + 0.01708921231329441, + 0.01916417106986046, + 0.014079325832426548, + 0.016660595312714577, + 0.013881787657737732, + 0.014967159368097782, + 0.008098728023469448, + 0.028215931728482246, + 0.021400965750217438, + 0.012742972932755947, + 0.01491890661418438, + 0.013856783509254456, + 0.014578986912965775, + 0.0214859489351511, + 0.021797213703393936, + 0.04629022628068924, + 0.0420539453625679, + 0.04395086318254471, + 0.045890018343925476, + 0.01006876677274704, + 0.007809914648532867, + 0.016817783936858177, + 0.008984842337667942, + 0.011484917253255844, + 0.007423782721161842, + 0.007948528043925762, + 0.007201243191957474, + 0.019690535962581635, + 0.017283527180552483, + 0.01393450703471899, + 0.014727246947586536, + 0.01752576418220997, + 0.02639893814921379, + 0.012482216581702232, + 0.016008250415325165, + 0.011110744439065456, + 0.007728120777755976, + 0.00377048272639513, + 0.004140972625464201, + 0.03621695190668106, + 0.030140630900859833, + 0.0079043535515666, + 0.0045792036689817905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 5 + ], + "selected_reward": -0.0046631875447928905 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up of dark, glossy, steaming black coffee in a white ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a fish-eye reflection of a weary office worker in his 30s with disheveled brown hair and gaunt features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, steady upward tilt synchronized with the cup being raised to the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man shown wearing a white dress shirt with the top button undone and a loosened navy blue tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt his head back slightly to take a long sip, showing his Adam's apple and swallow, against a soft-focus background of cubicle walls and computer monitors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008647613227367401 + ], + "opposite_logprob_tokens": [ + -4.758647441864014 + ], + "logprob_reward": -0.008647613227367401, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008647613227367401, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023300910368561745, + 0.00925273634493351, + 0.007829432375729084, + 0.03302912786602974, + 0.05845193564891815, + 0.033917736262083054, + 0.7639688849449158, + 0.020173177123069763, + 0.029068950563669205, + 0.021007120609283447 + ], + "credit_map": [ + 0.0035470626316964626, + 0.001844421261921525, + 0.001614679698832333, + 0.002279688138514757, + 0.004175154957920313, + 0.0026763661298900843, + 0.0013053196016699076, + 0.005858216434717178, + 0.001571877859532833, + 0.001136188511736691, + 0.0010554337641224265, + 0.0008111884235404432, + 0.0015765160787850618, + 0.0012477769050747156, + 0.0007080660434439778, + 0.0011456884676590562, + 0.001150125521235168, + 0.0009999361354857683, + 0.0008248179801739752, + 0.0007207946619018912, + 0.0011648887302726507, + 0.0010221623815596104, + 0.0007359235314652324, + 0.0012107840739190578, + 0.002323326887562871, + 0.0067461333237588406, + 0.003579500364139676, + 0.0025009666569530964, + 0.006457448936998844, + 0.004779299255460501, + 0.0025978772900998592, + 0.0040445770137012005, + 0.009249493479728699, + 0.004396989941596985, + 0.0067246295511722565, + 0.006492956075817347, + 0.005382633302360773, + 0.008147476240992546, + 0.005901358090341091, + 0.012156393378973007, + 0.0042806812562048435, + 0.004197356291115284, + 0.00449006911367178, + 0.003493329044431448, + 0.0072386483661830425, + 0.0037824998144060373, + 0.0029141930863261223, + 0.0035209613852202892, + 0.0567682720720768, + 0.04117187112569809, + 0.07882221788167953, + 0.07483629882335663, + 0.14241497218608856, + 0.11236266046762466, + 0.12659502029418945, + 0.13099756836891174, + 0.002600547857582569, + 0.0021288027055561543, + 0.002229181583970785, + 0.00205051782540977, + 0.002458799397572875, + 0.002732028253376484, + 0.002621009945869446, + 0.003352290252223611, + 0.004914282821118832, + 0.0039020823314785957, + 0.0028440163005143404, + 0.0028361640870571136, + 0.0030892048962414265, + 0.00482680881395936, + 0.0029384889639914036, + 0.0037179035134613514, + 0.00202829297631979, + 0.0020008219871670008, + 0.0008095110533758998, + 0.0006614653393626213, + 0.007966834120452404, + 0.004265024792402983, + 0.0022883471101522446, + 0.0009868235792964697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -0.008647613227367401 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up of dark, glossy, steaming black coffee in a white ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a fish-eye reflection of a weary office worker in his 30s with disheveled brown hair and gaunt features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, steady upward tilt synchronized with the cup being raised to the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man shown wearing a white dress shirt with the top button undone and a loosened navy blue tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt his head back slightly to take a long sip, showing his Adam's apple and swallow, against a soft-focus background of cubicle walls and computer monitors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018195616081357002 + ], + "opposite_logprob_tokens": [ + -4.018195629119873 + ], + "logprob_reward": -0.018195616081357002, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018195616081357002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04615481570363045, + 0.014864850789308548, + 0.025206344202160835, + 0.03252650424838066, + 0.031499169766902924, + 0.0784933939576149, + 0.0997428223490715, + 0.047230325639247894, + 0.06106195226311684, + 0.5632198452949524 + ], + "credit_map": [ + 0.010766434483230114, + 0.0031702162232249975, + 0.003603733377531171, + 0.005216785706579685, + 0.008453378453850746, + 0.004767885431647301, + 0.0022688903845846653, + 0.00790749303996563, + 0.0020616839174181223, + 0.0018696716288104653, + 0.0023335565347224474, + 0.001438060193322599, + 0.0016348097706213593, + 0.0025629347655922174, + 0.0009544420754536986, + 0.0020096919033676386, + 0.0028560173232108355, + 0.003119864035397768, + 0.002150997519493103, + 0.0024069459177553654, + 0.003049092832952738, + 0.005015621893107891, + 0.004091157577931881, + 0.002516647567972541, + 0.0028810438234359026, + 0.0031069356482475996, + 0.005938122048974037, + 0.0023337702732533216, + 0.003267838852480054, + 0.006233625579625368, + 0.004199380055069923, + 0.004565788898617029, + 0.005514226388186216, + 0.0034219075459986925, + 0.003476638114079833, + 0.0030688790138810873, + 0.004274048376828432, + 0.0038392620626837015, + 0.003273761598393321, + 0.004630445037037134, + 0.009296681731939316, + 0.00916964653879404, + 0.008720095269382, + 0.007038125768303871, + 0.012270838022232056, + 0.011895395815372467, + 0.009706255979835987, + 0.010396353900432587, + 0.008141216821968555, + 0.008143794722855091, + 0.007171350531280041, + 0.010504871606826782, + 0.01690146140754223, + 0.01093178242444992, + 0.019301163032650948, + 0.01864718087017536, + 0.004576933104544878, + 0.005564799997955561, + 0.006508646998554468, + 0.00402678269892931, + 0.005953386891633272, + 0.008169499225914478, + 0.006416996009647846, + 0.006013282109051943, + 0.006584614980965853, + 0.011180575005710125, + 0.0063811978325247765, + 0.007289092987775803, + 0.009246587753295898, + 0.008254225365817547, + 0.0075383358635008335, + 0.004587320610880852, + 0.05281169340014458, + 0.043427322059869766, + 0.013771756552159786, + 0.011331716552376747, + 0.22189833223819733, + 0.13223561644554138, + 0.06505587697029114, + 0.0226875152438879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -0.018195616081357002 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An initially static, high-angle, extreme close-up focuses on the dark, glossy surface of steaming black coffee in a plain, white ceramic cup. Distorted within the liquid's surface is the distorted, fish-eye reflection of a weary office worker in his 30s, his slightly disheveled brown hair and gaunt features illuminated by the cool, sterile glow of the overhead fluorescent lights. As he gazes into the cup, his hand, revealing the faint lines of a long day's work, enters the frame, purposefully wrapped around the warm ceramic. Then, the camera begins a slow, steady upward tilt, perfectly synchronized with the cup's movement as he raises it to his face. As the cup is raised, the reflection disappears, and the man himself comes into focus in an eye-level shot; he wears a crisp but slightly wrinkled white dress shirt, the top button undone, his navy blue tie loosened. He closes his eyes, and in a moment of weary anticipation, his head tilted back slightly as he takes a long, deep sip of the drink, his Adam's apple and the slight bob that marks his swallow. The background remains a soft-focus blur of cubicle walls and the glow of computer monitors, creating a palpable atmosphere of quiet solitude and exhaustion, interrupted only by this small, restorative ritual.", + "seed": 24057, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_6/index_8250_seed_24057_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.036890757363289596, + "vlm_logprob": -0.036890757363289596, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.036890757363289596, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up of dark, glossy, steaming black coffee in a white ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a fish-eye reflection of a weary office worker in his 30s with disheveled brown hair and gaunt features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, steady upward tilt synchronized with the cup being raised to the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man shown wearing a white dress shirt with the top button undone and a loosened navy blue tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt his head back slightly to take a long sip, showing his Adam's apple and swallow, against a soft-focus background of cubicle walls and computer monitors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02480958215892315 + ], + "opposite_logprob_tokens": [ + -4.14980936050415 + ], + "logprob_reward": -0.02480958215892315, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.02480958215892315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08042967319488525, + 0.03806612268090248, + 0.016728686168789864, + 0.057830799371004105, + 0.19584614038467407, + 0.1872231811285019, + 0.03092299774289131, + 0.06994421035051346, + 0.17910541594028473, + 0.14390277862548828 + ], + "credit_map": [ + 0.005680544301867485, + 0.013755417428910732, + 0.004915042780339718, + 0.006974175572395325, + 0.019071323797106743, + 0.008100519888103008, + 0.013705408200621605, + 0.008227241225540638, + 0.0029303370974957943, + 0.0061043621972203255, + 0.002359325299039483, + 0.0023567627649754286, + 0.00935870036482811, + 0.004941198974847794, + 0.006667709443718195, + 0.0033477244433015585, + 0.0014980861451476812, + 0.0024992465041577816, + 0.0013902797363698483, + 0.0012894583633169532, + 0.003989960066974163, + 0.0018589997198432684, + 0.0027615244034677744, + 0.0014411297161132097, + 0.0069540939293801785, + 0.007761158514767885, + 0.007565419655293226, + 0.006318969186395407, + 0.009573460556566715, + 0.008610645309090614, + 0.006343564484268427, + 0.004703490063548088, + 0.02177933044731617, + 0.006944361142814159, + 0.00983775220811367, + 0.012961994856595993, + 0.03756774589419365, + 0.01877305842936039, + 0.0737551748752594, + 0.014226725324988365, + 0.014916143380105495, + 0.015077323652803898, + 0.00513352919369936, + 0.005347732454538345, + 0.010101204738020897, + 0.06777714192867279, + 0.013928858563303947, + 0.054941244423389435, + 0.0030967448838055134, + 0.003539056284353137, + 0.0025753905065357685, + 0.004950248636305332, + 0.00391822773963213, + 0.005734763573855162, + 0.002826202195137739, + 0.00428236648440361, + 0.007360604591667652, + 0.008250158280134201, + 0.006251403596252203, + 0.01103510707616806, + 0.007948375307023525, + 0.011622881516814232, + 0.008166282437741756, + 0.00930939894169569, + 0.020631669089198112, + 0.021126097068190575, + 0.0180142093449831, + 0.022994976490736008, + 0.03154822438955307, + 0.022027945145964622, + 0.02200370840728283, + 0.02075859159231186, + 0.01637820526957512, + 0.012321015819907188, + 0.017012737691402435, + 0.008466673083603382, + 0.014952797442674637, + 0.03779046609997749, + 0.01302274689078331, + 0.023958129808306694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 8 + ], + "selected_reward": -0.02480958215892315 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up of dark, glossy, steaming black coffee in a white ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a fish-eye reflection of a weary office worker in his 30s with disheveled brown hair and gaunt features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, steady upward tilt synchronized with the cup being raised to the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man shown wearing a white dress shirt with the top button undone and a loosened navy blue tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt his head back slightly to take a long sip, showing his Adam's apple and swallow, against a soft-focus background of cubicle walls and computer monitors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1130855605006218 + ], + "opposite_logprob_tokens": [ + -2.2380855083465576 + ], + "logprob_reward": -0.1130855605006218, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1130855605006218, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0576888769865036, + 0.06628140807151794, + 0.025965692475438118, + 0.09039181470870972, + 0.13158582150936127, + 0.13035817444324493, + 0.10019445419311523, + 0.09360066801309586, + 0.22352750599384308, + 0.08040556311607361 + ], + "credit_map": [ + 0.0036016693338751793, + 0.010069510899484158, + 0.0033563142642378807, + 0.006426217500120401, + 0.013434179127216339, + 0.006277540698647499, + 0.009941799566149712, + 0.0045816488564014435, + 0.005619592498987913, + 0.011087432503700256, + 0.005638576112687588, + 0.005760252010077238, + 0.009502341970801353, + 0.009758752770721912, + 0.009496919810771942, + 0.009417545050382614, + 0.00250068842433393, + 0.004894941579550505, + 0.0029245074838399887, + 0.0022748885676264763, + 0.0031971121206879616, + 0.00299825775437057, + 0.00358328758738935, + 0.003592010820284486, + 0.012333410792052746, + 0.012432227842509747, + 0.014509250409901142, + 0.008967909030616283, + 0.012325385585427284, + 0.011241725645959377, + 0.00862796138972044, + 0.009953944943845272, + 0.01622871123254299, + 0.006520072463899851, + 0.007910220883786678, + 0.012768137268722057, + 0.018392406404018402, + 0.015483368188142776, + 0.04127804934978485, + 0.013004857115447521, + 0.014636298641562462, + 0.012652583420276642, + 0.008267178200185299, + 0.006148355547338724, + 0.017305869609117508, + 0.028119588270783424, + 0.013350828550755978, + 0.029877470806241035, + 0.009261677972972393, + 0.009864838793873787, + 0.010445124469697475, + 0.011537359096109867, + 0.011636604554951191, + 0.021141396835446358, + 0.012112938798964024, + 0.014194521121680737, + 0.005756230093538761, + 0.011691675521433353, + 0.00843367725610733, + 0.01384852547198534, + 0.011708677746355534, + 0.019711961969733238, + 0.009753204882144928, + 0.01269671879708767, + 0.037097420543432236, + 0.024454016238451004, + 0.022694239392876625, + 0.033324144780635834, + 0.03015483357012272, + 0.033560317009687424, + 0.022899087518453598, + 0.01934344507753849, + 0.01086159236729145, + 0.010305873118340969, + 0.008564767427742481, + 0.006070234347134829, + 0.0074583180248737335, + 0.016110649332404137, + 0.01056639850139618, + 0.010467730462551117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 5 + ], + "selected_reward": -0.1130855605006218 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up of dark, glossy, steaming black coffee in a white ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a fish-eye reflection of a weary office worker in his 30s with disheveled brown hair and gaunt features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, steady upward tilt synchronized with the cup being raised to the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man shown wearing a white dress shirt with the top button undone and a loosened navy blue tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt his head back slightly to take a long sip, showing his Adam's apple and swallow, against a soft-focus background of cubicle walls and computer monitors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076541476882994175 + ], + "opposite_logprob_tokens": [ + -4.882654190063477 + ], + "logprob_reward": -0.0076541476882994175, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076541476882994175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04509942606091499, + 0.07122307270765305, + 0.03082806058228016, + 0.07842361927032471, + 0.257795125246048, + 0.20788204669952393, + 0.0901104062795639, + 0.12148400396108627, + 0.06347716599702835, + 0.033677052706480026 + ], + "credit_map": [ + 0.003911670297384262, + 0.006644210312515497, + 0.004674658644944429, + 0.002629709430038929, + 0.011641748249530792, + 0.0036402896512299776, + 0.0074027120135724545, + 0.004554427694529295, + 0.004647703375667334, + 0.01034952886402607, + 0.004046867601573467, + 0.005232064984738827, + 0.023988524451851845, + 0.0048033008351922035, + 0.014267995022237301, + 0.0038870880380272865, + 0.0038371365517377853, + 0.004257754422724247, + 0.0034888614900410175, + 0.0020415829494595528, + 0.004268265329301357, + 0.0031498714815825224, + 0.004922200459986925, + 0.00486238906159997, + 0.009717458859086037, + 0.008189031854271889, + 0.011411241255700588, + 0.00819371361285448, + 0.015431837178766727, + 0.008719833567738533, + 0.008681642822921276, + 0.008078859187662601, + 0.040077149868011475, + 0.009988189674913883, + 0.017897121608257294, + 0.018173258751630783, + 0.03552943095564842, + 0.02514706924557686, + 0.09278186410665512, + 0.018201032653450966, + 0.02050032839179039, + 0.01828879304230213, + 0.01909640058875084, + 0.009727309457957745, + 0.023007018491625786, + 0.05203232169151306, + 0.02088553085923195, + 0.044344350695610046, + 0.006235636305063963, + 0.010596273466944695, + 0.012959134764969349, + 0.011191966012120247, + 0.011562245897948742, + 0.01734965108335018, + 0.008731411769986153, + 0.011484089307487011, + 0.008779844269156456, + 0.015134202316403389, + 0.013556718826293945, + 0.02484208345413208, + 0.013612516224384308, + 0.01574089378118515, + 0.014577596448361874, + 0.0152401402592659, + 0.0137775344774127, + 0.00629434734582901, + 0.0063169533386826515, + 0.004786679055541754, + 0.008189287036657333, + 0.012217342853546143, + 0.007005293387919664, + 0.004889728035777807, + 0.004727753344923258, + 0.0049516260623931885, + 0.0055072070099413395, + 0.004036041907966137, + 0.0041337222792208195, + 0.0031836009584367275, + 0.0037859047297388315, + 0.0033511961810290813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.0076541476882994175 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up of dark, glossy, steaming black coffee in a white ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a fish-eye reflection of a weary office worker in his 30s with disheveled brown hair and gaunt features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, steady upward tilt synchronized with the cup being raised to the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man shown wearing a white dress shirt with the top button undone and a loosened navy blue tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt his head back slightly to take a long sip, showing his Adam's apple and swallow, against a soft-focus background of cubicle walls and computer monitors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012549244798719883 + ], + "opposite_logprob_tokens": [ + -4.38754940032959 + ], + "logprob_reward": -0.012549244798719883, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012549244798719883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04318942129611969, + 0.021436531096696854, + 0.007825109176337719, + 0.02674640156328678, + 0.07496785372495651, + 0.45504534244537354, + 0.14659324288368225, + 0.08507256209850311, + 0.045361053198575974, + 0.09376244992017746 + ], + "credit_map": [ + 0.003851545276120305, + 0.005538045894354582, + 0.0049238502979278564, + 0.0027322075329720974, + 0.008873502723872662, + 0.004827393218874931, + 0.006364373490214348, + 0.006078503560274839, + 0.0015213926089927554, + 0.003350395243614912, + 0.002235779305920005, + 0.0014148858608677983, + 0.006293822545558214, + 0.0015961466124281287, + 0.003787045832723379, + 0.0012370619224384427, + 0.0009558163583278656, + 0.0010923343943431973, + 0.0008321859058924019, + 0.0005133440718054771, + 0.0011329920962452888, + 0.0009563795174472034, + 0.0011620238656178117, + 0.0011800334323197603, + 0.003508083987981081, + 0.0028595856856554747, + 0.004481659270823002, + 0.002655835822224617, + 0.004898370709270239, + 0.0027598231099545956, + 0.0027565546333789825, + 0.0028264906723052263, + 0.007437492720782757, + 0.0022465912625193596, + 0.002872475888580084, + 0.005035653710365295, + 0.01465860940515995, + 0.007332641165703535, + 0.030312033370137215, + 0.005072354804724455, + 0.03227870911359787, + 0.03306712210178375, + 0.008330268785357475, + 0.01140932459384203, + 0.023811431601643562, + 0.17516754567623138, + 0.03551294654607773, + 0.1354680210351944, + 0.013964877463877201, + 0.013767625205218792, + 0.015165552496910095, + 0.01792735606431961, + 0.018790697678923607, + 0.029324540868401527, + 0.01967143639922142, + 0.017981164157390594, + 0.005111185368150473, + 0.01012327615171671, + 0.009111925028264523, + 0.016394460573792458, + 0.010677727870643139, + 0.013927311636507511, + 0.008489181287586689, + 0.011237489990890026, + 0.005849816370755434, + 0.005456065759062767, + 0.004635681863874197, + 0.00435677682980895, + 0.006991859059780836, + 0.006962715182453394, + 0.006195141933858395, + 0.004912996664643288, + 0.011212041601538658, + 0.008499306626617908, + 0.011885251849889755, + 0.007306596729904413, + 0.01133279874920845, + 0.018883028998970985, + 0.011966866441071033, + 0.012676557525992393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 9 + ], + "selected_reward": -0.012549244798719883 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up of dark, glossy, steaming black coffee in a white ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a fish-eye reflection of a weary office worker in his 30s with disheveled brown hair and gaunt features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, steady upward tilt synchronized with the cup being raised to the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man shown wearing a white dress shirt with the top button undone and a loosened navy blue tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt his head back slightly to take a long sip, showing his Adam's apple and swallow, against a soft-focus background of cubicle walls and computer monitors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026355251669883728 + ], + "opposite_logprob_tokens": [ + -3.651355266571045 + ], + "logprob_reward": -0.026355251669883728, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026355251669883728, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0445900596678257, + 0.01836620271205902, + 0.013039844110608101, + 0.03508244454860687, + 0.048535559326410294, + 0.3275439143180847, + 0.039868272840976715, + 0.056811098009347916, + 0.1363062709569931, + 0.2798563539981842 + ], + "credit_map": [ + 0.0035191967617720366, + 0.006005150731652975, + 0.005284939426928759, + 0.002801458118483424, + 0.010124234482645988, + 0.004818640183657408, + 0.006760824471712112, + 0.005275614093989134, + 0.0023057761136442423, + 0.0028271765913814306, + 0.0027087966445833445, + 0.0016638878732919693, + 0.0030804721172899008, + 0.0020003155805170536, + 0.0025675022043287754, + 0.001212274539284408, + 0.001440751482732594, + 0.002045220695436001, + 0.001500164158642292, + 0.0008413541363552213, + 0.0018241878133267164, + 0.0013276080135256052, + 0.0020909965969622135, + 0.0019695612136274576, + 0.00433895131573081, + 0.0040140943601727486, + 0.005302701611071825, + 0.00367511878721416, + 0.005484801717102528, + 0.0040444000624120235, + 0.003734082682058215, + 0.004488295875489712, + 0.007413126993924379, + 0.002372890245169401, + 0.0031546778045594692, + 0.0037674757186323404, + 0.00722375325858593, + 0.0045892647467553616, + 0.016227111220359802, + 0.003787259105592966, + 0.023388279601931572, + 0.024560827761888504, + 0.006038956344127655, + 0.008671307004988194, + 0.01786123774945736, + 0.12583625316619873, + 0.02600124478340149, + 0.0951857939362526, + 0.0029396561440080404, + 0.004660474136471748, + 0.004500227048993111, + 0.006146072875708342, + 0.005597879644483328, + 0.006535626482218504, + 0.00429358659312129, + 0.005194749217480421, + 0.003730220254510641, + 0.00823215488344431, + 0.00665240827947855, + 0.012551474384963512, + 0.006521837320178747, + 0.005415215622633696, + 0.00576821668073535, + 0.007939569652080536, + 0.02255244180560112, + 0.013757064938545227, + 0.013706259429454803, + 0.015841234475374222, + 0.023800751194357872, + 0.020519807934761047, + 0.012559841386973858, + 0.01356886513531208, + 0.03392267972230911, + 0.02666253224015236, + 0.03314607962965965, + 0.01741446740925312, + 0.030348297208547592, + 0.06764470040798187, + 0.025177234783768654, + 0.045540355145931244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.026355251669883728 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An initially static, high-angle, extreme close-up focuses on the dark, glossy surface of steaming black coffee in a plain, white ceramic cup. Distorted within the liquid's surface is the distorted, fish-eye reflection of a weary office worker in his 30s, his slightly disheveled brown hair and gaunt features illuminated by the cool, sterile glow of the overhead fluorescent lights. As he gazes into the cup, his hand, revealing the faint lines of a long day's work, enters the frame, purposefully wrapped around the warm ceramic. Then, the camera begins a slow, steady upward tilt, perfectly synchronized with the cup's movement as he raises it to his face. As the cup is raised, the reflection disappears, and the man himself comes into focus in an eye-level shot; he wears a crisp but slightly wrinkled white dress shirt, the top button undone, his navy blue tie loosened. He closes his eyes, and in a moment of weary anticipation, his head tilted back slightly as he takes a long, deep sip of the drink, his Adam's apple and the slight bob that marks his swallow. The background remains a soft-focus blur of cubicle walls and the glow of computer monitors, creating a palpable atmosphere of quiet solitude and exhaustion, interrupted only by this small, restorative ritual.", + "seed": 24058, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_6/index_8250_seed_24058_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03571347836405039, + "vlm_logprob": -0.03571347836405039, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.03571347836405039, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up of dark, glossy, steaming black coffee in a white ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a fish-eye reflection of a weary office worker in his 30s with disheveled brown hair and gaunt features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, steady upward tilt synchronized with the cup being raised to the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man shown wearing a white dress shirt with the top button undone and a loosened navy blue tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt his head back slightly to take a long sip, showing his Adam's apple and swallow, against a soft-focus background of cubicle walls and computer monitors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018782848492264748 + ], + "opposite_logprob_tokens": [ + -4.518782615661621 + ], + "logprob_reward": -0.018782848492264748, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.018782848492264748, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5331511497497559, + 0.05528337508440018, + 0.0938548892736435, + 0.04875601828098297, + 0.0415484681725502, + 0.09286567568778992, + 0.022278305143117905, + 0.06073034927248955, + 0.02014334313571453, + 0.03138839080929756 + ], + "credit_map": [ + 0.06628811359405518, + 0.1335291713476181, + 0.039824724197387695, + 0.05532078444957733, + 0.049352068454027176, + 0.08190850168466568, + 0.023426255211234093, + 0.08350154757499695, + 0.005403895862400532, + 0.012757701799273491, + 0.005997162312269211, + 0.006372946314513683, + 0.005077783018350601, + 0.006683145649731159, + 0.006609318312257528, + 0.0063814218156039715, + 0.010393502190709114, + 0.01122475415468216, + 0.01643768884241581, + 0.00873477291315794, + 0.013355429284274578, + 0.014253225177526474, + 0.009718436747789383, + 0.009737079963088036, + 0.005118841305375099, + 0.006935495883226395, + 0.011082067154347897, + 0.005055729765444994, + 0.004199396353214979, + 0.00444515235722065, + 0.0024387382436543703, + 0.009480600245296955, + 0.00420351093634963, + 0.0070830583572387695, + 0.007712647784501314, + 0.006570012774318457, + 0.004751111380755901, + 0.005110892467200756, + 0.002565585309639573, + 0.0035516470670700073, + 0.008704476058483124, + 0.009738671593368053, + 0.005728091578930616, + 0.008595462888479233, + 0.00879698432981968, + 0.009887710213661194, + 0.019187508150935173, + 0.02222677506506443, + 0.0030539033468812704, + 0.003809359623119235, + 0.002494141226634383, + 0.0032501330133527517, + 0.002084267558529973, + 0.003400907153263688, + 0.00217708433046937, + 0.0020085086580365896, + 0.004672515206038952, + 0.006643138825893402, + 0.004260594490915537, + 0.00729184178635478, + 0.015349986962974072, + 0.008557171560823917, + 0.007986160926520824, + 0.005968939512968063, + 0.002926107030361891, + 0.002500212285667658, + 0.0019161500968039036, + 0.0025907615199685097, + 0.004473546985536814, + 0.0020048401784151793, + 0.001967264339327812, + 0.001764461980201304, + 0.0031698443926870823, + 0.0033495535608381033, + 0.0033907180186361074, + 0.004634188953787088, + 0.005727256182581186, + 0.004148972686380148, + 0.004438433330506086, + 0.002529422054067254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.018782848492264748 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up of dark, glossy, steaming black coffee in a white ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a fish-eye reflection of a weary office worker in his 30s with disheveled brown hair and gaunt features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, steady upward tilt synchronized with the cup being raised to the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man shown wearing a white dress shirt with the top button undone and a loosened navy blue tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt his head back slightly to take a long sip, showing his Adam's apple and swallow, against a soft-focus background of cubicle walls and computer monitors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12714432179927826 + ], + "opposite_logprob_tokens": [ + -2.1271443367004395 + ], + "logprob_reward": -0.12714432179927826, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12714432179927826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2042108178138733, + 0.09005776792764664, + 0.15981580317020416, + 0.14838972687721252, + 0.06922368705272675, + 0.08540987223386765, + 0.04153858870267868, + 0.08227936178445816, + 0.04565977677702904, + 0.073414646089077 + ], + "credit_map": [ + 0.02431402914226055, + 0.04658344388008118, + 0.0176884513348341, + 0.024056032299995422, + 0.016540173441171646, + 0.029974671080708504, + 0.01299477182328701, + 0.03205924108624458, + 0.010762251913547516, + 0.008823537267744541, + 0.012522503733634949, + 0.009196586906909943, + 0.00960773229598999, + 0.01029221061617136, + 0.013982747681438923, + 0.014870195649564266, + 0.018094083294272423, + 0.019020168110728264, + 0.025183670222759247, + 0.016272826120257378, + 0.02302427962422371, + 0.02413293533027172, + 0.016022659838199615, + 0.018065189942717552, + 0.017700370401144028, + 0.0350094735622406, + 0.022435862571001053, + 0.01677933894097805, + 0.0208220724016428, + 0.013212900608778, + 0.010850131511688232, + 0.011579572223126888, + 0.007332813460379839, + 0.013179507106542587, + 0.009627573192119598, + 0.00910252146422863, + 0.011952569708228111, + 0.006281863898038864, + 0.005498033482581377, + 0.006248808931559324, + 0.007780267391353846, + 0.006577847991138697, + 0.004989685490727425, + 0.007604208774864674, + 0.00839005783200264, + 0.011186152696609497, + 0.017301028594374657, + 0.021580621600151062, + 0.004804980009794235, + 0.005854369606822729, + 0.004163019359111786, + 0.0060111405327916145, + 0.003614610992372036, + 0.006750978529453278, + 0.00662451796233654, + 0.0037149733398109674, + 0.007010647561401129, + 0.009546051733195782, + 0.005815171170979738, + 0.007991797290742397, + 0.014948631636798382, + 0.008814697153866291, + 0.009955938905477524, + 0.01819642260670662, + 0.0031520610209554434, + 0.0035973016638308764, + 0.007506755646318197, + 0.0052365693263709545, + 0.008118820376694202, + 0.005149182863533497, + 0.0060889991000294685, + 0.006810087710618973, + 0.0053808121010661125, + 0.004875297658145428, + 0.0045946575701236725, + 0.008295011706650257, + 0.015609223395586014, + 0.012176894582808018, + 0.015019082464277744, + 0.0074636684730648994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.12714432179927826 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up of dark, glossy, steaming black coffee in a white ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a fish-eye reflection of a weary office worker in his 30s with disheveled brown hair and gaunt features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, steady upward tilt synchronized with the cup being raised to the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man shown wearing a white dress shirt with the top button undone and a loosened navy blue tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt his head back slightly to take a long sip, showing his Adam's apple and swallow, against a soft-focus background of cubicle walls and computer monitors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004659034311771393 + ], + "opposite_logprob_tokens": [ + -5.379659175872803 + ], + "logprob_reward": -0.004659034311771393, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004659034311771393, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12423107028007507, + 0.18296785652637482, + 0.11044551432132721, + 0.13813801109790802, + 0.10100538283586502, + 0.10875296592712402, + 0.038541100919246674, + 0.11663396656513214, + 0.04085003212094307, + 0.03843408450484276 + ], + "credit_map": [ + 0.015282545238733292, + 0.025413574650883675, + 0.009260066784918308, + 0.01126890815794468, + 0.014248416759073734, + 0.023210270330309868, + 0.005800286773592234, + 0.01974700577557087, + 0.014510991983115673, + 0.011386365629732609, + 0.02813052013516426, + 0.016216956079006195, + 0.013080710545182228, + 0.02596292644739151, + 0.03878407180309296, + 0.03489530831575394, + 0.01318495161831379, + 0.013186945579946041, + 0.017344191670417786, + 0.013019019737839699, + 0.0147663289681077, + 0.016087565571069717, + 0.011708281002938747, + 0.011148236691951752, + 0.01479974202811718, + 0.023753637447953224, + 0.019390426576137543, + 0.019737180322408676, + 0.01671391911804676, + 0.017031272873282433, + 0.011934585869312286, + 0.014777245931327343, + 0.009006649255752563, + 0.012413903139531612, + 0.015583733096718788, + 0.01141833234578371, + 0.012163279578089714, + 0.01374799944460392, + 0.011297010816633701, + 0.015374468639492989, + 0.015197140164673328, + 0.014237076975405216, + 0.008870171383023262, + 0.014456295408308506, + 0.013079528696835041, + 0.013821166008710861, + 0.011619980446994305, + 0.017471613362431526, + 0.004334730561822653, + 0.0060852388851344585, + 0.0045318896882236, + 0.0046936203725636005, + 0.003733274294063449, + 0.005997533909976482, + 0.0050809625536203384, + 0.004083850886672735, + 0.010566778481006622, + 0.011077516712248325, + 0.008502057753503323, + 0.011014861986041069, + 0.020611688494682312, + 0.015523617155849934, + 0.012749695219099522, + 0.026587743312120438, + 0.0035502463579177856, + 0.004055290948599577, + 0.004882958251982927, + 0.004719823598861694, + 0.008739271201193333, + 0.006155645474791527, + 0.003932449035346508, + 0.0048143439926207066, + 0.0035177937243133783, + 0.003722087014466524, + 0.00394797045737505, + 0.0063303313218057156, + 0.0056232125498354435, + 0.00634184293448925, + 0.0055298032239079475, + 0.0034210451412945986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.004659034311771393 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up of dark, glossy, steaming black coffee in a white ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a fish-eye reflection of a weary office worker in his 30s with disheveled brown hair and gaunt features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, steady upward tilt synchronized with the cup being raised to the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man shown wearing a white dress shirt with the top button undone and a loosened navy blue tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt his head back slightly to take a long sip, showing his Adam's apple and swallow, against a soft-focus background of cubicle walls and computer monitors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009789316914975643 + ], + "opposite_logprob_tokens": [ + -4.63478946685791 + ], + "logprob_reward": -0.009789316914975643, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009789316914975643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08665616065263748, + 0.018595807254314423, + 0.054431572556495667, + 0.2476554661989212, + 0.03769730404019356, + 0.4077679514884949, + 0.02369905821979046, + 0.07158862799406052, + 0.027720658108592033, + 0.02418735809624195 + ], + "credit_map": [ + 0.008632106706500053, + 0.015973780304193497, + 0.005524997133761644, + 0.008631162345409393, + 0.005426101386547089, + 0.01815337873995304, + 0.0038629546761512756, + 0.020451679825782776, + 0.002371096983551979, + 0.0029788424726575613, + 0.0020145142916589975, + 0.0017136929091066122, + 0.0020891567692160606, + 0.0024762072134763002, + 0.002090971916913986, + 0.0028613239992409945, + 0.005568382795900106, + 0.00624126847833395, + 0.010433988645672798, + 0.0077260201796889305, + 0.006963585037738085, + 0.006819932721555233, + 0.005665754899382591, + 0.005012637935578823, + 0.020449208095669746, + 0.014611012302339077, + 0.07260065525770187, + 0.022583860903978348, + 0.015286922454833984, + 0.022927403450012207, + 0.01172737404704094, + 0.0674690380692482, + 0.004858873784542084, + 0.005885958671569824, + 0.006860209628939629, + 0.004511659499257803, + 0.004804021213203669, + 0.00427610706537962, + 0.0035441440995782614, + 0.0029563317075371742, + 0.02735547348856926, + 0.030007611960172653, + 0.019539711996912956, + 0.0271395705640316, + 0.03881160169839859, + 0.048828188329935074, + 0.10942176729440689, + 0.1066640242934227, + 0.004147811327129602, + 0.0034732401836663485, + 0.0026512693148106337, + 0.002439626958221197, + 0.0023868835996836424, + 0.0033541808370500803, + 0.002431601984426379, + 0.0028144451789557934, + 0.009434310719370842, + 0.006480577401816845, + 0.0048979418352246284, + 0.007590792141854763, + 0.01044258289039135, + 0.006401669234037399, + 0.006896464619785547, + 0.019444288685917854, + 0.002658406039699912, + 0.0037108336109668016, + 0.0029457933269441128, + 0.004543148446828127, + 0.00428925221785903, + 0.002910131588578224, + 0.003002986777573824, + 0.003660106798633933, + 0.002472544554620981, + 0.0022485111840069294, + 0.002456917427480221, + 0.003510246751829982, + 0.003405465744435787, + 0.0038523832336068153, + 0.004059046506881714, + 0.0021822426933795214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.009789316914975643 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up of dark, glossy, steaming black coffee in a white ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a fish-eye reflection of a weary office worker in his 30s with disheveled brown hair and gaunt features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, steady upward tilt synchronized with the cup being raised to the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man shown wearing a white dress shirt with the top button undone and a loosened navy blue tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt his head back slightly to take a long sip, showing his Adam's apple and swallow, against a soft-focus background of cubicle walls and computer monitors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0181918703019619 + ], + "opposite_logprob_tokens": [ + -4.018191814422607 + ], + "logprob_reward": -0.0181918703019619, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0181918703019619, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14484241604804993, + 0.03638104721903801, + 0.06037415936589241, + 0.04422396048903465, + 0.046392664313316345, + 0.07071177661418915, + 0.05086756497621536, + 0.1468653678894043, + 0.1256428211927414, + 0.27369824051856995 + ], + "credit_map": [ + 0.014008753933012486, + 0.025678614154458046, + 0.008859675377607346, + 0.01485557109117508, + 0.017149411141872406, + 0.025585534051060677, + 0.008486713282763958, + 0.03021812252700329, + 0.0026849820278584957, + 0.003790592774748802, + 0.004862962756305933, + 0.0031223183032125235, + 0.0034411230590194464, + 0.004972474649548531, + 0.007345800753682852, + 0.006160787772387266, + 0.007928446866571903, + 0.008857669308781624, + 0.0069245691411197186, + 0.006507192738354206, + 0.010861369781196117, + 0.006510311737656593, + 0.006195837631821632, + 0.006588757038116455, + 0.0037923227064311504, + 0.00765891233459115, + 0.008460653014481068, + 0.006472749635577202, + 0.003646459663286805, + 0.004286502953618765, + 0.0027721384540200233, + 0.007134219165891409, + 0.0043267980217933655, + 0.005793269257992506, + 0.00588520523160696, + 0.0070473989471793175, + 0.006652841344475746, + 0.006265455856919289, + 0.006108839996159077, + 0.0043128477409482, + 0.004436330404132605, + 0.00714102340862155, + 0.0029034113977104425, + 0.005012177862226963, + 0.006252788007259369, + 0.00778523925691843, + 0.018627695739269257, + 0.018553102388978004, + 0.0068986499682068825, + 0.007988596335053444, + 0.004746952094137669, + 0.006273824255913496, + 0.008388292975723743, + 0.006517126224935055, + 0.005158924963325262, + 0.004895193036645651, + 0.018838712945580482, + 0.01240147091448307, + 0.026889082044363022, + 0.015871796756982803, + 0.012855089269578457, + 0.01205216720700264, + 0.012171491049230099, + 0.03578553721308708, + 0.010669701732695103, + 0.012465260922908783, + 0.018487099558115005, + 0.01789185404777527, + 0.026403171941637993, + 0.012266742065548897, + 0.014908934943377972, + 0.012550042942166328, + 0.033883873373270035, + 0.0294868852943182, + 0.022559458389878273, + 0.03659161552786827, + 0.031086161732673645, + 0.058810558170080185, + 0.0383274219930172, + 0.022952262312173843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.0181918703019619 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An initially static, high-angle, extreme close-up focuses on the dark, glossy surface of steaming black coffee in a plain, white ceramic cup. Distorted within the liquid's surface is the distorted, fish-eye reflection of a weary office worker in his 30s, his slightly disheveled brown hair and gaunt features illuminated by the cool, sterile glow of the overhead fluorescent lights. As he gazes into the cup, his hand, revealing the faint lines of a long day's work, enters the frame, purposefully wrapped around the warm ceramic. Then, the camera begins a slow, steady upward tilt, perfectly synchronized with the cup's movement as he raises it to his face. As the cup is raised, the reflection disappears, and the man himself comes into focus in an eye-level shot; he wears a crisp but slightly wrinkled white dress shirt, the top button undone, his navy blue tie loosened. He closes his eyes, and in a moment of weary anticipation, his head tilted back slightly as he takes a long, deep sip of the drink, his Adam's apple and the slight bob that marks his swallow. The background remains a soft-focus blur of cubicle walls and the glow of computer monitors, creating a palpable atmosphere of quiet solitude and exhaustion, interrupted only by this small, restorative ritual.", + "seed": 24059, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_6/index_8250_seed_24059_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.032062410376966, + "vlm_logprob": -0.032062410376966, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.032062410376966, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up of dark, glossy, steaming black coffee in a white ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a fish-eye reflection of a weary office worker in his 30s with disheveled brown hair and gaunt features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, steady upward tilt synchronized with the cup being raised to the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man shown wearing a white dress shirt with the top button undone and a loosened navy blue tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt his head back slightly to take a long sip, showing his Adam's apple and swallow, against a soft-focus background of cubicle walls and computer monitors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.031054137274622917 + ], + "opposite_logprob_tokens": [ + -3.7810540199279785 + ], + "logprob_reward": -0.031054137274622917, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.031054137274622917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32218700647354126, + 0.1506195068359375, + 0.10792797803878784, + 0.09749459475278854, + 0.05531520023941994, + 0.04392142966389656, + 0.07058770209550858, + 0.051414743065834045, + 0.06501690298318863, + 0.03551488369703293 + ], + "credit_map": [ + 0.03606313094496727, + 0.021605055779218674, + 0.06513572484254837, + 0.02249215543270111, + 0.039322976022958755, + 0.020213479176163673, + 0.08207947015762329, + 0.03527504950761795, + 0.024155547842383385, + 0.018225563690066338, + 0.019015377387404442, + 0.012881508097052574, + 0.021397996693849564, + 0.015263727866113186, + 0.02265884354710579, + 0.017020950093865395, + 0.014118291437625885, + 0.014293363317847252, + 0.015339675359427929, + 0.013869456015527248, + 0.015135332942008972, + 0.01386674027889967, + 0.011541980318725109, + 0.009763134643435478, + 0.00708699831739068, + 0.013783207163214684, + 0.009189995937049389, + 0.010688768699765205, + 0.0077502247877418995, + 0.02671549841761589, + 0.007632054854184389, + 0.014647850766777992, + 0.007025833707302809, + 0.005642574746161699, + 0.006516640540212393, + 0.00995601061731577, + 0.0071633909828960896, + 0.005724239628762007, + 0.005402823910117149, + 0.007883688434958458, + 0.005041284486651421, + 0.0030905765015631914, + 0.004746876657009125, + 0.00645905127748847, + 0.007112124469131231, + 0.006896129809319973, + 0.005369717720896006, + 0.005205669440329075, + 0.006561099085956812, + 0.006939462386071682, + 0.005639111623167992, + 0.00728771835565567, + 0.010221192613244057, + 0.009020477533340454, + 0.014454266987740993, + 0.010464377701282501, + 0.005828748922795057, + 0.005708050914108753, + 0.00782667938619852, + 0.005249655339866877, + 0.007989702746272087, + 0.008368270471692085, + 0.004709829110652208, + 0.005733804311603308, + 0.010135083459317684, + 0.008581292815506458, + 0.006747954059392214, + 0.004784550983458757, + 0.007372282911092043, + 0.017641454935073853, + 0.005362747702747583, + 0.0043915314599871635, + 0.009273095987737179, + 0.003018604824319482, + 0.0036880834959447384, + 0.003230152651667595, + 0.0044264476746320724, + 0.006835087668150663, + 0.0025072721764445305, + 0.00253613805398345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.031054137274622917 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up of dark, glossy, steaming black coffee in a white ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a fish-eye reflection of a weary office worker in his 30s with disheveled brown hair and gaunt features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, steady upward tilt synchronized with the cup being raised to the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man shown wearing a white dress shirt with the top button undone and a loosened navy blue tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt his head back slightly to take a long sip, showing his Adam's apple and swallow, against a soft-focus background of cubicle walls and computer monitors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10046008974313736 + ], + "opposite_logprob_tokens": [ + -2.3504600524902344 + ], + "logprob_reward": -0.10046008974313736, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10046008974313736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14427641034126282, + 0.06744643300771713, + 0.1384822428226471, + 0.11326614022254944, + 0.14204545319080353, + 0.10525859147310257, + 0.06332074850797653, + 0.06431695073843002, + 0.07507374882698059, + 0.08651328086853027 + ], + "credit_map": [ + 0.014286533929407597, + 0.010493305511772633, + 0.026553280651569366, + 0.020989904180169106, + 0.016871867701411247, + 0.010776497423648834, + 0.0313173271715641, + 0.012987680733203888, + 0.01073063351213932, + 0.006968943867832422, + 0.010514563880860806, + 0.005252147559076548, + 0.007468467112630606, + 0.007152441423386335, + 0.009145421907305717, + 0.010213813744485378, + 0.016743766143918037, + 0.016906801611185074, + 0.017162449657917023, + 0.0171163622289896, + 0.02419895865023136, + 0.016170453280210495, + 0.02020159922540188, + 0.009981859475374222, + 0.007921930402517319, + 0.015617587603628635, + 0.010636872611939907, + 0.012526411563158035, + 0.012098563835024834, + 0.0252483282238245, + 0.012617629952728748, + 0.016598815098404884, + 0.024425122886896133, + 0.01170791033655405, + 0.019213559105992317, + 0.02014942467212677, + 0.017759934067726135, + 0.01652858778834343, + 0.010769683867692947, + 0.021491240710020065, + 0.010512999258935452, + 0.007727964781224728, + 0.011324986815452576, + 0.01823236793279648, + 0.015330272726714611, + 0.01588444411754608, + 0.013748358935117722, + 0.012497201561927795, + 0.008329196833074093, + 0.006896447855979204, + 0.006660031154751778, + 0.006663199979811907, + 0.010592292062938213, + 0.010667676106095314, + 0.006344684399664402, + 0.0071672191843390465, + 0.007462185341864824, + 0.007436503656208515, + 0.011882711201906204, + 0.007447234820574522, + 0.006926573347300291, + 0.010812995955348015, + 0.0072123268619179726, + 0.005136424675583839, + 0.0056025199592113495, + 0.01186837162822485, + 0.009018730372190475, + 0.006855142302811146, + 0.009170283563435078, + 0.018005872145295143, + 0.009255764074623585, + 0.005297064781188965, + 0.020628752186894417, + 0.009091373533010483, + 0.00821923278272152, + 0.010936588980257511, + 0.008589686825871468, + 0.014640559442341328, + 0.006150231231004, + 0.008256855420768261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.10046008974313736 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up of dark, glossy, steaming black coffee in a white ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a fish-eye reflection of a weary office worker in his 30s with disheveled brown hair and gaunt features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, steady upward tilt synchronized with the cup being raised to the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man shown wearing a white dress shirt with the top button undone and a loosened navy blue tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt his head back slightly to take a long sip, showing his Adam's apple and swallow, against a soft-focus background of cubicle walls and computer monitors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059745293110609055 + ], + "opposite_logprob_tokens": [ + -5.130974292755127 + ], + "logprob_reward": -0.0059745293110609055, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059745293110609055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0694209560751915, + 0.07324133068323135, + 0.3870707154273987, + 0.17179234325885773, + 0.05287836492061615, + 0.03631935641169548, + 0.04519549757242203, + 0.047162145376205444, + 0.05309131368994713, + 0.0638280063867569 + ], + "credit_map": [ + 0.00936302449554205, + 0.004980092402547598, + 0.008821996860206127, + 0.01184471882879734, + 0.009937377646565437, + 0.005604542791843414, + 0.010368426330387592, + 0.008500778116285801, + 0.012105753645300865, + 0.00759775610640645, + 0.01614142209291458, + 0.0060074543580412865, + 0.007656481582671404, + 0.0071933516301214695, + 0.007741774898022413, + 0.008797330781817436, + 0.04450884833931923, + 0.07617577165365219, + 0.04243507236242294, + 0.05299177020788193, + 0.05297289043664932, + 0.03776297718286514, + 0.048494286835193634, + 0.0317290723323822, + 0.01275571994483471, + 0.019970664754509926, + 0.017522817477583885, + 0.019491059705615044, + 0.017068153247237206, + 0.048699475824832916, + 0.016406841576099396, + 0.019877605140209198, + 0.010923517867922783, + 0.0043290480971336365, + 0.007399639580398798, + 0.00701657310128212, + 0.005062970798462629, + 0.005907148122787476, + 0.005217166151851416, + 0.007022299338132143, + 0.003003144869580865, + 0.0026054545305669308, + 0.003721124492585659, + 0.00824666302651167, + 0.004478144459426403, + 0.004394346382468939, + 0.00462864525616169, + 0.005241834092885256, + 0.005834574345499277, + 0.004290655255317688, + 0.005152574740350246, + 0.005568015389144421, + 0.0060091461054980755, + 0.0071393344551324844, + 0.0062392414547502995, + 0.00496195163577795, + 0.005275165196508169, + 0.004925661254674196, + 0.00378225720487535, + 0.005414682440459728, + 0.010242941789329052, + 0.004481709096580744, + 0.00838702917098999, + 0.004652700386941433, + 0.008833550848066807, + 0.009470215998589993, + 0.005425333511084318, + 0.003888408886268735, + 0.008328564465045929, + 0.008296559564769268, + 0.005088774487376213, + 0.003759905928745866, + 0.011353296227753162, + 0.006515205837786198, + 0.006281880661845207, + 0.006434964016079903, + 0.014330713078379631, + 0.006573084741830826, + 0.005327540449798107, + 0.007011321373283863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.0059745293110609055 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up of dark, glossy, steaming black coffee in a white ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a fish-eye reflection of a weary office worker in his 30s with disheveled brown hair and gaunt features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, steady upward tilt synchronized with the cup being raised to the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man shown wearing a white dress shirt with the top button undone and a loosened navy blue tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt his head back slightly to take a long sip, showing his Adam's apple and swallow, against a soft-focus background of cubicle walls and computer monitors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067454250529408455 + ], + "opposite_logprob_tokens": [ + -5.006745338439941 + ], + "logprob_reward": -0.0067454250529408455, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067454250529408455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10629972815513611, + 0.06155272200703621, + 0.07674917578697205, + 0.10440675169229507, + 0.15387313067913055, + 0.09565472602844238, + 0.21930690109729767, + 0.07588054239749908, + 0.06104947254061699, + 0.04522688686847687 + ], + "credit_map": [ + 0.015705469995737076, + 0.00711458595469594, + 0.016508154571056366, + 0.010867775417864323, + 0.01968715339899063, + 0.007116301450878382, + 0.014608588069677353, + 0.014691703021526337, + 0.012161253951489925, + 0.005350479390472174, + 0.0169124286621809, + 0.004279194865375757, + 0.006050945725291967, + 0.004398815333843231, + 0.006456255912780762, + 0.0059433490969240665, + 0.010788257233798504, + 0.010525279678404331, + 0.011236592195928097, + 0.013748996891081333, + 0.00916756596416235, + 0.00808577798306942, + 0.007193024270236492, + 0.006003684364259243, + 0.00857054814696312, + 0.012102354317903519, + 0.01005544327199459, + 0.01383630745112896, + 0.009826441295444965, + 0.02477264404296875, + 0.01177177857607603, + 0.013471230864524841, + 0.01936551183462143, + 0.019576672464609146, + 0.01811979152262211, + 0.02039491944015026, + 0.022017862647771835, + 0.022747261449694633, + 0.01297787670046091, + 0.018673237413167953, + 0.008117260411381721, + 0.00683231046423316, + 0.008181353099644184, + 0.01792372204363346, + 0.013901232741773129, + 0.013462347909808159, + 0.015790339559316635, + 0.011446154676377773, + 0.02081383764743805, + 0.027919935062527657, + 0.016535647213459015, + 0.021892428398132324, + 0.031247248873114586, + 0.02003282867372036, + 0.05673830583691597, + 0.024126669391989708, + 0.0072818719781935215, + 0.010251256637275219, + 0.006383772008121014, + 0.007735054939985275, + 0.01467717532068491, + 0.00811509694904089, + 0.009941712953150272, + 0.011494601145386696, + 0.00485206488519907, + 0.00945314858108759, + 0.010963549837470055, + 0.004696674644947052, + 0.007108112331479788, + 0.016635246574878693, + 0.004158487543463707, + 0.0031821918673813343, + 0.007295425981283188, + 0.004436420276761055, + 0.004123936872929335, + 0.004153950605541468, + 0.007316404953598976, + 0.005981967784464359, + 0.006932128686457872, + 0.004986647050827742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.0067454250529408455 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up of dark, glossy, steaming black coffee in a white ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a fish-eye reflection of a weary office worker in his 30s with disheveled brown hair and gaunt features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, steady upward tilt synchronized with the cup being raised to the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man shown wearing a white dress shirt with the top button undone and a loosened navy blue tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt his head back slightly to take a long sip, showing his Adam's apple and swallow, against a soft-focus background of cubicle walls and computer monitors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607787050306797 + ], + "opposite_logprob_tokens": [ + -4.141077995300293 + ], + "logprob_reward": -0.01607787050306797, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607787050306797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08854061365127563, + 0.05922944098711014, + 0.07800852507352829, + 0.07731226831674576, + 0.07720030099153519, + 0.08457312732934952, + 0.04205495864152908, + 0.08196140080690384, + 0.14680235087871552, + 0.26431700587272644 + ], + "credit_map": [ + 0.010998138226568699, + 0.007093748077750206, + 0.011312882415950298, + 0.010319016873836517, + 0.014183010905981064, + 0.006068546324968338, + 0.019333450123667717, + 0.009231813251972198, + 0.00960006657987833, + 0.00786831695586443, + 0.01592056080698967, + 0.00470666354522109, + 0.006139402277767658, + 0.005584970116615295, + 0.004725669976323843, + 0.004683786537498236, + 0.01253979280591011, + 0.011361371725797653, + 0.010405500419437885, + 0.015762705355882645, + 0.00961998663842678, + 0.0069980924017727375, + 0.005961543880403042, + 0.00535953463986516, + 0.004669089335948229, + 0.01588735729455948, + 0.005908173508942127, + 0.012431299313902855, + 0.004967723973095417, + 0.013388339430093765, + 0.0050476170144975185, + 0.015012669377028942, + 0.010708585381507874, + 0.009172460995614529, + 0.011906532570719719, + 0.010452643036842346, + 0.012127858586609364, + 0.007861655205488205, + 0.00600651977583766, + 0.008964044973254204, + 0.008813908323645592, + 0.005664633587002754, + 0.011241909116506577, + 0.015069040469825268, + 0.008565934374928474, + 0.014869882725179195, + 0.009643937461078167, + 0.01070388127118349, + 0.004994017072021961, + 0.0042326245456933975, + 0.005387791432440281, + 0.0040224757976830006, + 0.0071245902217924595, + 0.004633668810129166, + 0.008138567209243774, + 0.0035212223883718252, + 0.009040468372404575, + 0.008074382320046425, + 0.017035340890288353, + 0.010174376890063286, + 0.008592468686401844, + 0.015820471569895744, + 0.007427181117236614, + 0.005796713288873434, + 0.013781803660094738, + 0.023402247577905655, + 0.030204083770513535, + 0.012266767211258411, + 0.012960663065314293, + 0.036572497338056564, + 0.009056024253368378, + 0.008558254688978195, + 0.06955738365650177, + 0.028012804687023163, + 0.02011149190366268, + 0.02904391847550869, + 0.031168848276138306, + 0.04831099882721901, + 0.017266053706407547, + 0.02084551565349102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.01607787050306797 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_8/index_tensor([4540, 4540])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_8/index_tensor([4540, 4540])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9a9a1ff209421964ab35c33ca9fb8a2e20837890 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_8/index_tensor([4540, 4540])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, 360-degree rotating camera shot, rendered in a whimsical Disney Pixar style, gracefully spirals downward, capturing a busy yet cozy kitchen scene. The camera's movement centers on a large, dark cast iron cauldron on the stove, its matte surface slightly worn and charmingly imperfect. Inside, water is at a vibrant, roiling boil, with giant, stylized bubbles bursting joyfully on the surface. Thick, cottony clouds of stylized steam billow upward, catching the warm, golden afternoon light streaming in from the nearby paneled windows. Wide, plump, belt-like noodles, with an expressive, chewy texture, dance and tumble merrily in the bubbling water, their shimmering surfaces bending and twisting as if filled with joy. A mesmerizing swirl of color forms around them, with vibrant flecks of ruby-red chili oil and perfectly sliced rings of bright emerald green scallions swirling swiftly and playfully, creating a scene of pure, vivid culinary magic. The backdrop showcases a charmingly cluttered kitchen with warm wooden cabinets and rustic countertops, all bathed in a soft, inviting glow, creating a joyful and welcoming atmosphere.", + "seed": 19683, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_8/index_4540_seed_19683_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.3077582687139511, + "vlm_logprob": -0.3077582687139511, + "vlm_yes_no_margin": 1.7249999523162842, + "vlm_token_credit": -0.3077582687139511, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Disney Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a 360-degree rotating shot that spirals downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark cast iron cauldron visible on the stove with a matte, slightly worn surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wide, plump, belt-like noodles with an expressive, chewy texture visible dancing in the boiling water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a cluttered kitchen featuring warm wooden cabinets and rustic countertops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.478446900844574 + ], + "opposite_logprob_tokens": [ + -0.978446900844574 + ], + "logprob_reward": -0.478446900844574, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.478446900844574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18734881281852722, + 0.052107904106378555, + 0.05043031647801399, + 0.12650911509990692, + 0.13686324656009674, + 0.2614455819129944, + 0.05018864572048187, + 0.032547712326049805, + 0.06162223219871521, + 0.040936414152383804 + ], + "credit_map": [ + 0.014436749741435051, + 0.01901618018746376, + 0.020362766459584236, + 0.036161478608846664, + 0.03236590698361397, + 0.022494062781333923, + 0.026114225387573242, + 0.01639745756983757, + 0.007598194759339094, + 0.006792793981730938, + 0.005076420493423939, + 0.007627036888152361, + 0.00939639937132597, + 0.007876011542975903, + 0.003915390465408564, + 0.003825657768175006, + 0.007566652726382017, + 0.006821244955062866, + 0.0030197815503925085, + 0.008082885295152664, + 0.009936045855283737, + 0.004148080013692379, + 0.006096266210079193, + 0.004759358707815409, + 0.024616727605462074, + 0.029481906443834305, + 0.008167175576090813, + 0.013431819155812263, + 0.016353026032447815, + 0.012171040289103985, + 0.009819303639233112, + 0.012468121945858002, + 0.03882550075650215, + 0.013591512106359005, + 0.005216672085225582, + 0.04023963585495949, + 0.009273506700992584, + 0.007436373736709356, + 0.015746083110570908, + 0.0065339673310518265, + 0.051344480365514755, + 0.033602871000766754, + 0.016362793743610382, + 0.04425278306007385, + 0.046615660190582275, + 0.026502422988414764, + 0.028449920937418938, + 0.014314636588096619, + 0.007353432942181826, + 0.005009775515645742, + 0.003828963963314891, + 0.010066979564726353, + 0.005039773415774107, + 0.005803622771054506, + 0.008074517361819744, + 0.005011583678424358, + 0.005625958554446697, + 0.0038651812355965376, + 0.002774445340037346, + 0.00619553541764617, + 0.0032911733724176884, + 0.002985266735777259, + 0.004331851378083229, + 0.003478302853181958, + 0.00741164805367589, + 0.00770993996411562, + 0.003937057219445705, + 0.004820212256163359, + 0.014032921753823757, + 0.008164431899785995, + 0.004398454446345568, + 0.011147568002343178, + 0.005856209900230169, + 0.003430256387218833, + 0.00289476802572608, + 0.00590552156791091, + 0.007914387620985508, + 0.004862996283918619, + 0.0042244624346494675, + 0.0058478135615587234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.478446900844574 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Disney Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a 360-degree rotating shot that spirals downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark cast iron cauldron visible on the stove with a matte, slightly worn surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wide, plump, belt-like noodles with an expressive, chewy texture visible dancing in the boiling water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a cluttered kitchen featuring warm wooden cabinets and rustic countertops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8260906934738159 + ], + "opposite_logprob_tokens": [ + -0.5760906934738159 + ], + "logprob_reward": -0.8260906934738159, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8260906934738159, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1218428909778595, + 0.08300473541021347, + 0.10488350689411163, + 0.046271659433841705, + 0.4164195656776428, + 0.03761153295636177, + 0.045870836824178696, + 0.04817015677690506, + 0.05984732508659363, + 0.03607781231403351 + ], + "credit_map": [ + 0.015296771191060543, + 0.016526440158486366, + 0.009512786753475666, + 0.015353167429566383, + 0.031364548951387405, + 0.011381049640476704, + 0.010525185614824295, + 0.011882943101227283, + 0.013465623371303082, + 0.012916061095893383, + 0.006733066868036985, + 0.00978532712906599, + 0.018231699243187904, + 0.007487359922379255, + 0.005980552174150944, + 0.008405046537518501, + 0.020371412858366966, + 0.022694267332553864, + 0.009665622375905514, + 0.008835572749376297, + 0.01802034303545952, + 0.008163200691342354, + 0.009872028604149818, + 0.007261060643941164, + 0.00547655439004302, + 0.008146476000547409, + 0.0045439512468874454, + 0.005713941063731909, + 0.006221389397978783, + 0.00474167475476861, + 0.00622539920732379, + 0.00520227150991559, + 0.12294965982437134, + 0.04379652813076973, + 0.020783551037311554, + 0.12911401689052582, + 0.026291916146874428, + 0.01870267651975155, + 0.0391266830265522, + 0.01565452478826046, + 0.008548719808459282, + 0.005677065812051296, + 0.0024715655017644167, + 0.0033720789942890406, + 0.007032151799649, + 0.004012841731309891, + 0.003833892522379756, + 0.002663217019289732, + 0.0073545523919165134, + 0.006179888732731342, + 0.0028065636288374662, + 0.004938081838190556, + 0.009490757249295712, + 0.0054562948644161224, + 0.004833744373172522, + 0.004810956306755543, + 0.00966452807188034, + 0.005776365753263235, + 0.004300469532608986, + 0.008789265528321266, + 0.004411517176777124, + 0.004967670887708664, + 0.005593125708401203, + 0.004667214117944241, + 0.009269003756344318, + 0.009337913244962692, + 0.0036868315655738115, + 0.006585169583559036, + 0.01371940691024065, + 0.006070618983358145, + 0.004582277964800596, + 0.006596100516617298, + 0.005528916139155626, + 0.00270098983310163, + 0.0016799308359622955, + 0.004041021224111319, + 0.0069523463025689125, + 0.006592277903109789, + 0.0032200468704104424, + 0.005362280178815126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.8260906934738159 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Disney Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a 360-degree rotating shot that spirals downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark cast iron cauldron visible on the stove with a matte, slightly worn surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wide, plump, belt-like noodles with an expressive, chewy texture visible dancing in the boiling water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a cluttered kitchen featuring warm wooden cabinets and rustic countertops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903378993272781 + ], + "opposite_logprob_tokens": [ + -2.464033842086792 + ], + "logprob_reward": -0.08903378993272781, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903378993272781, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17418992519378662, + 0.0665413960814476, + 0.12811416387557983, + 0.1233905553817749, + 0.1354902684688568, + 0.12853559851646423, + 0.057103779166936874, + 0.11055444926023483, + 0.05756518617272377, + 0.01851459965109825 + ], + "credit_map": [ + 0.015136566944420338, + 0.015417416580021381, + 0.026429926976561546, + 0.02390509471297264, + 0.023325353860855103, + 0.015312332659959793, + 0.0315861813724041, + 0.02307705767452717, + 0.007883396930992603, + 0.006718830205500126, + 0.013653297908604145, + 0.010022478178143501, + 0.006506064441055059, + 0.007275099866092205, + 0.007455785758793354, + 0.007026445586234331, + 0.024178236722946167, + 0.02541521191596985, + 0.010683148168027401, + 0.013307725079357624, + 0.017553530633449554, + 0.009782540611922741, + 0.01560320146381855, + 0.011590569280087948, + 0.019866356626152992, + 0.021032391116023064, + 0.010880148969590664, + 0.017232079058885574, + 0.01784461922943592, + 0.013066752813756466, + 0.010120224207639694, + 0.013347984291613102, + 0.04174601286649704, + 0.011828877031803131, + 0.006297233980149031, + 0.04289332032203674, + 0.007824393920600414, + 0.006321287713944912, + 0.013366969302296638, + 0.00521217891946435, + 0.016769874840974808, + 0.017724180594086647, + 0.009479865431785583, + 0.025242086499929428, + 0.014264170080423355, + 0.01520018931478262, + 0.02024848572909832, + 0.009606747888028622, + 0.009768775664269924, + 0.0059749349020421505, + 0.004134560935199261, + 0.010636351071298122, + 0.007164346054196358, + 0.007017419673502445, + 0.008315743878483772, + 0.004091646987944841, + 0.019794706255197525, + 0.014071526005864143, + 0.011777914129197598, + 0.021810287609696388, + 0.010199146345257759, + 0.00869800429791212, + 0.015194942243397236, + 0.009007930755615234, + 0.007992553524672985, + 0.013714014552533627, + 0.0037166534457355738, + 0.00845814123749733, + 0.00822275411337614, + 0.006260685622692108, + 0.004439404234290123, + 0.004760978743433952, + 0.002704076934605837, + 0.001838041702285409, + 0.001899660099297762, + 0.0027860281988978386, + 0.0023709184024482965, + 0.0025782554876059294, + 0.0022611217573285103, + 0.0020764966029673815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.08903378993272781 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Disney Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a 360-degree rotating shot that spirals downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark cast iron cauldron visible on the stove with a matte, slightly worn surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wide, plump, belt-like noodles with an expressive, chewy texture visible dancing in the boiling water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a cluttered kitchen featuring warm wooden cabinets and rustic countertops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018189646303653717 + ], + "opposite_logprob_tokens": [ + -4.018189430236816 + ], + "logprob_reward": -0.018189646303653717, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018189646303653717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16454669833183289, + 0.03810761868953705, + 0.03933371976017952, + 0.12654505670070648, + 0.1207595020532608, + 0.23413681983947754, + 0.04649858549237251, + 0.0987243801355362, + 0.07663287967443466, + 0.05471465364098549 + ], + "credit_map": [ + 0.02085454948246479, + 0.021532973274588585, + 0.021407578140497208, + 0.017569497227668762, + 0.03763112053275108, + 0.01589800789952278, + 0.014820595271885395, + 0.014832389540970325, + 0.004380475729703903, + 0.00441016023978591, + 0.005446972791105509, + 0.00645732507109642, + 0.005136815365403891, + 0.00271196523681283, + 0.004932323005050421, + 0.004631581250578165, + 0.005274547263979912, + 0.006337911821901798, + 0.003545931773260236, + 0.007666409946978092, + 0.005826887674629688, + 0.003204192267730832, + 0.004101058002561331, + 0.003376784734427929, + 0.011110902763903141, + 0.015497521497309208, + 0.010033631697297096, + 0.033383771777153015, + 0.01377530861645937, + 0.01159706525504589, + 0.0205879844725132, + 0.010558881796896458, + 0.018226034939289093, + 0.018841983750462532, + 0.007383855991065502, + 0.020163290202617645, + 0.015276500023901463, + 0.011573798023164272, + 0.01812264695763588, + 0.01117140706628561, + 0.04176938906311989, + 0.03231753036379814, + 0.015153407119214535, + 0.04197919741272926, + 0.0385332889854908, + 0.022763743996620178, + 0.028252296149730682, + 0.013367991894483566, + 0.0082895178347826, + 0.005224564112722874, + 0.0029731413815170527, + 0.008679602295160294, + 0.005955792963504791, + 0.0057363552041351795, + 0.006109218113124371, + 0.0035303987096995115, + 0.018940025940537453, + 0.010422573424875736, + 0.009697536937892437, + 0.020032841712236404, + 0.007589313201606274, + 0.00904394406825304, + 0.013179122470319271, + 0.0098190288990736, + 0.008953331038355827, + 0.0061896489933133125, + 0.0042619602754712105, + 0.0063179186545312405, + 0.016647810116410255, + 0.009774751029908657, + 0.005955114960670471, + 0.018532348796725273, + 0.006954962853342295, + 0.004339041654020548, + 0.0041910880245268345, + 0.006122620776295662, + 0.009102418087422848, + 0.008884811773896217, + 0.0049756960943341255, + 0.010144016705453396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.018189646303653717 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Disney Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a 360-degree rotating shot that spirals downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark cast iron cauldron visible on the stove with a matte, slightly worn surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wide, plump, belt-like noodles with an expressive, chewy texture visible dancing in the boiling water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a cluttered kitchen featuring warm wooden cabinets and rustic countertops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703031301498413 + ], + "opposite_logprob_tokens": [ + -2.127030372619629 + ], + "logprob_reward": -0.12703031301498413, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703031301498413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43206125497817993, + 0.055782340466976166, + 0.039402298629283905, + 0.04051685705780983, + 0.07920313626527786, + 0.05353587865829468, + 0.06759490817785263, + 0.10298392921686172, + 0.07187771052122116, + 0.057041656225919724 + ], + "credit_map": [ + 0.06742347776889801, + 0.06596731394529343, + 0.02324073016643524, + 0.0375303216278553, + 0.11821935325860977, + 0.04842556267976761, + 0.03755950555205345, + 0.03369499742984772, + 0.007132645230740309, + 0.006760551128536463, + 0.009591941721737385, + 0.007359962444752455, + 0.007144019939005375, + 0.007918802089989185, + 0.004545474424958229, + 0.005328945815563202, + 0.0052022142335772514, + 0.005538536701351404, + 0.0033917685505002737, + 0.006061420775949955, + 0.006460628937929869, + 0.0038011609576642513, + 0.004590898752212524, + 0.004355668555945158, + 0.0068745845928788185, + 0.008651068434119225, + 0.003218750236555934, + 0.005507314577698708, + 0.00525211077183485, + 0.003984460607171059, + 0.003479215782135725, + 0.0035493487957865, + 0.019529875367879868, + 0.01060594990849495, + 0.004548733588308096, + 0.020142102614045143, + 0.008014630526304245, + 0.004361520521342754, + 0.008354744873940945, + 0.003645580494776368, + 0.007969268597662449, + 0.007038195617496967, + 0.005285803694278002, + 0.009823772124946117, + 0.0062592220492661, + 0.0068442486226558685, + 0.006732461974024773, + 0.0035829117987304926, + 0.014726916328072548, + 0.00731817027553916, + 0.005263833329081535, + 0.00676489993929863, + 0.011189073324203491, + 0.009750280529260635, + 0.005756075959652662, + 0.006825657095760107, + 0.017810771241784096, + 0.012615635059773922, + 0.01070280559360981, + 0.021891269832849503, + 0.008014400489628315, + 0.008332043886184692, + 0.013023576699197292, + 0.010593423619866371, + 0.011944952420890331, + 0.013663976453244686, + 0.003502551931887865, + 0.008586281910538673, + 0.009877805598080158, + 0.011886555701494217, + 0.0049269841983914375, + 0.00748860090970993, + 0.00795151386409998, + 0.005915368907153606, + 0.002800894435495138, + 0.007229756098240614, + 0.007940965704619884, + 0.01212575938552618, + 0.00528509309515357, + 0.0077923075295984745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.12703031301498413 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, 360-degree rotating camera shot, rendered in a whimsical Disney Pixar style, gracefully spirals downward, capturing a busy yet cozy kitchen scene. The camera's movement centers on a large, dark cast iron cauldron on the stove, its matte surface slightly worn and charmingly imperfect. Inside, water is at a vibrant, roiling boil, with giant, stylized bubbles bursting joyfully on the surface. Thick, cottony clouds of stylized steam billow upward, catching the warm, golden afternoon light streaming in from the nearby paneled windows. Wide, plump, belt-like noodles, with an expressive, chewy texture, dance and tumble merrily in the bubbling water, their shimmering surfaces bending and twisting as if filled with joy. A mesmerizing swirl of color forms around them, with vibrant flecks of ruby-red chili oil and perfectly sliced rings of bright emerald green scallions swirling swiftly and playfully, creating a scene of pure, vivid culinary magic. The backdrop showcases a charmingly cluttered kitchen with warm wooden cabinets and rustic countertops, all bathed in a soft, inviting glow, creating a joyful and welcoming atmosphere.", + "seed": 19684, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_8/index_4540_seed_19684_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.35106565412133933, + "vlm_logprob": -0.35106565412133933, + "vlm_yes_no_margin": 1.9, + "vlm_token_credit": -0.35106565412133933, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Disney Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a 360-degree rotating shot that spirals downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark cast iron cauldron visible on the stove with a matte, slightly worn surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wide, plump, belt-like noodles with an expressive, chewy texture visible dancing in the boiling water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a cluttered kitchen featuring warm wooden cabinets and rustic countertops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4368859827518463 + ], + "opposite_logprob_tokens": [ + -1.061885952949524 + ], + "logprob_reward": -0.4368859827518463, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4368859827518463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34839534759521484, + 0.0571262389421463, + 0.06730643659830093, + 0.05945055931806564, + 0.12508746981620789, + 0.09096384793519974, + 0.09132761508226395, + 0.039013125002384186, + 0.07265393435955048, + 0.04867544025182724 + ], + "credit_map": [ + 0.03390858322381973, + 0.049954477697610855, + 0.026769092306494713, + 0.03927384316921234, + 0.0895385667681694, + 0.04623659327626228, + 0.02984512411057949, + 0.032869063317775726, + 0.00593810947611928, + 0.005122427828609943, + 0.003737036371603608, + 0.011411004699766636, + 0.006171470507979393, + 0.011087593622505665, + 0.006959946826100349, + 0.006698646582663059, + 0.008337300270795822, + 0.01080384012311697, + 0.003906416706740856, + 0.010946495458483696, + 0.012579711154103279, + 0.010833045467734337, + 0.005218741949647665, + 0.004680891055613756, + 0.011050017550587654, + 0.007489400915801525, + 0.005718171130865812, + 0.006157927215099335, + 0.012045862153172493, + 0.006147986743599176, + 0.006051566451787949, + 0.004789625760167837, + 0.013612896203994751, + 0.01345102209597826, + 0.014841234311461449, + 0.01854017935693264, + 0.024379609152674675, + 0.01391957700252533, + 0.012228249572217464, + 0.014114701189100742, + 0.017796484753489494, + 0.01320534385740757, + 0.004928145091980696, + 0.010288970544934273, + 0.01912318915128708, + 0.01106928288936615, + 0.007658551912754774, + 0.0068938820622861385, + 0.009876999072730541, + 0.010592109523713589, + 0.011591784656047821, + 0.013386201113462448, + 0.014590061269700527, + 0.00839961040765047, + 0.012982291169464588, + 0.00990855973213911, + 0.003232610411942005, + 0.004529085475951433, + 0.004285405855625868, + 0.007124422118067741, + 0.006103843916207552, + 0.0033457616809755564, + 0.006150603294372559, + 0.0042413934133946896, + 0.011837339960038662, + 0.006331956014037132, + 0.006538841873407364, + 0.00901494175195694, + 0.00943412072956562, + 0.009918369352817535, + 0.008193832822144032, + 0.011384534649550915, + 0.006511845625936985, + 0.004857430700212717, + 0.004472787957638502, + 0.006689503323286772, + 0.010549324564635754, + 0.005978744942694902, + 0.004988823551684618, + 0.0046269819140434265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.4368859827518463 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Disney Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a 360-degree rotating shot that spirals downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark cast iron cauldron visible on the stove with a matte, slightly worn surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wide, plump, belt-like noodles with an expressive, chewy texture visible dancing in the boiling water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a cluttered kitchen featuring warm wooden cabinets and rustic countertops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9742587804794312 + ], + "opposite_logprob_tokens": [ + -0.47425878047943115 + ], + "logprob_reward": -0.9742587804794312, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9742587804794312, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2829884886741638, + 0.17448748648166656, + 0.17954809963703156, + 0.10501612722873688, + 0.11416158080101013, + 0.04772130399942398, + 0.023372499272227287, + 0.02842305228114128, + 0.022048693150281906, + 0.022232653573155403 + ], + "credit_map": [ + 0.03274208679795265, + 0.03270421922206879, + 0.03545558825135231, + 0.032593052834272385, + 0.036201074719429016, + 0.04120926186442375, + 0.025311965495347977, + 0.04677121341228485, + 0.01672982983291149, + 0.01911201886832714, + 0.01137667614966631, + 0.0371270515024662, + 0.02983039990067482, + 0.02627698704600334, + 0.019604356959462166, + 0.014430158771574497, + 0.023927025496959686, + 0.019401922821998596, + 0.011456157080829144, + 0.02564781904220581, + 0.04862411692738533, + 0.02427227422595024, + 0.015065806917846203, + 0.011152985505759716, + 0.013772816397249699, + 0.013416994363069534, + 0.00997899565845728, + 0.0125091802328825, + 0.027692416682839394, + 0.010134917683899403, + 0.009962433949112892, + 0.007548378314822912, + 0.017715727910399437, + 0.012082603760063648, + 0.010883845388889313, + 0.018632903695106506, + 0.018157927319407463, + 0.01214523334056139, + 0.013561525382101536, + 0.010981806553900242, + 0.007861437276005745, + 0.006166599690914154, + 0.00498074758797884, + 0.006548123899847269, + 0.008728533051908016, + 0.00461907172575593, + 0.0047994437627494335, + 0.004017344210296869, + 0.002865757793188095, + 0.00272102328017354, + 0.0018520911689847708, + 0.003949952777475119, + 0.003661978291347623, + 0.0020699910819530487, + 0.003899799194186926, + 0.002351907314732671, + 0.002805035561323166, + 0.003670116188004613, + 0.00259781489148736, + 0.005022277589887381, + 0.004822548944503069, + 0.0025544380769133568, + 0.00404957914724946, + 0.0029012400191277266, + 0.0032318257726728916, + 0.0020455888006836176, + 0.0018622592324391007, + 0.0034137035254389048, + 0.002945060608908534, + 0.0024505150504410267, + 0.002877168357372284, + 0.0032225721515715122, + 0.0022826693020761013, + 0.002017716411501169, + 0.0015082522295415401, + 0.0038139824755489826, + 0.004116085823625326, + 0.0030210716649889946, + 0.003395598381757736, + 0.002077277982607484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.9742587804794312 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Disney Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a 360-degree rotating shot that spirals downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark cast iron cauldron visible on the stove with a matte, slightly worn surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wide, plump, belt-like noodles with an expressive, chewy texture visible dancing in the boiling water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a cluttered kitchen featuring warm wooden cabinets and rustic countertops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018198776990175247 + ], + "opposite_logprob_tokens": [ + -4.0181989669799805 + ], + "logprob_reward": -0.018198776990175247, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018198776990175247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24739348888397217, + 0.07186321914196014, + 0.13697151839733124, + 0.29221999645233154, + 0.08939969539642334, + 0.05349781736731529, + 0.05618545040488243, + 0.011847969144582748, + 0.017161859199404716, + 0.02345893159508705 + ], + "credit_map": [ + 0.027973104268312454, + 0.03305871784687042, + 0.01723523437976837, + 0.024632934480905533, + 0.05479088053107262, + 0.04235439747571945, + 0.02163263037800789, + 0.025715582072734833, + 0.00487815635278821, + 0.008476486429572105, + 0.004912760574370623, + 0.01962941884994507, + 0.010718287900090218, + 0.008102585561573505, + 0.011136648245155811, + 0.004008880350738764, + 0.016882427036762238, + 0.017436353489756584, + 0.008832000195980072, + 0.012825938872992992, + 0.040325552225112915, + 0.017506476491689682, + 0.012717616744339466, + 0.010445154272019863, + 0.033207859843969345, + 0.03770655766129494, + 0.023743856698274612, + 0.0486828051507473, + 0.06659740209579468, + 0.02884257398545742, + 0.03300715610384941, + 0.02043178677558899, + 0.014207614585757256, + 0.009564261883497238, + 0.00927555188536644, + 0.013404165394604206, + 0.013497657142579556, + 0.010281831957399845, + 0.009906452149152756, + 0.009262168779969215, + 0.009255258366465569, + 0.00796233955770731, + 0.004640991799533367, + 0.007107947953045368, + 0.00729746837168932, + 0.00587456626817584, + 0.005845945794135332, + 0.005513296462595463, + 0.007507329806685448, + 0.006210105028003454, + 0.0060839042998850346, + 0.006706210784614086, + 0.009856928139925003, + 0.005331648513674736, + 0.006485882215201855, + 0.00800343882292509, + 0.0012533720582723618, + 0.0010946601396426558, + 0.0015163448406383395, + 0.0019690408371388912, + 0.0017261315369978547, + 0.001016084454022348, + 0.0019648256711661816, + 0.0013075100723654032, + 0.002096515381708741, + 0.001529068686068058, + 0.0012894095852971077, + 0.0022332477383315563, + 0.0025609571021050215, + 0.0020227469503879547, + 0.0017618704587221146, + 0.003668043063953519, + 0.0028680122923105955, + 0.00166377704590559, + 0.002096083015203476, + 0.0034127491526305676, + 0.003712319303303957, + 0.002827226184308529, + 0.002901938743889332, + 0.003976825624704361 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.018198776990175247 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Disney Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a 360-degree rotating shot that spirals downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark cast iron cauldron visible on the stove with a matte, slightly worn surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wide, plump, belt-like noodles with an expressive, chewy texture visible dancing in the boiling water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a cluttered kitchen featuring warm wooden cabinets and rustic countertops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545241974294186 + ], + "opposite_logprob_tokens": [ + -4.387545108795166 + ], + "logprob_reward": -0.012545241974294186, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545241974294186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20917025208473206, + 0.046615712344646454, + 0.13732585310935974, + 0.11542508751153946, + 0.0921701192855835, + 0.07988256961107254, + 0.1613580286502838, + 0.06413207948207855, + 0.05563509464263916, + 0.03828520327806473 + ], + "credit_map": [ + 0.020858071744441986, + 0.025803938508033752, + 0.020594337955117226, + 0.022879566997289658, + 0.04688490182161331, + 0.030673429369926453, + 0.016708115115761757, + 0.02476789988577366, + 0.004826851189136505, + 0.006128484383225441, + 0.0034312792122364044, + 0.008860099129378796, + 0.008038751780986786, + 0.006854071747511625, + 0.005068873520940542, + 0.003407304408028722, + 0.01759023219347, + 0.01522368099540472, + 0.013828564435243607, + 0.017061876133084297, + 0.029387829825282097, + 0.014738214202225208, + 0.015214784070849419, + 0.014280665665864944, + 0.012609835714101791, + 0.015174509026110172, + 0.008585403673350811, + 0.019426200538873672, + 0.029287781566381454, + 0.010226953774690628, + 0.01193640660494566, + 0.008177993819117546, + 0.010822675190865993, + 0.011541462503373623, + 0.010276279412209988, + 0.013953969813883305, + 0.020107848569750786, + 0.009312284179031849, + 0.0076547060161828995, + 0.008500896394252777, + 0.01171957515180111, + 0.009475850500166416, + 0.006832505576312542, + 0.012734060175716877, + 0.011663942597806454, + 0.007826678454875946, + 0.009175756946206093, + 0.010454203002154827, + 0.017457855865359306, + 0.018364084884524345, + 0.017361581325531006, + 0.029950253665447235, + 0.01820611208677292, + 0.015770919620990753, + 0.020647550001740456, + 0.023599671199917793, + 0.004657495766878128, + 0.007550782989710569, + 0.006125092506408691, + 0.01349396351724863, + 0.007644489873200655, + 0.004923482891172171, + 0.011655203066766262, + 0.00808157492429018, + 0.004948237910866737, + 0.004892081953585148, + 0.004654520191252232, + 0.009068036451935768, + 0.007395720575004816, + 0.005699296481907368, + 0.008875568397343159, + 0.01010163314640522, + 0.0047255526296794415, + 0.0037079027388244867, + 0.003970620688050985, + 0.006364094093441963, + 0.005461872089654207, + 0.0032973280176520348, + 0.0063330503180623055, + 0.004424781538546085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.012545241974294186 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Disney Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a 360-degree rotating shot that spirals downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark cast iron cauldron visible on the stove with a matte, slightly worn surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wide, plump, belt-like noodles with an expressive, chewy texture visible dancing in the boiling water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a cluttered kitchen featuring warm wooden cabinets and rustic countertops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134394884109497 + ], + "opposite_logprob_tokens": [ + -1.3134394884109497 + ], + "logprob_reward": -0.3134394884109497, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134394884109497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49579882621765137, + 0.08767296373844147, + 0.09153854846954346, + 0.06670822948217392, + 0.09458163380622864, + 0.05373934283852577, + 0.019745884463191032, + 0.0223858579993248, + 0.029587849974632263, + 0.038240883499383926 + ], + "credit_map": [ + 0.05622483044862747, + 0.08378439396619797, + 0.03461046516895294, + 0.047295086085796356, + 0.11908326297998428, + 0.07200000435113907, + 0.04248158633708954, + 0.040319204330444336, + 0.007839960046112537, + 0.007079164031893015, + 0.006868953816592693, + 0.01823507808148861, + 0.01356220431625843, + 0.013733901083469391, + 0.012145751155912876, + 0.008207953535020351, + 0.012885145843029022, + 0.011570310220122337, + 0.005257916171103716, + 0.01167214009910822, + 0.022374385967850685, + 0.014159937389194965, + 0.00826295092701912, + 0.005355764180421829, + 0.009414498694241047, + 0.009057793766260147, + 0.004603154491633177, + 0.009708914905786514, + 0.01548689603805542, + 0.007698038127273321, + 0.006372367963194847, + 0.0043665640987455845, + 0.01316629908978939, + 0.009617825038731098, + 0.010793831199407578, + 0.015881283208727837, + 0.01484500989317894, + 0.010280271992087364, + 0.009674107655882835, + 0.01032300665974617, + 0.008446973748505116, + 0.006827116943895817, + 0.005592326633632183, + 0.006758089642971754, + 0.00943000428378582, + 0.0055628567934036255, + 0.005962487775832415, + 0.005159489344805479, + 0.0023053190670907497, + 0.0023829687852412462, + 0.0017602852312847972, + 0.0029240085277706385, + 0.003380881855264306, + 0.0015467072371393442, + 0.0034666259307414293, + 0.0019790867809206247, + 0.002149315783753991, + 0.002454320900142193, + 0.0014607763150706887, + 0.0047203004360198975, + 0.0031442055478692055, + 0.0019463448552414775, + 0.004340082872658968, + 0.0021705101244151592, + 0.005252622999250889, + 0.0026189605705440044, + 0.0026074862107634544, + 0.0044434573501348495, + 0.005008326843380928, + 0.0033298837952315807, + 0.0036709243431687355, + 0.0026561887934803963, + 0.004234254360198975, + 0.004109230823814869, + 0.004027398303151131, + 0.005456825252622366, + 0.007339450065046549, + 0.003796051722019911, + 0.004994797986000776, + 0.004282873123884201 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.3134394884109497 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, 360-degree rotating camera shot, rendered in a whimsical Disney Pixar style, gracefully spirals downward, capturing a busy yet cozy kitchen scene. The camera's movement centers on a large, dark cast iron cauldron on the stove, its matte surface slightly worn and charmingly imperfect. Inside, water is at a vibrant, roiling boil, with giant, stylized bubbles bursting joyfully on the surface. Thick, cottony clouds of stylized steam billow upward, catching the warm, golden afternoon light streaming in from the nearby paneled windows. Wide, plump, belt-like noodles, with an expressive, chewy texture, dance and tumble merrily in the bubbling water, their shimmering surfaces bending and twisting as if filled with joy. A mesmerizing swirl of color forms around them, with vibrant flecks of ruby-red chili oil and perfectly sliced rings of bright emerald green scallions swirling swiftly and playfully, creating a scene of pure, vivid culinary magic. The backdrop showcases a charmingly cluttered kitchen with warm wooden cabinets and rustic countertops, all bathed in a soft, inviting glow, creating a joyful and welcoming atmosphere.", + "seed": 19685, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_8/index_4540_seed_19685_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.3316132635809481, + "vlm_logprob": -0.3316132635809481, + "vlm_yes_no_margin": 2.025, + "vlm_token_credit": -0.3316132635809481, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Disney Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a 360-degree rotating shot that spirals downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark cast iron cauldron visible on the stove with a matte, slightly worn surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wide, plump, belt-like noodles with an expressive, chewy texture visible dancing in the boiling water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a cluttered kitchen featuring warm wooden cabinets and rustic countertops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.762434720993042 + ], + "opposite_logprob_tokens": [ + -0.637434720993042 + ], + "logprob_reward": -0.762434720993042, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.762434720993042, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2450931966304779, + 0.08732873201370239, + 0.35781994462013245, + 0.104719378054142, + 0.04190995916724205, + 0.035556063055992126, + 0.01683695800602436, + 0.016464009881019592, + 0.05876027047634125, + 0.03551148250699043 + ], + "credit_map": [ + 0.00802066270262003, + 0.08914361894130707, + 0.009302576072514057, + 0.008154909126460552, + 0.09095576405525208, + 0.027169303968548775, + 0.006133498623967171, + 0.006212885491549969, + 0.002753105014562607, + 0.025184359401464462, + 0.0045352233573794365, + 0.008849742822349072, + 0.02543560042977333, + 0.010532524436712265, + 0.006517479661852121, + 0.0035206954926252365, + 0.010465744882822037, + 0.1517259180545807, + 0.005140524357557297, + 0.004474654793739319, + 0.1552024483680725, + 0.022259686142206192, + 0.004146777093410492, + 0.004404196981340647, + 0.008037828840315342, + 0.02562675252556801, + 0.003952313214540482, + 0.00801463145762682, + 0.03347232565283775, + 0.013885283842682838, + 0.007425860036164522, + 0.004304380156099796, + 0.005142688285559416, + 0.005779034923762083, + 0.003280947683379054, + 0.004621847532689571, + 0.008217787370085716, + 0.007034302223473787, + 0.004250474274158478, + 0.0035828801337629557, + 0.004870683886110783, + 0.005180293694138527, + 0.004111267626285553, + 0.004392425529658794, + 0.004615298472344875, + 0.0049621048383414745, + 0.0032378544565290213, + 0.004186135716736317, + 0.0019813785329461098, + 0.0018982605542987585, + 0.0022831428796052933, + 0.0022312102373689413, + 0.002673064824193716, + 0.0015950206434354186, + 0.0028427618090063334, + 0.0013321199221536517, + 0.0020192337688058615, + 0.0015934741823002696, + 0.00385727034881711, + 0.0026820136699825525, + 0.0015520656015723944, + 0.0017920663813129067, + 0.0016660885885357857, + 0.0013017971068620682, + 0.00487488741055131, + 0.006924876477569342, + 0.005351189989596605, + 0.004999586846679449, + 0.005967624485492706, + 0.01622326485812664, + 0.007526914123445749, + 0.006891925353556871, + 0.001696050283499062, + 0.010379865765571594, + 0.0014784280210733414, + 0.0036572939716279507, + 0.0033173253759741783, + 0.009170981124043465, + 0.003976828884333372, + 0.0018347103614360094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.762434720993042 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Disney Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a 360-degree rotating shot that spirals downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark cast iron cauldron visible on the stove with a matte, slightly worn surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wide, plump, belt-like noodles with an expressive, chewy texture visible dancing in the boiling water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a cluttered kitchen featuring warm wooden cabinets and rustic countertops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233029723167419 + ], + "opposite_logprob_tokens": [ + -0.8983029723167419 + ], + "logprob_reward": -0.5233029723167419, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233029723167419, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04710410535335541, + 0.13988125324249268, + 0.327790230512619, + 0.15565581619739532, + 0.052354320883750916, + 0.17731045186519623, + 0.026898108422756195, + 0.01370044145733118, + 0.03518761321902275, + 0.02411770634353161 + ], + "credit_map": [ + 0.002775257220491767, + 0.012699868530035019, + 0.0025518969632685184, + 0.004451362881809473, + 0.01256615575402975, + 0.005882760044187307, + 0.003630678867921233, + 0.0025461239274591208, + 0.00539760198444128, + 0.05348991975188255, + 0.005416775122284889, + 0.006631691474467516, + 0.05427069589495659, + 0.006565309129655361, + 0.004907590802758932, + 0.0032016781624406576, + 0.00837860070168972, + 0.14002685248851776, + 0.0050101205706596375, + 0.005081393290311098, + 0.1441499888896942, + 0.017517561092972755, + 0.003987854812294245, + 0.003637836780399084, + 0.01325001660734415, + 0.03730883076786995, + 0.006760665215551853, + 0.010339641012251377, + 0.045825038105249405, + 0.025136154145002365, + 0.011050192639231682, + 0.005985275376588106, + 0.006468724925071001, + 0.006337239407002926, + 0.011082605458796024, + 0.00872654840350151, + 0.007080866489559412, + 0.0041715484112501144, + 0.0060504465363919735, + 0.0024363405536860228, + 0.016489453613758087, + 0.03209078311920166, + 0.017371660098433495, + 0.025945302098989487, + 0.028756430372595787, + 0.02437281608581543, + 0.01685548946261406, + 0.015428530052304268, + 0.0026559128891676664, + 0.003304356010630727, + 0.0038241150323301554, + 0.004180786665529013, + 0.0043836659751832485, + 0.0029731113463640213, + 0.0033118478022515774, + 0.0022643129341304302, + 0.0013893345603719354, + 0.0013628361048176885, + 0.001441151020117104, + 0.002238063607364893, + 0.0017652141395956278, + 0.0015020016580820084, + 0.0028276625089347363, + 0.0011741770431399345, + 0.003027592785656452, + 0.0037280104588717222, + 0.002706714440137148, + 0.002753587206825614, + 0.0033140387386083603, + 0.011785767041146755, + 0.004267176613211632, + 0.003604726167395711, + 0.0015041329897940159, + 0.0063653429970145226, + 0.0012297085486352444, + 0.0021619959734380245, + 0.0018734269542619586, + 0.006591371260583401, + 0.0022984063252806664, + 0.0020933211781084538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 3 + ], + "selected_reward": -0.5233029723167419 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Disney Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a 360-degree rotating shot that spirals downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark cast iron cauldron visible on the stove with a matte, slightly worn surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wide, plump, belt-like noodles with an expressive, chewy texture visible dancing in the boiling water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a cluttered kitchen featuring warm wooden cabinets and rustic countertops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007635456044226885 + ], + "opposite_logprob_tokens": [ + -4.882635593414307 + ], + "logprob_reward": -0.007635456044226885, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007635456044226885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28517353534698486, + 0.29369303584098816, + 0.09924735873937607, + 0.04962591454386711, + 0.05249974504113197, + 0.037213295698165894, + 0.038757290691137314, + 0.026161063462495804, + 0.07705460488796234, + 0.04057420790195465 + ], + "credit_map": [ + 0.022480394691228867, + 0.0872284546494484, + 0.014495337381958961, + 0.018074126914143562, + 0.09087568521499634, + 0.021060163155198097, + 0.01613154262304306, + 0.014827779494225979, + 0.01609715074300766, + 0.08393196761608124, + 0.017374079674482346, + 0.02033279277384281, + 0.08350145071744919, + 0.043185390532016754, + 0.018588252365589142, + 0.010681945830583572, + 0.006580443121492863, + 0.03483740612864494, + 0.005294058937579393, + 0.004158083349466324, + 0.03571869060397148, + 0.0062006874941289425, + 0.0037121297791600227, + 0.0027458551339805126, + 0.003061160445213318, + 0.01176462508738041, + 0.0029905016999691725, + 0.005068589001893997, + 0.013422916643321514, + 0.007572044152766466, + 0.0037996317259967327, + 0.0019464395008981228, + 0.006650203373283148, + 0.00816284865140915, + 0.004360172897577286, + 0.006581489462405443, + 0.011521372944116592, + 0.006729052867740393, + 0.005497344303876162, + 0.002997257746756077, + 0.004965862259268761, + 0.005654248408973217, + 0.005316740367561579, + 0.004569215700030327, + 0.005235740449279547, + 0.004646936431527138, + 0.0035340350586920977, + 0.003290515625849366, + 0.0041551170870661736, + 0.006106998305767775, + 0.0038189077749848366, + 0.005880636628717184, + 0.0056094336323440075, + 0.0040640272200107574, + 0.00611708452925086, + 0.003005081554874778, + 0.002899622544646263, + 0.002909586066380143, + 0.0031359847635030746, + 0.004862517584115267, + 0.0029150599148124456, + 0.0039502400904893875, + 0.0034247166477143764, + 0.0020633351523429155, + 0.006229642778635025, + 0.008823316544294357, + 0.0063760108314454556, + 0.006599996704608202, + 0.008254677057266235, + 0.022310050204396248, + 0.010098223574459553, + 0.008362676948308945, + 0.0022452734410762787, + 0.011227876879274845, + 0.002067696303129196, + 0.0030404115095734596, + 0.003154961159452796, + 0.011727572418749332, + 0.0031386951450258493, + 0.003971719183027744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.007635456044226885 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Disney Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a 360-degree rotating shot that spirals downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark cast iron cauldron visible on the stove with a matte, slightly worn surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wide, plump, belt-like noodles with an expressive, chewy texture visible dancing in the boiling water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a cluttered kitchen featuring warm wooden cabinets and rustic countertops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607177034020424 + ], + "opposite_logprob_tokens": [ + -4.141071796417236 + ], + "logprob_reward": -0.01607177034020424, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607177034020424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23894834518432617, + 0.1060771718621254, + 0.21300765872001648, + 0.07594595104455948, + 0.06446871161460876, + 0.05352843552827835, + 0.03568299114704132, + 0.07814476639032364, + 0.08079219609498978, + 0.05340372025966644 + ], + "credit_map": [ + 0.010712745599448681, + 0.08618901669979095, + 0.007905881851911545, + 0.00991799496114254, + 0.08876358717679977, + 0.021647118031978607, + 0.006495675537735224, + 0.007316322065889835, + 0.0030612966511398554, + 0.04133453220129013, + 0.0031076855957508087, + 0.002682929392904043, + 0.042238179594278336, + 0.008900977671146393, + 0.002201485214754939, + 0.0025500901974737644, + 0.00741082988679409, + 0.08861605077981949, + 0.004216471221297979, + 0.004128929227590561, + 0.09156294167041779, + 0.009586588479578495, + 0.004387731663882732, + 0.0030981164891272783, + 0.005676692817360163, + 0.01837332919239998, + 0.004707967396825552, + 0.004547478631138802, + 0.024774890393018723, + 0.009505266323685646, + 0.005229651462286711, + 0.0031306741293519735, + 0.004617653787136078, + 0.012551919557154179, + 0.005267612636089325, + 0.004804682452231646, + 0.018243296071887016, + 0.009336820803582668, + 0.005521851126104593, + 0.0041248733177781105, + 0.004881335888057947, + 0.009625817649066448, + 0.004433339461684227, + 0.008246134966611862, + 0.00785155687481165, + 0.008138359524309635, + 0.005623962264508009, + 0.004727930296212435, + 0.0035390325356274843, + 0.006104261614382267, + 0.0050470782443881035, + 0.003640093607828021, + 0.006012642756104469, + 0.0037065676879137754, + 0.0038639202248305082, + 0.003769396338611841, + 0.007063057739287615, + 0.00822538509964943, + 0.016572006046772003, + 0.017567958682775497, + 0.005991228390485048, + 0.010018588043749332, + 0.008249254897236824, + 0.004457289818674326, + 0.006339459214359522, + 0.009421451948583126, + 0.007320425473153591, + 0.009708947502076626, + 0.009430922567844391, + 0.017389290034770966, + 0.01262288261204958, + 0.00855881255120039, + 0.0028258650563657284, + 0.014357549138367176, + 0.0026674186810851097, + 0.004196530673652887, + 0.005706522148102522, + 0.013890964910387993, + 0.005367015488445759, + 0.004391856025904417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.01607177034020424 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Disney Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a 360-degree rotating shot that spirals downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark cast iron cauldron visible on the stove with a matte, slightly worn surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wide, plump, belt-like noodles with an expressive, chewy texture visible dancing in the boiling water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a cluttered kitchen featuring warm wooden cabinets and rustic countertops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486213982105255 + ], + "opposite_logprob_tokens": [ + -1.2236213684082031 + ], + "logprob_reward": -0.3486213982105255, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486213982105255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09864369034767151, + 0.4133731424808502, + 0.14469893276691437, + 0.08552759140729904, + 0.055526845157146454, + 0.02655000425875187, + 0.027432579547166824, + 0.03495556488633156, + 0.07350197434425354, + 0.039789654314517975 + ], + "credit_map": [ + 0.010064482688903809, + 0.028365066275000572, + 0.006776247173547745, + 0.006200110539793968, + 0.02852327562868595, + 0.007720776367932558, + 0.005734556820243597, + 0.0052591729909181595, + 0.01589803583920002, + 0.15632997453212738, + 0.01887977495789528, + 0.013686569407582283, + 0.15630051493644714, + 0.028870142996311188, + 0.01166529767215252, + 0.011742839589715004, + 0.003167208516970277, + 0.06071672588586807, + 0.003464382840320468, + 0.0035568801686167717, + 0.06261123716831207, + 0.005751390941441059, + 0.0032669531647115946, + 0.00216414756141603, + 0.008042442612349987, + 0.020397990942001343, + 0.004142850637435913, + 0.005127144046127796, + 0.024877242743968964, + 0.013430871069431305, + 0.006395085249096155, + 0.003113966668024659, + 0.0073834373615682125, + 0.009661346673965454, + 0.006336831953376532, + 0.0060104671865701675, + 0.01429703738540411, + 0.005055053159594536, + 0.003762896405532956, + 0.003019777126610279, + 0.0018664401723071933, + 0.0068581136874854565, + 0.002306997310370207, + 0.002520087407901883, + 0.006781763397157192, + 0.0022564709652215242, + 0.0016463741194456816, + 0.0023137580137699842, + 0.0028889065142720938, + 0.00385944708250463, + 0.003277810523286462, + 0.004376178607344627, + 0.004348034504801035, + 0.0030730171129107475, + 0.0028769380878657103, + 0.002732247579842806, + 0.004802817013114691, + 0.0029496653005480766, + 0.005164261907339096, + 0.006353811826556921, + 0.003404241055250168, + 0.005395876709371805, + 0.0035097762010991573, + 0.003375113243237138, + 0.00629099365323782, + 0.007476397790014744, + 0.005813900846987963, + 0.005639417562633753, + 0.007709996309131384, + 0.025066513568162918, + 0.00798815954476595, + 0.007516596931964159, + 0.003235012758523226, + 0.006901902612298727, + 0.0031798214185982943, + 0.003921116702258587, + 0.004085407126694918, + 0.009439484216272831, + 0.003888080595061183, + 0.005138831213116646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.3486213982105255 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, 360-degree rotating camera shot, rendered in a whimsical Disney Pixar style, gracefully spirals downward, capturing a busy yet cozy kitchen scene. The camera's movement centers on a large, dark cast iron cauldron on the stove, its matte surface slightly worn and charmingly imperfect. Inside, water is at a vibrant, roiling boil, with giant, stylized bubbles bursting joyfully on the surface. Thick, cottony clouds of stylized steam billow upward, catching the warm, golden afternoon light streaming in from the nearby paneled windows. Wide, plump, belt-like noodles, with an expressive, chewy texture, dance and tumble merrily in the bubbling water, their shimmering surfaces bending and twisting as if filled with joy. A mesmerizing swirl of color forms around them, with vibrant flecks of ruby-red chili oil and perfectly sliced rings of bright emerald green scallions swirling swiftly and playfully, creating a scene of pure, vivid culinary magic. The backdrop showcases a charmingly cluttered kitchen with warm wooden cabinets and rustic countertops, all bathed in a soft, inviting glow, creating a joyful and welcoming atmosphere.", + "seed": 19686, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_8/index_4540_seed_19686_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.15243166694417595, + "vlm_logprob": -0.15243166694417595, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.15243166694417595, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Disney Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a 360-degree rotating shot that spirals downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark cast iron cauldron visible on the stove with a matte, slightly worn surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wide, plump, belt-like noodles with an expressive, chewy texture visible dancing in the boiling water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a cluttered kitchen featuring warm wooden cabinets and rustic countertops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07453089952468872 + ], + "opposite_logprob_tokens": [ + -2.699530839920044 + ], + "logprob_reward": -0.07453089952468872, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07453089952468872, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17821279168128967, + 0.10725647956132889, + 0.02234761044383049, + 0.04844749718904495, + 0.058385271579027176, + 0.1900959610939026, + 0.27044686675071716, + 0.06376466155052185, + 0.016553455963730812, + 0.04448944330215454 + ], + "credit_map": [ + 0.021931011229753494, + 0.022123165428638458, + 0.013698309659957886, + 0.03218500316143036, + 0.03125254064798355, + 0.017444288358092308, + 0.02014017477631569, + 0.019438274204730988, + 0.010682687163352966, + 0.016568779945373535, + 0.007501927670091391, + 0.02297792211174965, + 0.024319708347320557, + 0.008103061467409134, + 0.010068114846944809, + 0.007034264970570803, + 0.0026227240450680256, + 0.0021232522558420897, + 0.002096629235893488, + 0.003737645223736763, + 0.0030104450415819883, + 0.0027235797606408596, + 0.0033860106486827135, + 0.0026473228354007006, + 0.006480928510427475, + 0.0061904615722596645, + 0.003282892983406782, + 0.005123959854245186, + 0.01011747494339943, + 0.00876226369291544, + 0.003453752724453807, + 0.005035757087171078, + 0.012627080082893372, + 0.015484445728361607, + 0.0031126849353313446, + 0.008399423211812973, + 0.006510543171316385, + 0.0028497076127678156, + 0.005981179885566235, + 0.0034202071838080883, + 0.024281103163957596, + 0.04788321629166603, + 0.01493037585169077, + 0.02661348320543766, + 0.025684623047709465, + 0.018425744026899338, + 0.019549502059817314, + 0.012727905064821243, + 0.02513209544122219, + 0.03375537693500519, + 0.01940450444817543, + 0.037739064544439316, + 0.0431547611951828, + 0.04161227494478226, + 0.039443280547857285, + 0.030205516144633293, + 0.006180065218359232, + 0.013100526295602322, + 0.003046529833227396, + 0.00813322328031063, + 0.008520817384123802, + 0.007405768148601055, + 0.00804800633341074, + 0.009329722262918949, + 0.0024788700975477695, + 0.0023273697588592768, + 0.0013405504869297147, + 0.0027492493391036987, + 0.002157037379220128, + 0.0018028334015980363, + 0.0020059742964804173, + 0.0016915699234232306, + 0.00719456234946847, + 0.004407244734466076, + 0.006297074258327484, + 0.007550653535872698, + 0.00460857804864645, + 0.004398707300424576, + 0.0050552221946418285, + 0.004977394361048937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.07453089952468872 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Disney Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a 360-degree rotating shot that spirals downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark cast iron cauldron visible on the stove with a matte, slightly worn surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wide, plump, belt-like noodles with an expressive, chewy texture visible dancing in the boiling water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a cluttered kitchen featuring warm wooden cabinets and rustic countertops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761018991470337 + ], + "opposite_logprob_tokens": [ + -0.8261018991470337 + ], + "logprob_reward": -0.5761018991470337, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761018991470337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1351304054260254, + 0.03286971524357796, + 0.06656049191951752, + 0.05742831155657768, + 0.16552141308784485, + 0.06020315736532211, + 0.2143731415271759, + 0.11509856581687927, + 0.08326275646686554, + 0.06955201923847198 + ], + "credit_map": [ + 0.014441120438277721, + 0.01626105047762394, + 0.011001354083418846, + 0.029193855822086334, + 0.02309119701385498, + 0.012474359944462776, + 0.016080282628536224, + 0.012587185949087143, + 0.002785372780635953, + 0.003828431712463498, + 0.0027308620046824217, + 0.005441740620881319, + 0.005853390786796808, + 0.0029630810022354126, + 0.004089016001671553, + 0.005177819170057774, + 0.0084375636652112, + 0.00811165664345026, + 0.006271963939070702, + 0.007187933195382357, + 0.006720854435116053, + 0.009637607261538506, + 0.010431270115077496, + 0.009761637076735497, + 0.010138904675841331, + 0.006752632092684507, + 0.00474546430632472, + 0.007858799770474434, + 0.009917831979691982, + 0.007740947417914867, + 0.0044437856413424015, + 0.005829946603626013, + 0.05352615565061569, + 0.06551135331392288, + 0.005737730301916599, + 0.014365573413670063, + 0.010708307847380638, + 0.004540059715509415, + 0.0064387694001197815, + 0.00469345785677433, + 0.007062409073114395, + 0.014348802156746387, + 0.004945285152643919, + 0.008115286938846111, + 0.008651508949697018, + 0.004638994112610817, + 0.00698771933093667, + 0.005453148391097784, + 0.024695521220564842, + 0.03266499936580658, + 0.01811867579817772, + 0.020754700526595116, + 0.038118377327919006, + 0.032550059258937836, + 0.025505831465125084, + 0.02196498028934002, + 0.013614405877888203, + 0.030481286346912384, + 0.0057072811760008335, + 0.011514193378388882, + 0.01460675522685051, + 0.01009775884449482, + 0.009957235306501389, + 0.01911965012550354, + 0.017404012382030487, + 0.013285251334309578, + 0.006341136991977692, + 0.008454984053969383, + 0.009960340335965157, + 0.012858793139457703, + 0.00765911303460598, + 0.007299121469259262, + 0.006313098594546318, + 0.008422238752245903, + 0.007880084216594696, + 0.011509759351611137, + 0.00994888786226511, + 0.009578496217727661, + 0.00961644109338522, + 0.006283016409724951 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.5761018991470337 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Disney Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a 360-degree rotating shot that spirals downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark cast iron cauldron visible on the stove with a matte, slightly worn surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wide, plump, belt-like noodles with an expressive, chewy texture visible dancing in the boiling water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a cluttered kitchen featuring warm wooden cabinets and rustic countertops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003619549795985222 + ], + "opposite_logprob_tokens": [ + -5.62861967086792 + ], + "logprob_reward": -0.003619549795985222, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003619549795985222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09127506613731384, + 0.17371724545955658, + 0.03700694441795349, + 0.0862281396985054, + 0.24032507836818695, + 0.23297527432441711, + 0.054648250341415405, + 0.021679431200027466, + 0.024238625541329384, + 0.03790592402219772 + ], + "credit_map": [ + 0.00768340053036809, + 0.014029477722942829, + 0.009085203520953655, + 0.012511743232607841, + 0.014867976307868958, + 0.009898406453430653, + 0.011308503337204456, + 0.01189036201685667, + 0.01349796075373888, + 0.023831835016608238, + 0.015269048511981964, + 0.03275313228368759, + 0.0317305363714695, + 0.019541708752512932, + 0.018557962030172348, + 0.0185350701212883, + 0.004825177136808634, + 0.0036076745018363, + 0.0027910724747925997, + 0.006220795679837465, + 0.005057597532868385, + 0.0034273893106728792, + 0.006205686368048191, + 0.004871554672718048, + 0.009401309303939342, + 0.015071109868586063, + 0.007884706370532513, + 0.006331238429993391, + 0.02569117583334446, + 0.010212012566626072, + 0.005878221243619919, + 0.005758369341492653, + 0.08118221163749695, + 0.09618605673313141, + 0.008418705314397812, + 0.020736021921038628, + 0.016054784879088402, + 0.005030504893511534, + 0.007094662170857191, + 0.005622152704745531, + 0.030588602647185326, + 0.06214024871587753, + 0.015816031023859978, + 0.03332848846912384, + 0.031100211665034294, + 0.01957249827682972, + 0.023387698456645012, + 0.017041508108377457, + 0.005209654103964567, + 0.006982407532632351, + 0.005950023420155048, + 0.008316051214933395, + 0.007049868814647198, + 0.009142706170678139, + 0.0057246447540819645, + 0.006272900849580765, + 0.003646638011559844, + 0.002466670935973525, + 0.0019173211185261607, + 0.003141880501061678, + 0.0024140821769833565, + 0.0028793434612452984, + 0.002806752920150757, + 0.0024067452177405357, + 0.0030970475636422634, + 0.002614595228806138, + 0.0016635982319712639, + 0.00516686774790287, + 0.00310836685821414, + 0.003223018953576684, + 0.002823190065100789, + 0.0025419448502361774, + 0.006036367267370224, + 0.004288587719202042, + 0.0035842275246977806, + 0.005488688591867685, + 0.004574482794851065, + 0.004391387104988098, + 0.004333436954766512, + 0.005208748392760754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 1 + ], + "selected_reward": -0.003619549795985222 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Disney Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a 360-degree rotating shot that spirals downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark cast iron cauldron visible on the stove with a matte, slightly worn surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wide, plump, belt-like noodles with an expressive, chewy texture visible dancing in the boiling water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a cluttered kitchen featuring warm wooden cabinets and rustic countertops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007626938167959452 + ], + "opposite_logprob_tokens": [ + -4.882627010345459 + ], + "logprob_reward": -0.007626938167959452, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007626938167959452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07901376485824585, + 0.06707508862018585, + 0.02133009023964405, + 0.047023553401231766, + 0.46602845191955566, + 0.06788076460361481, + 0.05352672562003136, + 0.10614626854658127, + 0.049944400787353516, + 0.04203085973858833 + ], + "credit_map": [ + 0.011697276495397091, + 0.010644279420375824, + 0.007505430839955807, + 0.011747682467103004, + 0.012188891880214214, + 0.008240405470132828, + 0.008634060621261597, + 0.008355734869837761, + 0.005355942528694868, + 0.009430957958102226, + 0.0056336913257837296, + 0.015831435099244118, + 0.012344501912593842, + 0.005225756671279669, + 0.007422549184411764, + 0.0058302488178014755, + 0.0025618334766477346, + 0.002547571202740073, + 0.0023803627118468285, + 0.0035104257985949516, + 0.0029733420815318823, + 0.002464868128299713, + 0.0028218936640769243, + 0.0020697941072285175, + 0.004678912460803986, + 0.008012036792933941, + 0.004702948499470949, + 0.004392051137983799, + 0.013804865069687366, + 0.005320814438164234, + 0.0032485807314515114, + 0.0028633440379053354, + 0.15503795444965363, + 0.1901342272758484, + 0.014023824594914913, + 0.040815144777297974, + 0.033676039427518845, + 0.010600343346595764, + 0.012970378622412682, + 0.008770517073571682, + 0.007313166745007038, + 0.01465985644608736, + 0.005409988109022379, + 0.00999966636300087, + 0.008969968184828758, + 0.00682140002027154, + 0.008452587760984898, + 0.006254131905734539, + 0.006992586888372898, + 0.0061904811300337315, + 0.005516205448657274, + 0.009131669998168945, + 0.005949494894593954, + 0.005515040829777718, + 0.008795831352472305, + 0.0054354132153093815, + 0.01449668686836958, + 0.011791818775236607, + 0.007060649339109659, + 0.013833069242537022, + 0.015512293204665184, + 0.011374050751328468, + 0.022199951112270355, + 0.009877748787403107, + 0.006567794363945723, + 0.007068604696542025, + 0.00355735095217824, + 0.00824989378452301, + 0.006522074807435274, + 0.005985582247376442, + 0.006684166379272938, + 0.005308931227773428, + 0.00563027523458004, + 0.005266362801194191, + 0.0054021417163312435, + 0.006146055646240711, + 0.005328756291419268, + 0.005168695468455553, + 0.00485775712877512, + 0.004230812191963196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.007626938167959452 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Disney Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a 360-degree rotating shot that spirals downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark cast iron cauldron visible on the stove with a matte, slightly worn surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wide, plump, belt-like noodles with an expressive, chewy texture visible dancing in the boiling water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a cluttered kitchen featuring warm wooden cabinets and rustic countertops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10027904808521271 + ], + "opposite_logprob_tokens": [ + -2.3502790927886963 + ], + "logprob_reward": -0.10027904808521271, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10027904808521271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25057050585746765, + 0.11841170489788055, + 0.09403375536203384, + 0.08982612937688828, + 0.02757549285888672, + 0.05513855069875717, + 0.0933496281504631, + 0.13210096955299377, + 0.09895621240139008, + 0.04003707692027092 + ], + "credit_map": [ + 0.028755011036992073, + 0.030251121148467064, + 0.02468501403927803, + 0.05396053567528725, + 0.03851690888404846, + 0.024847520515322685, + 0.029749788343906403, + 0.019804563373327255, + 0.01229267381131649, + 0.02065674588084221, + 0.008339092135429382, + 0.018533142283558846, + 0.0267069973051548, + 0.008704344742000103, + 0.013160111382603645, + 0.010018587112426758, + 0.010796334594488144, + 0.009954513050615788, + 0.008394274860620499, + 0.01668063923716545, + 0.01496619451791048, + 0.009511466138064861, + 0.011714701540768147, + 0.012015622109174728, + 0.013208855874836445, + 0.011406239122152328, + 0.006099502556025982, + 0.010860334150493145, + 0.017804540693759918, + 0.011048011481761932, + 0.010160621255636215, + 0.009238015860319138, + 0.0037334472872316837, + 0.007402608171105385, + 0.00275039323605597, + 0.00533861480653286, + 0.00317003414966166, + 0.0017115824157372117, + 0.002101606922224164, + 0.0013672026107087731, + 0.006839497480541468, + 0.00857273768633604, + 0.004132709000259638, + 0.009212839417159557, + 0.006811647675931454, + 0.006042329594492912, + 0.008751293644309044, + 0.004775495268404484, + 0.009014014154672623, + 0.011824117042124271, + 0.008052609860897064, + 0.011322984471917152, + 0.015684006735682487, + 0.01505754142999649, + 0.013300820253789425, + 0.00909352395683527, + 0.014671770855784416, + 0.029793206602334976, + 0.006997340824455023, + 0.019731640815734863, + 0.015118568204343319, + 0.01429341547191143, + 0.012816345319151878, + 0.01867867074906826, + 0.012681268155574799, + 0.012291218154132366, + 0.007919046096503735, + 0.01918446272611618, + 0.011966357938945293, + 0.010320681147277355, + 0.012952504679560661, + 0.011640669777989388, + 0.005331798456609249, + 0.005337501410394907, + 0.004568182863295078, + 0.005905570462346077, + 0.006288595497608185, + 0.004247662611305714, + 0.005103354342281818, + 0.0032544087152928114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.10027904808521271 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_9/index_tensor([2806, 2806])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_9/index_tensor([2806, 2806])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ee46b4afae72c9c15e507f25d6531442cadd3429 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_9/index_tensor([2806, 2806])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "steady, slightly elevated wide-angle lens slowly tilts right, capturing a vibrant marina bathed in the warm glow of the afternoon sun. Countless sailboats, hoisting a row of colorful sails, are moored at the wooden pier, their masts forming a forest against the sky. In the foreground, a sleek sailboat with a striking red and black sail begins its journey; the sails are taut with the wind, and across their black sections, the words \"Flying Emirates\" are prominently emblazoned in bold white lettering. As the boat smoothly and gracefully departs the dock, it gently tilts, its white fiberglass hull gliding smoothly through the gentle cuts of the sea, heading out into the vast expanse. As the camera follows its movement, a vast modern cityscape unfolds in the background, featuring towering skyscrapers and a majestic suspension bridge, their shapes softened by the golden hour haze. In the midfield, another sailboat with a similar red and black sail also pulls away from the dock, creating a sense of parallel movement. Other boats with red and white sails dot the background, adding depth and the feeling of a lively regatta as the leading sailboat glides further into the open water.", + "seed": 29066, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_9/index_2806_seed_29066_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01405518688261509, + "vlm_logprob": -0.01405518688261509, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.01405518688261509, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a steady, slightly elevated wide-angle view that slowly tilts to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a vibrant marina during the afternoon with warm sunlight and a golden hour haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sleek sailboat in the foreground with a red and black sail featuring the white text 'Flying Emirates'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground sailboat depart the wooden pier with taut sails, tilting gently as its white fiberglass hull glides through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast modern cityscape with towering skyscrapers and a suspension bridge visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029169822111725807 + ], + "opposite_logprob_tokens": [ + -4.779170036315918 + ], + "logprob_reward": -0.029169822111725807, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.029169822111725807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07191962003707886, + 0.02832968346774578, + 0.2308797687292099, + 0.043553102761507034, + 0.043003469705581665, + 0.06349041312932968, + 0.07558106631040573, + 0.13181068003177643, + 0.019482288509607315, + 0.2919498383998871 + ], + "credit_map": [ + 0.010805202648043633, + 0.009830075316131115, + 0.010913965292274952, + 0.013131973333656788, + 0.007064745295792818, + 0.006467892322689295, + 0.005891934037208557, + 0.007813840173184872, + 0.005598977208137512, + 0.003924382384866476, + 0.00378029583953321, + 0.0025875663850456476, + 0.0033041031565517187, + 0.0038638636469841003, + 0.0025253493804484606, + 0.002745148492977023, + 0.08023346960544586, + 0.012827882543206215, + 0.005885627120733261, + 0.011745304800570011, + 0.008603209629654884, + 0.09644129127264023, + 0.008002749644219875, + 0.007140247616916895, + 0.009059392847120762, + 0.005343191791325808, + 0.004522865172475576, + 0.004541771486401558, + 0.004555871710181236, + 0.006373098585754633, + 0.00542164733633399, + 0.0037352661602199078, + 0.006888074800372124, + 0.004778637550771236, + 0.007675750181078911, + 0.004674861673265696, + 0.002844186034053564, + 0.006073566619306803, + 0.003385020885616541, + 0.006683372892439365, + 0.014035925269126892, + 0.006879148539155722, + 0.007181533612310886, + 0.008664176799356937, + 0.003993055783212185, + 0.009809957817196846, + 0.00462484173476696, + 0.00830177403986454, + 0.01403107400983572, + 0.012752617709338665, + 0.008524156175553799, + 0.005379597190767527, + 0.009371955879032612, + 0.0122001301497221, + 0.005820100661367178, + 0.0075014433823525906, + 0.027025913819670677, + 0.010946648195385933, + 0.011827146634459496, + 0.012645754963159561, + 0.021369008347392082, + 0.020334448665380478, + 0.01254673209041357, + 0.01511503104120493, + 0.0032675948459655046, + 0.001759098144248128, + 0.002011929638683796, + 0.0019598586950451136, + 0.002643100917339325, + 0.003252378199249506, + 0.001737650134600699, + 0.002850678749382496, + 0.017558980733156204, + 0.004799427464604378, + 0.0034138888586312532, + 0.003699138993397355, + 0.1271277666091919, + 0.12522046267986298, + 0.005334802903234959, + 0.004795406013727188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 7 + ], + "selected_reward": -0.029169822111725807 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a steady, slightly elevated wide-angle view that slowly tilts to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a vibrant marina during the afternoon with warm sunlight and a golden hour haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sleek sailboat in the foreground with a red and black sail featuring the white text 'Flying Emirates'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground sailboat depart the wooden pier with taut sails, tilting gently as its white fiberglass hull glides through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast modern cityscape with towering skyscrapers and a suspension bridge visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202469028532505 + ], + "opposite_logprob_tokens": [ + -4.26420259475708 + ], + "logprob_reward": -0.014202469028532505, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202469028532505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028406349942088127, + 0.015668367967009544, + 0.7782347202301025, + 0.025567954406142235, + 0.019761668518185616, + 0.0038286743219941854, + 0.0054491255432367325, + 0.027160588651895523, + 0.001922272494994104, + 0.09400025755167007 + ], + "credit_map": [ + 0.004424619022756815, + 0.0036469718907028437, + 0.006945427972823381, + 0.006825619377195835, + 0.0011454052291810513, + 0.0023296019062399864, + 0.0014058571541681886, + 0.0016828459920361638, + 0.0021378665696829557, + 0.002016544109210372, + 0.0015208199620246887, + 0.0021169374231249094, + 0.0016984428511932492, + 0.0022708114702254534, + 0.001747711910866201, + 0.0021592339035123587, + 0.28061339259147644, + 0.04002191871404648, + 0.017591865733265877, + 0.03930433467030525, + 0.022856950759887695, + 0.3324662446975708, + 0.023902999237179756, + 0.02147703431546688, + 0.005540119484066963, + 0.0034868752118200064, + 0.0034935115836560726, + 0.002326754154637456, + 0.0018124307971447706, + 0.003941628150641918, + 0.0022867426741868258, + 0.0026798932813107967, + 0.0048197596333920956, + 0.0022070403210818768, + 0.0024159362073987722, + 0.0021250429563224316, + 0.0012106253998354077, + 0.0035173017531633377, + 0.0015754752093926072, + 0.001890487503260374, + 0.0005492225172929466, + 0.0005168122588656843, + 0.00043016448034904897, + 0.00048075438826344907, + 0.00037513714050874114, + 0.000580400344915688, + 0.0003690343874040991, + 0.0005271486006677151, + 0.0007509459392167628, + 0.000970853550825268, + 0.0005376091576181352, + 0.00048477714881300926, + 0.0004942662199027836, + 0.000847577175591141, + 0.0007838315796107054, + 0.000579264888074249, + 0.004773594904690981, + 0.002418055897578597, + 0.002282568486407399, + 0.0030571676325052977, + 0.004435309674590826, + 0.0041482457891106606, + 0.0028717806562781334, + 0.0031738660763949156, + 0.0005141666624695063, + 0.00010977512283716351, + 0.00021509351790882647, + 0.00017076253425329924, + 0.00024582567857578397, + 0.00033901838469319046, + 0.00014211311645340174, + 0.0001855173904914409, + 0.00634178938344121, + 0.0011854154290631413, + 0.0008325627422891557, + 0.0010944424429908395, + 0.04152438044548035, + 0.04025130346417427, + 0.0015809587202966213, + 0.0011894026538357139 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.014202469028532505 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a steady, slightly elevated wide-angle view that slowly tilts to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a vibrant marina during the afternoon with warm sunlight and a golden hour haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sleek sailboat in the foreground with a red and black sail featuring the white text 'Flying Emirates'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground sailboat depart the wooden pier with taut sails, tilting gently as its white fiberglass hull glides through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast modern cityscape with towering skyscrapers and a suspension bridge visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004094788804650307 + ], + "opposite_logprob_tokens": [ + -5.50409460067749 + ], + "logprob_reward": -0.004094788804650307, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004094788804650307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02511606365442276, + 0.01255885325372219, + 0.1955512911081314, + 0.00796960573643446, + 0.007513042073696852, + 0.010216563940048218, + 0.02133193053305149, + 0.05780038237571716, + 0.010552349500358105, + 0.6513899564743042 + ], + "credit_map": [ + 0.004946135450154543, + 0.003816235577687621, + 0.002588005969300866, + 0.002430862747132778, + 0.0026861331425607204, + 0.002727321581915021, + 0.0030215312726795673, + 0.0028998381458222866, + 0.002540561370551586, + 0.0020714562851935625, + 0.0013270701747387648, + 0.0009993964340537786, + 0.0012772534973919392, + 0.002172175794839859, + 0.0011518693063408136, + 0.0010190694592893124, + 0.06905519962310791, + 0.01111193373799324, + 0.0051934365183115005, + 0.009832000359892845, + 0.007529300171881914, + 0.08012446761131287, + 0.006947947200387716, + 0.005756994243711233, + 0.00168532261159271, + 0.001151100266724825, + 0.0007651610067114234, + 0.0007925030658952892, + 0.0008127220207825303, + 0.0012854321394115686, + 0.0009193008881993592, + 0.0005580635042861104, + 0.0012230384163558483, + 0.000995713984593749, + 0.0008254467975348234, + 0.0007537942728959024, + 0.0009051229571923614, + 0.0011516015511006117, + 0.0008481394615955651, + 0.0008101846906356514, + 0.0014125820016488433, + 0.0015757721848785877, + 0.0010558445937931538, + 0.0011976259993389249, + 0.0010719646234065294, + 0.0014340626075863838, + 0.0010556451743468642, + 0.0014130668714642525, + 0.0025091965217143297, + 0.004331325180828571, + 0.0016093684826046228, + 0.0017379916971549392, + 0.0028224217239767313, + 0.0036691331770271063, + 0.002595240017399192, + 0.002057252684608102, + 0.012176379561424255, + 0.004704282619059086, + 0.006052524782717228, + 0.004446461331099272, + 0.00911715067923069, + 0.008893921971321106, + 0.005178680643439293, + 0.007230977062135935, + 0.0014256061986088753, + 0.0009138731984421611, + 0.0008787197875790298, + 0.0017881972016766667, + 0.0014125300804153085, + 0.001742656691931188, + 0.0014429378788918257, + 0.0009478284628130496, + 0.044323913753032684, + 0.007619679439812899, + 0.005748707801103592, + 0.006995115429162979, + 0.28781282901763916, + 0.2792092263698578, + 0.011034092865884304, + 0.008646344766020775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 7 + ], + "selected_reward": -0.004094788804650307 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a steady, slightly elevated wide-angle view that slowly tilts to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a vibrant marina during the afternoon with warm sunlight and a golden hour haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sleek sailboat in the foreground with a red and black sail featuring the white text 'Flying Emirates'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground sailboat depart the wooden pier with taut sails, tilting gently as its white fiberglass hull glides through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast modern cityscape with towering skyscrapers and a suspension bridge visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008630831725895405 + ], + "opposite_logprob_tokens": [ + -4.758630752563477 + ], + "logprob_reward": -0.008630831725895405, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008630831725895405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04830928146839142, + 0.021079134196043015, + 0.356813907623291, + 0.01933676190674305, + 0.03685103356838226, + 0.03486533835530281, + 0.06361386179924011, + 0.23802313208580017, + 0.019301079213619232, + 0.16180646419525146 + ], + "credit_map": [ + 0.00705413892865181, + 0.0063491868786513805, + 0.00790264829993248, + 0.006400829181075096, + 0.005251669324934483, + 0.005767310969531536, + 0.005365035496652126, + 0.004218466114252806, + 0.0032229293137788773, + 0.003001820994541049, + 0.0018899672431871295, + 0.0020941980183124542, + 0.0025332269724458456, + 0.0031430015806108713, + 0.0030201890040189028, + 0.0021738032810389996, + 0.12855841219425201, + 0.018674174323678017, + 0.008175114169716835, + 0.01815829798579216, + 0.010397150181233883, + 0.1516454666852951, + 0.011293846182525158, + 0.009911463595926762, + 0.0035143750719726086, + 0.0030888188630342484, + 0.0017025668639689684, + 0.0018488602945581079, + 0.0024700991343706846, + 0.003181261010468006, + 0.001828740001656115, + 0.0017020405502989888, + 0.007495697122067213, + 0.005860902834683657, + 0.004160059150308371, + 0.004567794967442751, + 0.0033393821213394403, + 0.006450501270592213, + 0.002561423694714904, + 0.0024152689147740602, + 0.004566645715385675, + 0.005485972855240107, + 0.003832356072962284, + 0.004190175794064999, + 0.004880718886852264, + 0.004922049585729837, + 0.002885189140215516, + 0.004102230537682772, + 0.008679485879838467, + 0.010352232493460178, + 0.00715654157102108, + 0.005612567532807589, + 0.007119929883629084, + 0.009737120941281319, + 0.007288953755050898, + 0.007667032536119223, + 0.04706352576613426, + 0.020592818036675453, + 0.02183961123228073, + 0.0231565423309803, + 0.03813944384455681, + 0.036560505628585815, + 0.02310291677713394, + 0.027567774057388306, + 0.00230862433090806, + 0.0022956382017582655, + 0.0019619613885879517, + 0.0032059308141469955, + 0.002890606177970767, + 0.0030401197727769613, + 0.0019990517757833004, + 0.0015991464024409652, + 0.01578880101442337, + 0.007373142521828413, + 0.005353153217583895, + 0.004659662954509258, + 0.059218522161245346, + 0.05654524266719818, + 0.00700565567240119, + 0.005862280260771513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 9 + ], + "selected_reward": -0.008630831725895405 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a steady, slightly elevated wide-angle view that slowly tilts to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a vibrant marina during the afternoon with warm sunlight and a golden hour haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sleek sailboat in the foreground with a red and black sail featuring the white text 'Flying Emirates'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground sailboat depart the wooden pier with taut sails, tilting gently as its white fiberglass hull glides through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast modern cityscape with towering skyscrapers and a suspension bridge visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014178022742271423 + ], + "opposite_logprob_tokens": [ + -4.2641777992248535 + ], + "logprob_reward": -0.014178022742271423, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014178022742271423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12943144142627716, + 0.0556049644947052, + 0.1832912266254425, + 0.02435925416648388, + 0.02475735731422901, + 0.01736167073249817, + 0.016392989084124565, + 0.031245818361639977, + 0.014364304952323437, + 0.5031909942626953 + ], + "credit_map": [ + 0.028138328343629837, + 0.019743941724300385, + 0.024215402081608772, + 0.020187631249427795, + 0.005798767786473036, + 0.014723943546414375, + 0.0058819204568862915, + 0.01074148714542389, + 0.00906915869563818, + 0.008482983335852623, + 0.006247793789952993, + 0.005481242667883635, + 0.0033599771559238434, + 0.012194500304758549, + 0.0035495779011398554, + 0.007219722028821707, + 0.06270620226860046, + 0.012265386059880257, + 0.005146834533661604, + 0.009244189597666264, + 0.008957160636782646, + 0.0705353245139122, + 0.006845829542726278, + 0.007590275723487139, + 0.004964119754731655, + 0.0031328946352005005, + 0.0026328791864216328, + 0.0025020965840667486, + 0.0022486322559416294, + 0.004108482040464878, + 0.0028213050682097673, + 0.0019488426623865962, + 0.003835099982097745, + 0.0047464752569794655, + 0.0029873696621507406, + 0.0029403255321085453, + 0.0025062565691769123, + 0.003916775807738304, + 0.0018102010944858193, + 0.0020148500334471464, + 0.002917961683124304, + 0.002447721315547824, + 0.002312375232577324, + 0.0017720286268740892, + 0.0017122785793617368, + 0.0024719845969229937, + 0.0014867469435557723, + 0.002240571891888976, + 0.0035183734726160765, + 0.001858858042396605, + 0.0021950385998934507, + 0.0012121060863137245, + 0.001996393548324704, + 0.002555143553763628, + 0.0013237388338893652, + 0.001733335666358471, + 0.006057942286133766, + 0.0031407284550368786, + 0.002662659389898181, + 0.003269926179200411, + 0.005129901226609945, + 0.004729785490781069, + 0.002655702643096447, + 0.0035991675686091185, + 0.0033585724886506796, + 0.0010390226962044835, + 0.0017867566784843802, + 0.0012116574216634035, + 0.00203902181237936, + 0.0024783622939139605, + 0.0011320087360218167, + 0.001318901078775525, + 0.03580009192228317, + 0.006535917054861784, + 0.005426814313977957, + 0.0056920358911156654, + 0.22056731581687927, + 0.21290011703968048, + 0.009322017431259155, + 0.0069467173889279366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.014178022742271423 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, slightly elevated wide-angle lens slowly tilts right, capturing a vibrant marina bathed in the warm glow of the afternoon sun. Countless sailboats, hoisting a row of colorful sails, are moored at the wooden pier, their masts forming a forest against the sky. In the foreground, a sleek sailboat with a striking red and black sail begins its journey; the sails are taut with the wind, and across their black sections, the words \"Flying Emirates\" are prominently emblazoned in bold white lettering. As the boat smoothly and gracefully departs the dock, it gently tilts, its white fiberglass hull gliding smoothly through the gentle cuts of the sea, heading out into the vast expanse. As the camera follows its movement, a vast modern cityscape unfolds in the background, featuring towering skyscrapers and a majestic suspension bridge, their shapes softened by the golden hour haze. In the midfield, another sailboat with a similar red and black sail also pulls away from the dock, creating a sense of parallel movement. Other boats with red and white sails dot the background, adding depth and the feeling of a lively regatta as the leading sailboat glides further into the open water.", + "seed": 29067, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_9/index_2806_seed_29067_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03896106444299221, + "vlm_logprob": -0.03896106444299221, + "vlm_yes_no_margin": 3.7749999523162843, + "vlm_token_credit": -0.03896106444299221, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a steady, slightly elevated wide-angle view that slowly tilts to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a vibrant marina during the afternoon with warm sunlight and a golden hour haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sleek sailboat in the foreground with a red and black sail featuring the white text 'Flying Emirates'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground sailboat depart the wooden pier with taut sails, tilting gently as its white fiberglass hull glides through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast modern cityscape with towering skyscrapers and a suspension bridge visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.032313112169504166 + ], + "opposite_logprob_tokens": [ + -4.282313346862793 + ], + "logprob_reward": -0.032313112169504166, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.032313112169504166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1683453768491745, + 0.056527674198150635, + 0.15125544369220734, + 0.07871844619512558, + 0.051847558468580246, + 0.12332629412412643, + 0.09676162153482437, + 0.09144191443920135, + 0.09462229907512665, + 0.087153360247612 + ], + "credit_map": [ + 0.029174359515309334, + 0.02680157870054245, + 0.01595718413591385, + 0.013268891721963882, + 0.026087770238518715, + 0.028656160458922386, + 0.012986855581402779, + 0.015412572771310806, + 0.008001964539289474, + 0.007987977005541325, + 0.006102063227444887, + 0.006208250764757395, + 0.008758111856877804, + 0.007753355428576469, + 0.006946549750864506, + 0.004769402090460062, + 0.02472655475139618, + 0.026941267773509026, + 0.015374485403299332, + 0.016796480864286423, + 0.01708376035094261, + 0.026927197352051735, + 0.009215575642883778, + 0.014190132729709148, + 0.00878502894192934, + 0.010983915999531746, + 0.00739712780341506, + 0.007163839414715767, + 0.013323194347321987, + 0.011555565521121025, + 0.01084231398999691, + 0.008667460642755032, + 0.005922294221818447, + 0.009643251076340675, + 0.0058028376661241055, + 0.005975157953798771, + 0.006046898663043976, + 0.0077214897610247135, + 0.00608005141839385, + 0.00465557724237442, + 0.029543045908212662, + 0.023221733048558235, + 0.016967255622148514, + 0.009737030602991581, + 0.00684867100790143, + 0.020831400528550148, + 0.006445490289479494, + 0.009731666184961796, + 0.015551583841443062, + 0.016433820128440857, + 0.013217421248555183, + 0.012579002417623997, + 0.011175992898643017, + 0.011763831600546837, + 0.009312103502452374, + 0.006727870088070631, + 0.012982477433979511, + 0.011489585041999817, + 0.01214577816426754, + 0.008710192516446114, + 0.009985702112317085, + 0.012163059785962105, + 0.013706971891224384, + 0.010258148424327374, + 0.012480227276682854, + 0.014657488092780113, + 0.013022687286138535, + 0.009960683062672615, + 0.012747643515467644, + 0.010269933380186558, + 0.011068620719015598, + 0.010415014810860157, + 0.010749006643891335, + 0.010473998263478279, + 0.009788352996110916, + 0.010379812680184841, + 0.009325092658400536, + 0.010598846711218357, + 0.011836608871817589, + 0.01400163397192955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.032313112169504166 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a steady, slightly elevated wide-angle view that slowly tilts to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a vibrant marina during the afternoon with warm sunlight and a golden hour haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sleek sailboat in the foreground with a red and black sail featuring the white text 'Flying Emirates'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground sailboat depart the wooden pier with taut sails, tilting gently as its white fiberglass hull glides through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast modern cityscape with towering skyscrapers and a suspension bridge visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038105905055999756 + ], + "opposite_logprob_tokens": [ + -3.2881059646606445 + ], + "logprob_reward": -0.038105905055999756, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038105905055999756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14334139227867126, + 0.12688492238521576, + 0.07294342666864395, + 0.09959817677736282, + 0.03637568652629852, + 0.04713226854801178, + 0.10961742699146271, + 0.037480659782886505, + 0.079990915954113, + 0.24663516879081726 + ], + "credit_map": [ + 0.02632857672870159, + 0.019110623747110367, + 0.016838818788528442, + 0.009044701233506203, + 0.01605275273323059, + 0.023177526891231537, + 0.015394741669297218, + 0.01739363558590412, + 0.019600210711359978, + 0.026719385758042336, + 0.011138931848108768, + 0.01431378722190857, + 0.014245681464672089, + 0.020612353459000587, + 0.010407940484583378, + 0.009846631437540054, + 0.0079783471301198, + 0.0092055918648839, + 0.009421797469258308, + 0.008719035424292088, + 0.006611432880163193, + 0.010882699862122536, + 0.01028954703360796, + 0.009834975004196167, + 0.012782423757016659, + 0.022895710542798042, + 0.007371560204774141, + 0.015057279728353024, + 0.010001718066632748, + 0.016918614506721497, + 0.007274153176695108, + 0.007296718657016754, + 0.004486052319407463, + 0.004666588269174099, + 0.005124683957546949, + 0.0031718146055936813, + 0.004618416074663401, + 0.005156274419277906, + 0.005580477882176638, + 0.0035713810939341784, + 0.007270338013768196, + 0.011149669997394085, + 0.003597358474507928, + 0.0065248627215623856, + 0.003262179670855403, + 0.00899659562855959, + 0.0025928174145519733, + 0.0037384475581347942, + 0.012897965498268604, + 0.022281723096966743, + 0.016784174367785454, + 0.011685396544635296, + 0.009850488975644112, + 0.021435169503092766, + 0.006113091018050909, + 0.008569414727389812, + 0.005454745143651962, + 0.004929684568196535, + 0.0039497665129601955, + 0.003440618747845292, + 0.004210950341075659, + 0.003550079185515642, + 0.007463161833584309, + 0.004481653682887554, + 0.009933249093592167, + 0.011254910379648209, + 0.015837332233786583, + 0.011030246503651142, + 0.009728123433887959, + 0.00695891585201025, + 0.00685019139200449, + 0.008397949859499931, + 0.040299754589796066, + 0.0327046774327755, + 0.0414622537791729, + 0.04110994562506676, + 0.02784785069525242, + 0.023083604872226715, + 0.023733118548989296, + 0.016393961384892464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.038105905055999756 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a steady, slightly elevated wide-angle view that slowly tilts to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a vibrant marina during the afternoon with warm sunlight and a golden hour haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sleek sailboat in the foreground with a red and black sail featuring the white text 'Flying Emirates'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground sailboat depart the wooden pier with taut sails, tilting gently as its white fiberglass hull glides through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast modern cityscape with towering skyscrapers and a suspension bridge visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947155877947807 + ], + "opposite_logprob_tokens": [ + -5.130947113037109 + ], + "logprob_reward": -0.005947155877947807, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947155877947807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16693498194217682, + 0.10154784470796585, + 0.07732629030942917, + 0.12018761783838272, + 0.06631924211978912, + 0.09659481048583984, + 0.0812298059463501, + 0.06793611496686935, + 0.08322688937187195, + 0.13869644701480865 + ], + "credit_map": [ + 0.03010573983192444, + 0.02853163704276085, + 0.017142659053206444, + 0.012894432991743088, + 0.0219850093126297, + 0.026973925530910492, + 0.016193026676774025, + 0.013108563609421253, + 0.015859754756093025, + 0.019759753718972206, + 0.008852760307490826, + 0.00901327095925808, + 0.013794135302305222, + 0.015622214414179325, + 0.008505233563482761, + 0.010140721686184406, + 0.01151187438517809, + 0.011447768658399582, + 0.007915740832686424, + 0.007889263331890106, + 0.012348571792244911, + 0.013486508280038834, + 0.0075781685300171375, + 0.00514839356765151, + 0.013724391348659992, + 0.022119065746665, + 0.010133243165910244, + 0.012556378729641438, + 0.016986418515443802, + 0.01966533623635769, + 0.013429158367216587, + 0.01157363224774599, + 0.009511698968708515, + 0.007415676489472389, + 0.008294016122817993, + 0.0068651242181658745, + 0.01155085302889347, + 0.00953568797558546, + 0.006371784023940563, + 0.0067744008265435696, + 0.012247265316545963, + 0.01921354979276657, + 0.007098238915205002, + 0.012586022727191448, + 0.010605021379888058, + 0.015268211252987385, + 0.006192510947585106, + 0.013383992947638035, + 0.009908445179462433, + 0.013449106365442276, + 0.00893631111830473, + 0.00880667194724083, + 0.008025134913623333, + 0.01128084771335125, + 0.013025890104472637, + 0.007797393016517162, + 0.007827844470739365, + 0.008058040402829647, + 0.007218833547085524, + 0.00837988592684269, + 0.007907789200544357, + 0.005935440771281719, + 0.011883769184350967, + 0.010724504478275776, + 0.011440194211900234, + 0.0127556836232543, + 0.01280840951949358, + 0.010649869218468666, + 0.008853165432810783, + 0.011050663888454437, + 0.006443812511861324, + 0.009225092828273773, + 0.01598099060356617, + 0.01645718328654766, + 0.014706579968333244, + 0.01709195412695408, + 0.017234554514288902, + 0.013464151881635189, + 0.023523151874542236, + 0.020237881690263748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.005947155877947807 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a steady, slightly elevated wide-angle view that slowly tilts to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a vibrant marina during the afternoon with warm sunlight and a golden hour haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sleek sailboat in the foreground with a red and black sail featuring the white text 'Flying Emirates'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground sailboat depart the wooden pier with taut sails, tilting gently as its white fiberglass hull glides through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast modern cityscape with towering skyscrapers and a suspension bridge visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818004623055458 + ], + "opposite_logprob_tokens": [ + -4.018179893493652 + ], + "logprob_reward": -0.01818004623055458, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01818004623055458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0732407420873642, + 0.07268684357404709, + 0.06488967686891556, + 0.11285672336816788, + 0.08370496332645416, + 0.16631513833999634, + 0.07393909990787506, + 0.17369621992111206, + 0.08489447832107544, + 0.09377603977918625 + ], + "credit_map": [ + 0.009890127927064896, + 0.009719869121909142, + 0.0078486492857337, + 0.005439579952508211, + 0.014852060936391354, + 0.010781936347484589, + 0.00894713494926691, + 0.005761387757956982, + 0.007219362072646618, + 0.007890995591878891, + 0.007484434172511101, + 0.009300538338720798, + 0.008304325863718987, + 0.00971402321010828, + 0.017211319878697395, + 0.0055618491023778915, + 0.008404993452131748, + 0.010564027354121208, + 0.0048782131634652615, + 0.007614597678184509, + 0.008461826480925083, + 0.01084565557539463, + 0.005614210851490498, + 0.008506156504154205, + 0.01311388798058033, + 0.025292737409472466, + 0.008796830661594868, + 0.01219224650412798, + 0.013430689461529255, + 0.018941955640912056, + 0.013567508198320866, + 0.007520879153162241, + 0.012277334928512573, + 0.01411457173526287, + 0.011211339384317398, + 0.009246543049812317, + 0.009393978863954544, + 0.011635195463895798, + 0.008101992309093475, + 0.007724013179540634, + 0.020087191835045815, + 0.04456370323896408, + 0.011648228392004967, + 0.022086720913648605, + 0.010410559363663197, + 0.033441733568906784, + 0.007974199019372463, + 0.016102822497487068, + 0.010620281100273132, + 0.012844596989452839, + 0.009796432219445705, + 0.008783001452684402, + 0.008783271536231041, + 0.008194065652787685, + 0.009791799820959568, + 0.005125656258314848, + 0.028236910700798035, + 0.022336354479193687, + 0.021321957930922508, + 0.013276108540594578, + 0.02011077105998993, + 0.020145263522863388, + 0.03024023398756981, + 0.01802862621843815, + 0.013172689825296402, + 0.010893248952925205, + 0.01697677932679653, + 0.015677964314818382, + 0.007957185618579388, + 0.005567953921854496, + 0.007125660311430693, + 0.007523003965616226, + 0.010740340687334538, + 0.010187514126300812, + 0.011730197817087173, + 0.010593395680189133, + 0.010620933026075363, + 0.009018965996801853, + 0.016955168917775154, + 0.013929533772170544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 3 + ], + "selected_reward": -0.01818004623055458 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a steady, slightly elevated wide-angle view that slowly tilts to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a vibrant marina during the afternoon with warm sunlight and a golden hour haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sleek sailboat in the foreground with a red and black sail featuring the white text 'Flying Emirates'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground sailboat depart the wooden pier with taut sails, tilting gently as its white fiberglass hull glides through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast modern cityscape with towering skyscrapers and a suspension bridge visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10025910288095474 + ], + "opposite_logprob_tokens": [ + -2.3502590656280518 + ], + "logprob_reward": -0.10025910288095474, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10025910288095474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045381445437669754, + 0.3431086242198944, + 0.045892808586359024, + 0.05782492831349373, + 0.029192255809903145, + 0.03883637860417366, + 0.0827532708644867, + 0.023579571396112442, + 0.19931361079216003, + 0.13411709666252136 + ], + "credit_map": [ + 0.00579230347648263, + 0.007645429112017155, + 0.005236810073256493, + 0.005208839196711779, + 0.005612511187791824, + 0.00663656834512949, + 0.004600290209054947, + 0.004648697562515736, + 0.05812721699476242, + 0.07164562493562698, + 0.032663606107234955, + 0.029203779995441437, + 0.03885800018906593, + 0.05790954828262329, + 0.024734515696763992, + 0.029966330155730247, + 0.006439106538891792, + 0.006235902663320303, + 0.0031782498117536306, + 0.004238890018314123, + 0.005281070247292519, + 0.0084043825045228, + 0.006691905204206705, + 0.005423298105597496, + 0.006246197037398815, + 0.013136889785528183, + 0.005391336977481842, + 0.007647381629794836, + 0.006207958795130253, + 0.009720352478325367, + 0.004347045440226793, + 0.005127768497914076, + 0.003492315299808979, + 0.005953940562903881, + 0.0033790695015341043, + 0.0037582723889499903, + 0.00264303432777524, + 0.005240354686975479, + 0.0027509434148669243, + 0.0019743251614272594, + 0.007055466528981924, + 0.007214734796434641, + 0.004914469085633755, + 0.005322128999978304, + 0.0022332502994686365, + 0.006751249078661203, + 0.00225301505997777, + 0.0030920663848519325, + 0.01215909793972969, + 0.014079890213906765, + 0.012790230102837086, + 0.013085835613310337, + 0.007255712058395147, + 0.012134182266891003, + 0.005946463439613581, + 0.005301863420754671, + 0.002664877800270915, + 0.0031457494478672743, + 0.0028554510790854692, + 0.004282564856112003, + 0.0027306037954986095, + 0.0027039994020015, + 0.0028419748414307833, + 0.0023543497081846, + 0.029459664598107338, + 0.029490049928426743, + 0.04956803470849991, + 0.030136242508888245, + 0.01533481851220131, + 0.012212339788675308, + 0.013396683149039745, + 0.01971578784286976, + 0.021175984293222427, + 0.018641527742147446, + 0.026210887357592583, + 0.019066989421844482, + 0.01226575393229723, + 0.013619274832308292, + 0.012309188023209572, + 0.010827488265931606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 9 + ], + "selected_reward": -0.10025910288095474 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, slightly elevated wide-angle lens slowly tilts right, capturing a vibrant marina bathed in the warm glow of the afternoon sun. Countless sailboats, hoisting a row of colorful sails, are moored at the wooden pier, their masts forming a forest against the sky. In the foreground, a sleek sailboat with a striking red and black sail begins its journey; the sails are taut with the wind, and across their black sections, the words \"Flying Emirates\" are prominently emblazoned in bold white lettering. As the boat smoothly and gracefully departs the dock, it gently tilts, its white fiberglass hull gliding smoothly through the gentle cuts of the sea, heading out into the vast expanse. As the camera follows its movement, a vast modern cityscape unfolds in the background, featuring towering skyscrapers and a majestic suspension bridge, their shapes softened by the golden hour haze. In the midfield, another sailboat with a similar red and black sail also pulls away from the dock, creating a sense of parallel movement. Other boats with red and white sails dot the background, adding depth and the feeling of a lively regatta as the leading sailboat glides further into the open water.", + "seed": 29068, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_9/index_2806_seed_29068_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.021840532962232827, + "vlm_logprob": -0.021840532962232827, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.021840532962232827, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a steady, slightly elevated wide-angle view that slowly tilts to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a vibrant marina during the afternoon with warm sunlight and a golden hour haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sleek sailboat in the foreground with a red and black sail featuring the white text 'Flying Emirates'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground sailboat depart the wooden pier with taut sails, tilting gently as its white fiberglass hull glides through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast modern cityscape with towering skyscrapers and a suspension bridge visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.034269966185092926 + ], + "opposite_logprob_tokens": [ + -4.5342698097229 + ], + "logprob_reward": -0.034269966185092926, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.034269966185092926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11634323000907898, + 0.07506293058395386, + 0.039314378052949905, + 0.038805264979600906, + 0.0422990657389164, + 0.40825679898262024, + 0.07127555459737778, + 0.034541476517915726, + 0.10350362211465836, + 0.07059773802757263 + ], + "credit_map": [ + 0.025960324332118034, + 0.02011062577366829, + 0.010999103076756, + 0.011272748932242393, + 0.010664467699825764, + 0.019834719598293304, + 0.009089314378798008, + 0.008411923423409462, + 0.008967138826847076, + 0.015698615461587906, + 0.009286877699196339, + 0.007926443591713905, + 0.00971131306141615, + 0.010986924171447754, + 0.006427730433642864, + 0.006057888735085726, + 0.005780315492302179, + 0.007749558426439762, + 0.004506331868469715, + 0.005006261169910431, + 0.004788847640156746, + 0.005549862515181303, + 0.0027044694870710373, + 0.0032287288922816515, + 0.005969904828816652, + 0.005721150431782007, + 0.006937719881534576, + 0.004521583206951618, + 0.0034862342290580273, + 0.0054084365256130695, + 0.003730474505573511, + 0.003029764164239168, + 0.005574013106524944, + 0.005601243581622839, + 0.0035909824073314667, + 0.0042742276564240456, + 0.005958157125860453, + 0.005950707942247391, + 0.007722095120698214, + 0.0036276362370699644, + 0.14914119243621826, + 0.03317325934767723, + 0.012377869337797165, + 0.007904007099568844, + 0.018559003248810768, + 0.1606970578432083, + 0.01538823451846838, + 0.011016150005161762, + 0.01815756969153881, + 0.009350563399493694, + 0.007510846015065908, + 0.005955932661890984, + 0.007287742104381323, + 0.012999203987419605, + 0.005439972039312124, + 0.004573721904307604, + 0.004942572675645351, + 0.003922266885638237, + 0.0034348939079791307, + 0.004190672188997269, + 0.0037025141064077616, + 0.004828264936804771, + 0.006227388978004456, + 0.0032929035369306803, + 0.019154174253344536, + 0.011115007102489471, + 0.022953622043132782, + 0.007518865168094635, + 0.009266527369618416, + 0.015428675338625908, + 0.00865321233868599, + 0.00941354501992464, + 0.006501747295260429, + 0.00935455597937107, + 0.007387616205960512, + 0.007028724066913128, + 0.015440462157130241, + 0.011629763059318066, + 0.007326182443648577, + 0.0059286910109221935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.034269966185092926 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a steady, slightly elevated wide-angle view that slowly tilts to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a vibrant marina during the afternoon with warm sunlight and a golden hour haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sleek sailboat in the foreground with a red and black sail featuring the white text 'Flying Emirates'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground sailboat depart the wooden pier with taut sails, tilting gently as its white fiberglass hull glides through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast modern cityscape with towering skyscrapers and a suspension bridge visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008647731505334377 + ], + "opposite_logprob_tokens": [ + -4.758647918701172 + ], + "logprob_reward": -0.008647731505334377, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008647731505334377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20446324348449707, + 0.11624139547348022, + 0.054939787834882736, + 0.037447553128004074, + 0.034715358167886734, + 0.20301783084869385, + 0.10450876504182816, + 0.09112492948770523, + 0.13413584232330322, + 0.01940523460507393 + ], + "credit_map": [ + 0.0469004288315773, + 0.038173362612724304, + 0.02142823114991188, + 0.02153879776597023, + 0.01879793033003807, + 0.027426468208432198, + 0.014425262808799744, + 0.01577276922762394, + 0.016743876039981842, + 0.02954457700252533, + 0.011965487152338028, + 0.010431562550365925, + 0.009097201749682426, + 0.025689734145998955, + 0.006419407203793526, + 0.006349551025778055, + 0.007678420748561621, + 0.011657213792204857, + 0.006627561990171671, + 0.005051932763308287, + 0.006336972117424011, + 0.009789526462554932, + 0.004754272755235434, + 0.003043888136744499, + 0.005176766309887171, + 0.007526583969593048, + 0.006086163222789764, + 0.003814989235252142, + 0.0032047023996710777, + 0.005176890641450882, + 0.003059233305975795, + 0.003402221016585827, + 0.0033028160687536, + 0.005517212674021721, + 0.0036532566882669926, + 0.005425282288342714, + 0.0027138912118971348, + 0.0051009817980229855, + 0.0038189480546861887, + 0.005182969383895397, + 0.07406684011220932, + 0.017590774223208427, + 0.007045440375804901, + 0.006118095945566893, + 0.00636640889570117, + 0.07994796335697174, + 0.006356362719088793, + 0.005525949411094189, + 0.025563577190041542, + 0.012505006045103073, + 0.011233487166464329, + 0.009756159968674183, + 0.008817515335977077, + 0.01749531738460064, + 0.009052537381649017, + 0.010085166431963444, + 0.01580161042511463, + 0.008699826896190643, + 0.015419984236359596, + 0.008895029313862324, + 0.007363899610936642, + 0.012457435950636864, + 0.01048034057021141, + 0.012006799690425396, + 0.023393699899315834, + 0.01307124923914671, + 0.02979947067797184, + 0.01004712749272585, + 0.012192168273031712, + 0.019535185769200325, + 0.012936479412019253, + 0.013160466216504574, + 0.002588847652077675, + 0.0026854239404201508, + 0.00217781332321465, + 0.0023904505651444197, + 0.0029166287276893854, + 0.002604458015412092, + 0.002242075279355049, + 0.0017995372181758285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.008647731505334377 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a steady, slightly elevated wide-angle view that slowly tilts to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a vibrant marina during the afternoon with warm sunlight and a golden hour haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sleek sailboat in the foreground with a red and black sail featuring the white text 'Flying Emirates'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground sailboat depart the wooden pier with taut sails, tilting gently as its white fiberglass hull glides through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast modern cityscape with towering skyscrapers and a suspension bridge visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463874451816082 + ], + "opposite_logprob_tokens": [ + -5.379638671875 + ], + "logprob_reward": -0.00463874451816082, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463874451816082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1405777633190155, + 0.050969984382390976, + 0.09118174761533737, + 0.10764898359775543, + 0.059778328984975815, + 0.31484854221343994, + 0.05944880470633507, + 0.08964423090219498, + 0.030483057722449303, + 0.05541851744055748 + ], + "credit_map": [ + 0.02274635247886181, + 0.022133037447929382, + 0.014531702734529972, + 0.02033180557191372, + 0.012513563968241215, + 0.02147139236330986, + 0.011980402283370495, + 0.014869515784084797, + 0.006927658803761005, + 0.009506401605904102, + 0.0055203367955982685, + 0.005719516426324844, + 0.006406273227185011, + 0.00780337443575263, + 0.004733231384307146, + 0.004353185649961233, + 0.0107087017968297, + 0.018916267901659012, + 0.010781822726130486, + 0.008972728624939919, + 0.01144004799425602, + 0.01547615509480238, + 0.008611631579697132, + 0.006274391897022724, + 0.010445566847920418, + 0.021525796502828598, + 0.00912456400692463, + 0.0070154001004993916, + 0.01708598993718624, + 0.02488991990685463, + 0.010133126750588417, + 0.007428615354001522, + 0.005956107284873724, + 0.0097004733979702, + 0.007058765739202499, + 0.011128842830657959, + 0.0048203254118561745, + 0.008162791840732098, + 0.005259340163320303, + 0.007691682782024145, + 0.12074651569128036, + 0.026998845860362053, + 0.008091235533356667, + 0.004922474268823862, + 0.008104225620627403, + 0.1303611844778061, + 0.007793226279318333, + 0.007830817252397537, + 0.010192287154495716, + 0.006947482004761696, + 0.004442952107638121, + 0.00869804248213768, + 0.005244055762887001, + 0.012435712851583958, + 0.006332163233309984, + 0.0051561035215854645, + 0.009828670881688595, + 0.007746865041553974, + 0.00994905550032854, + 0.009640313684940338, + 0.010286266915500164, + 0.010975227691233158, + 0.020651033148169518, + 0.010566793382167816, + 0.004872835241258144, + 0.0027864319272339344, + 0.004982118960469961, + 0.0026649117935448885, + 0.0029523305129259825, + 0.00395186897367239, + 0.004400426056236029, + 0.0038721368182450533, + 0.006452416535466909, + 0.008288557641208172, + 0.008129892870783806, + 0.006023885682225227, + 0.006366085261106491, + 0.007593405432999134, + 0.007165989838540554, + 0.005398288369178772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.00463874451816082 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a steady, slightly elevated wide-angle view that slowly tilts to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a vibrant marina during the afternoon with warm sunlight and a golden hour haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sleek sailboat in the foreground with a red and black sail featuring the white text 'Flying Emirates'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground sailboat depart the wooden pier with taut sails, tilting gently as its white fiberglass hull glides through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast modern cityscape with towering skyscrapers and a suspension bridge visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006730742286890745 + ], + "opposite_logprob_tokens": [ + -5.006730556488037 + ], + "logprob_reward": -0.006730742286890745, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006730742286890745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11627525091171265, + 0.05608339607715607, + 0.0657854676246643, + 0.03164484724402428, + 0.035568468272686005, + 0.42316335439682007, + 0.07646718621253967, + 0.051163267344236374, + 0.10578383505344391, + 0.038064923137426376 + ], + "credit_map": [ + 0.014430449344217777, + 0.01734706200659275, + 0.01046453882008791, + 0.010815093293786049, + 0.022629564628005028, + 0.017386894673109055, + 0.012720167636871338, + 0.01048148050904274, + 0.005381693597882986, + 0.00837530754506588, + 0.005373763851821423, + 0.005154013633728027, + 0.011232884600758553, + 0.009215417318046093, + 0.006717341486364603, + 0.00463297450914979, + 0.008736878633499146, + 0.010926933959126472, + 0.0116904117166996, + 0.009001905098557472, + 0.005435334052890539, + 0.009002730250358582, + 0.005691328085958958, + 0.005299943964928389, + 0.0037140613421797752, + 0.004040675703436136, + 0.0033400177489966154, + 0.0025289368350058794, + 0.007059353403747082, + 0.004099807236343622, + 0.004095909651368856, + 0.0027660848572850227, + 0.0044315592385828495, + 0.004497491288930178, + 0.005892637651413679, + 0.006126143038272858, + 0.005583459045737982, + 0.0037235512863844633, + 0.003186594694852829, + 0.002127031795680523, + 0.1484265923500061, + 0.038931820541620255, + 0.022275418043136597, + 0.011356960982084274, + 0.011878091841936111, + 0.16240283846855164, + 0.012481593526899815, + 0.015410046093165874, + 0.010731601156294346, + 0.008999626152217388, + 0.005034090485423803, + 0.014184233732521534, + 0.006751091219484806, + 0.014409419149160385, + 0.008743062615394592, + 0.007614057511091232, + 0.005645150318741798, + 0.006774066481739283, + 0.006686067208647728, + 0.004452149383723736, + 0.0068864659406244755, + 0.006268453784286976, + 0.008786479011178017, + 0.005664433818310499, + 0.014317044988274574, + 0.015115908347070217, + 0.020179616287350655, + 0.00863892026245594, + 0.010743158869445324, + 0.014788392931222916, + 0.012446320615708828, + 0.009554474614560604, + 0.004519364330917597, + 0.003188484348356724, + 0.005238511599600315, + 0.004153612535446882, + 0.007480768486857414, + 0.005484257359057665, + 0.004518221132457256, + 0.003481700085103512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.006730742286890745 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a steady, slightly elevated wide-angle view that slowly tilts to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a vibrant marina during the afternoon with warm sunlight and a golden hour haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sleek sailboat in the foreground with a red and black sail featuring the white text 'Flying Emirates'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground sailboat depart the wooden pier with taut sails, tilting gently as its white fiberglass hull glides through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast modern cityscape with towering skyscrapers and a suspension bridge visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05491548031568527 + ], + "opposite_logprob_tokens": [ + -2.929915428161621 + ], + "logprob_reward": -0.05491548031568527, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05491548031568527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4048769772052765, + 0.16490869224071503, + 0.03187202662229538, + 0.024130627512931824, + 0.050887320190668106, + 0.1753578931093216, + 0.05479612201452255, + 0.03460255637764931, + 0.026359815150499344, + 0.03220795840024948 + ], + "credit_map": [ + 0.09395700693130493, + 0.07690620422363281, + 0.04509640112519264, + 0.04202636703848839, + 0.0219940934330225, + 0.0676298588514328, + 0.01935582607984543, + 0.03791123256087303, + 0.01947307400405407, + 0.041322559118270874, + 0.01821824721992016, + 0.019280998036265373, + 0.01103919930756092, + 0.038188353180885315, + 0.009080203250050545, + 0.008306055329740047, + 0.005407598335295916, + 0.006367353722453117, + 0.004139972385019064, + 0.002965800929814577, + 0.0028971992433071136, + 0.004190273582935333, + 0.0028323486912995577, + 0.0030714801978319883, + 0.002801933791488409, + 0.003924322780221701, + 0.0034035213757306337, + 0.0038284591864794493, + 0.002126857405528426, + 0.003263824386522174, + 0.002475237473845482, + 0.0023064680863171816, + 0.007758546620607376, + 0.011920487508177757, + 0.005555868614464998, + 0.004775793757289648, + 0.004515559412539005, + 0.008753621950745583, + 0.0034350762143731117, + 0.004172366578131914, + 0.06046370789408684, + 0.01801246590912342, + 0.009259718470275402, + 0.006230800878256559, + 0.005994190461933613, + 0.06404433399438858, + 0.005887961946427822, + 0.005464725662022829, + 0.013553110882639885, + 0.00543673150241375, + 0.005941470619291067, + 0.006439847405999899, + 0.004187353421002626, + 0.009776105172932148, + 0.004927073139697313, + 0.004534431267529726, + 0.004655130673199892, + 0.003989486023783684, + 0.005386386066675186, + 0.0033248558174818754, + 0.003099477384239435, + 0.004492860287427902, + 0.0046210866421461105, + 0.005033272784203291, + 0.0053848098032176495, + 0.0031495369039475918, + 0.004734193440526724, + 0.0018926900811493397, + 0.0025696936063468456, + 0.0041875517927110195, + 0.002078684978187084, + 0.00236265710555017, + 0.0031482663471251726, + 0.003246407024562359, + 0.003437256906181574, + 0.002466874895617366, + 0.007777808699756861, + 0.0054763732478022575, + 0.003555923467501998, + 0.003099048277363181 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.05491548031568527 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, slightly elevated wide-angle lens slowly tilts right, capturing a vibrant marina bathed in the warm glow of the afternoon sun. Countless sailboats, hoisting a row of colorful sails, are moored at the wooden pier, their masts forming a forest against the sky. In the foreground, a sleek sailboat with a striking red and black sail begins its journey; the sails are taut with the wind, and across their black sections, the words \"Flying Emirates\" are prominently emblazoned in bold white lettering. As the boat smoothly and gracefully departs the dock, it gently tilts, its white fiberglass hull gliding smoothly through the gentle cuts of the sea, heading out into the vast expanse. As the camera follows its movement, a vast modern cityscape unfolds in the background, featuring towering skyscrapers and a majestic suspension bridge, their shapes softened by the golden hour haze. In the midfield, another sailboat with a similar red and black sail also pulls away from the dock, creating a sense of parallel movement. Other boats with red and white sails dot the background, adding depth and the feeling of a lively regatta as the leading sailboat glides further into the open water.", + "seed": 29069, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_9/index_2806_seed_29069_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.021013982873409987, + "vlm_logprob": -0.021013982873409987, + "vlm_yes_no_margin": 4.224999952316284, + "vlm_token_credit": -0.021013982873409987, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a steady, slightly elevated wide-angle view that slowly tilts to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a vibrant marina during the afternoon with warm sunlight and a golden hour haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sleek sailboat in the foreground with a red and black sail featuring the white text 'Flying Emirates'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground sailboat depart the wooden pier with taut sails, tilting gently as its white fiberglass hull glides through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast modern cityscape with towering skyscrapers and a suspension bridge visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014571674168109894 + ], + "opposite_logprob_tokens": [ + -5.139571666717529 + ], + "logprob_reward": -0.014571674168109894, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.014571674168109894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4505782425403595, + 0.07730947434902191, + 0.0818709135055542, + 0.08585795015096664, + 0.050836000591516495, + 0.03105786256492138, + 0.04144565388560295, + 0.057828307151794434, + 0.06318219751119614, + 0.06003342196345329 + ], + "credit_map": [ + 0.09256511926651001, + 0.055866651237010956, + 0.1560528725385666, + 0.058274347335100174, + 0.013462145812809467, + 0.02191060595214367, + 0.024032538756728172, + 0.02841399610042572, + 0.011409591883420944, + 0.008659773506224155, + 0.02028083987534046, + 0.010102261789143085, + 0.008356724865734577, + 0.005908136256039143, + 0.006410420406609774, + 0.006181733217090368, + 0.007802661973983049, + 0.007843319326639175, + 0.014110851101577282, + 0.01973673887550831, + 0.007173565682023764, + 0.007609239779412746, + 0.012090985663235188, + 0.005503551103174686, + 0.013653124682605267, + 0.010943720117211342, + 0.00980342272669077, + 0.009036147966980934, + 0.015700586140155792, + 0.012548242695629597, + 0.008068582974374294, + 0.006104124244302511, + 0.007252723444253206, + 0.006222786381840706, + 0.007445774506777525, + 0.007141052279621363, + 0.0076346630230546, + 0.00510812783613801, + 0.0067198132164776325, + 0.003311059670522809, + 0.006216117646545172, + 0.004007525276392698, + 0.004724700935184956, + 0.0037922882474958897, + 0.003470058087259531, + 0.0026719633024185896, + 0.0030612023547291756, + 0.003114007180556655, + 0.00433460995554924, + 0.004835933912545443, + 0.007240341044962406, + 0.006631956901401281, + 0.006303494796156883, + 0.004206112585961819, + 0.0056643798016011715, + 0.002228825818747282, + 0.008328182622790337, + 0.010019980370998383, + 0.005643686279654503, + 0.006350994575768709, + 0.00977024994790554, + 0.004785962402820587, + 0.006364049389958382, + 0.006565194111317396, + 0.011646844446659088, + 0.009982571005821228, + 0.008004152216017246, + 0.009077548049390316, + 0.007876857183873653, + 0.005825407803058624, + 0.006764108315110207, + 0.004004706162959337, + 0.007184115704149008, + 0.009349753148853779, + 0.007412302307784557, + 0.005844193510711193, + 0.010710498318076134, + 0.006753814872354269, + 0.008662889711558819, + 0.004115858115255833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.014571674168109894 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a steady, slightly elevated wide-angle view that slowly tilts to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a vibrant marina during the afternoon with warm sunlight and a golden hour haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sleek sailboat in the foreground with a red and black sail featuring the white text 'Flying Emirates'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground sailboat depart the wooden pier with taut sails, tilting gently as its white fiberglass hull glides through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast modern cityscape with towering skyscrapers and a suspension bridge visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818917691707611 + ], + "opposite_logprob_tokens": [ + -4.018188953399658 + ], + "logprob_reward": -0.01818917691707611, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01818917691707611, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09679500758647919, + 0.027899743989109993, + 0.03731175884604454, + 0.1281348466873169, + 0.03598908334970474, + 0.14205056428909302, + 0.09060888737440109, + 0.1290806531906128, + 0.10620944201946259, + 0.2059200406074524 + ], + "credit_map": [ + 0.0189554113894701, + 0.01144717913120985, + 0.029504021629691124, + 0.011251959018409252, + 0.004363827873021364, + 0.006846265867352486, + 0.006564069539308548, + 0.007862276397645473, + 0.004142486024647951, + 0.005778016056865454, + 0.003749784315004945, + 0.0030044112354516983, + 0.0035785804502665997, + 0.0024285719264298677, + 0.002589993644505739, + 0.0026279031299054623, + 0.005903296638280153, + 0.005032834131270647, + 0.005359622649848461, + 0.007166956085711718, + 0.004484568722546101, + 0.00279624224640429, + 0.004080474376678467, + 0.002487760502845049, + 0.015071352012455463, + 0.02056996524333954, + 0.01641819439828396, + 0.014763908460736275, + 0.02990119159221649, + 0.009812049567699432, + 0.01323242112994194, + 0.00836576521396637, + 0.003884610254317522, + 0.005442698020488024, + 0.004638405051082373, + 0.005083293188363314, + 0.005767008289694786, + 0.0028812771197408438, + 0.005981267895549536, + 0.0023105256259441376, + 0.02934839390218258, + 0.023035535588860512, + 0.018443256616592407, + 0.014838766306638718, + 0.027393385767936707, + 0.010021055117249489, + 0.01263797003775835, + 0.00633220374584198, + 0.0114859314635396, + 0.011581037193536758, + 0.010985897853970528, + 0.009897499345242977, + 0.020211510360240936, + 0.013497775420546532, + 0.007210904732346535, + 0.005738326348364353, + 0.015619602054357529, + 0.021175697445869446, + 0.014757398515939713, + 0.013575151562690735, + 0.028225883841514587, + 0.013409505598247051, + 0.014082375913858414, + 0.008235043846070766, + 0.014157646335661411, + 0.016970288008451462, + 0.016067612916231155, + 0.011093039065599442, + 0.01838032528758049, + 0.008791700005531311, + 0.01242461521178484, + 0.008324217982590199, + 0.037627119570970535, + 0.03513962775468826, + 0.017537521198391914, + 0.016952350735664368, + 0.018681243062019348, + 0.04891078919172287, + 0.020456695929169655, + 0.010614684782922268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 7 + ], + "selected_reward": -0.01818917691707611 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a steady, slightly elevated wide-angle view that slowly tilts to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a vibrant marina during the afternoon with warm sunlight and a golden hour haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sleek sailboat in the foreground with a red and black sail featuring the white text 'Flying Emirates'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground sailboat depart the wooden pier with taut sails, tilting gently as its white fiberglass hull glides through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast modern cityscape with towering skyscrapers and a suspension bridge visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016065435484051704 + ], + "opposite_logprob_tokens": [ + -4.14106559753418 + ], + "logprob_reward": -0.016065435484051704, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016065435484051704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0934881940484047, + 0.08174949139356613, + 0.11778075993061066, + 0.27802330255508423, + 0.10568656027317047, + 0.07380548864603043, + 0.057065241038799286, + 0.0723610445857048, + 0.07934078574180603, + 0.040699124336242676 + ], + "credit_map": [ + 0.014560850337147713, + 0.017216389998793602, + 0.016038576140999794, + 0.012713206000626087, + 0.008135441690683365, + 0.008521916344761848, + 0.007718897424638271, + 0.008582918904721737, + 0.013306102715432644, + 0.02010069414973259, + 0.012363418936729431, + 0.010929370298981667, + 0.007841401733458042, + 0.004928880371153355, + 0.006665355525910854, + 0.0056142657995224, + 0.017455488443374634, + 0.017118407413363457, + 0.015348401851952076, + 0.018236231058835983, + 0.013408067636191845, + 0.010491747409105301, + 0.016040293499827385, + 0.009682122617959976, + 0.0636684000492096, + 0.040324866771698, + 0.04765482619404793, + 0.022864652797579765, + 0.05088426172733307, + 0.017878444865345955, + 0.0214843712747097, + 0.0132634611800313, + 0.013182749971747398, + 0.01612352766096592, + 0.012714169919490814, + 0.013248750939965248, + 0.01732722297310829, + 0.007233283016830683, + 0.014511678367853165, + 0.011345178820192814, + 0.010120831429958344, + 0.008555598556995392, + 0.008735168725252151, + 0.007537951692938805, + 0.01062273420393467, + 0.007718388456851244, + 0.009183612652122974, + 0.011331205256283283, + 0.007263524923473597, + 0.006240865681320429, + 0.008756276220083237, + 0.005306881386786699, + 0.008022172376513481, + 0.0061567397788167, + 0.006060910411179066, + 0.009257872588932514, + 0.011747161857783794, + 0.013001225888729095, + 0.007255631498992443, + 0.005719940643757582, + 0.018144359812140465, + 0.004778041038662195, + 0.005942619871348143, + 0.00577206676825881, + 0.011338291689753532, + 0.010091930627822876, + 0.009258950129151344, + 0.00772528862580657, + 0.020483270287513733, + 0.006999504286795855, + 0.00827396847307682, + 0.005169581156224012, + 0.00641388026997447, + 0.00698951305821538, + 0.003998543601483107, + 0.00496778404340148, + 0.004592087585479021, + 0.007321583107113838, + 0.004441170021891594, + 0.001974566839635372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 4 + ], + "selected_reward": -0.016065435484051704 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a steady, slightly elevated wide-angle view that slowly tilts to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a vibrant marina during the afternoon with warm sunlight and a golden hour haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sleek sailboat in the foreground with a red and black sail featuring the white text 'Flying Emirates'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground sailboat depart the wooden pier with taut sails, tilting gently as its white fiberglass hull glides through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast modern cityscape with towering skyscrapers and a suspension bridge visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007622915785759687 + ], + "opposite_logprob_tokens": [ + -4.882622718811035 + ], + "logprob_reward": -0.007622915785759687, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007622915785759687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23371390998363495, + 0.05748433619737625, + 0.0799357146024704, + 0.12654872238636017, + 0.05428061634302139, + 0.15571530163288116, + 0.1477983146905899, + 0.04315769299864769, + 0.07049496471881866, + 0.030870432034134865 + ], + "credit_map": [ + 0.03922885283827782, + 0.03229258209466934, + 0.06687333434820175, + 0.03244359791278839, + 0.014927888289093971, + 0.016652371734380722, + 0.014053352177143097, + 0.017241908237338066, + 0.007332619745284319, + 0.00891492422670126, + 0.0059076775796711445, + 0.004670437891036272, + 0.009016286581754684, + 0.009057808667421341, + 0.005640081595629454, + 0.00694449245929718, + 0.010385384783148766, + 0.010287304408848286, + 0.008768342435359955, + 0.00866100937128067, + 0.011315847747027874, + 0.014905894175171852, + 0.00749071454629302, + 0.008121212013065815, + 0.02854030765593052, + 0.019287245348095894, + 0.013946611434221268, + 0.01143838744610548, + 0.023351047188043594, + 0.011166011914610863, + 0.011153429746627808, + 0.0076656644232571125, + 0.006884896662086248, + 0.00882245134562254, + 0.005318948533385992, + 0.006771625019609928, + 0.010296162217855453, + 0.00489678094163537, + 0.007085155230015516, + 0.004204594064503908, + 0.0349581316113472, + 0.023715157061815262, + 0.020527830347418785, + 0.013138161040842533, + 0.027018340304493904, + 0.013330785557627678, + 0.013485210947692394, + 0.009541677311062813, + 0.018771300092339516, + 0.019977644085884094, + 0.013432374224066734, + 0.013210276141762733, + 0.018669798970222473, + 0.021908339112997055, + 0.02697775512933731, + 0.014850812964141369, + 0.006617581006139517, + 0.007730670273303986, + 0.0056199124082922935, + 0.004226962570101023, + 0.007956387475132942, + 0.0030080254655331373, + 0.0044078528881073, + 0.0035902955569326878, + 0.008567762561142445, + 0.009800158441066742, + 0.006156284827739, + 0.005785680841654539, + 0.020328525453805923, + 0.006047467701137066, + 0.008264421485364437, + 0.005544650834053755, + 0.0035261402372270823, + 0.004131290130317211, + 0.003933989908546209, + 0.004398311022669077, + 0.003341002156957984, + 0.003519567660987377, + 0.004809678997844458, + 0.0032104486599564552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.007622915785759687 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a steady, slightly elevated wide-angle view that slowly tilts to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a vibrant marina during the afternoon with warm sunlight and a golden hour haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sleek sailboat in the foreground with a red and black sail featuring the white text 'Flying Emirates'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground sailboat depart the wooden pier with taut sails, tilting gently as its white fiberglass hull glides through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast modern cityscape with towering skyscrapers and a suspension bridge visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048620712012052536 + ], + "opposite_logprob_tokens": [ + -3.0486207008361816 + ], + "logprob_reward": -0.048620712012052536, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048620712012052536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06417899578809738, + 0.029658827930688858, + 0.0450739860534668, + 0.09363565593957901, + 0.04380333423614502, + 0.04363887757062912, + 0.14869874715805054, + 0.22117364406585693, + 0.20536448061466217, + 0.10477342456579208 + ], + "credit_map": [ + 0.011879898607730865, + 0.0092529421672225, + 0.018145078793168068, + 0.007749465294182301, + 0.003156276186928153, + 0.00481816241517663, + 0.004103970713913441, + 0.005073202773928642, + 0.004860908258706331, + 0.004825243726372719, + 0.006808553356677294, + 0.003626772668212652, + 0.0031111459247767925, + 0.002172670094296336, + 0.002082284539937973, + 0.0021712512243539095, + 0.007159981410950422, + 0.005813729949295521, + 0.006842122878879309, + 0.007341114804148674, + 0.005537381861358881, + 0.004636112134903669, + 0.004215477500110865, + 0.003528067609295249, + 0.016243450343608856, + 0.014475362375378609, + 0.014862060546875, + 0.010457062162458897, + 0.020724456757307053, + 0.005215224344283342, + 0.007181358989328146, + 0.00447667995467782, + 0.004465760663151741, + 0.006739199627190828, + 0.0051117874681949615, + 0.006859125569462776, + 0.009640966542065144, + 0.0025726391468197107, + 0.0059763225726783276, + 0.002437530318275094, + 0.008805367164313793, + 0.006122804246842861, + 0.005089868325740099, + 0.0050756661221385, + 0.007096033543348312, + 0.0037396640982478857, + 0.005315498448908329, + 0.0023939746897667646, + 0.019058747217059135, + 0.01932135410606861, + 0.025919491425156593, + 0.0168612077832222, + 0.03104141727089882, + 0.01535702869296074, + 0.013544254936277866, + 0.007595260627567768, + 0.029940184205770493, + 0.03647822514176369, + 0.023069942370057106, + 0.018283870071172714, + 0.052383508533239365, + 0.02516142837703228, + 0.01906483620405197, + 0.01679164730012417, + 0.02452710270881653, + 0.026884647086262703, + 0.021312588825821877, + 0.02219761349260807, + 0.05675438418984413, + 0.01969720982015133, + 0.021284623071551323, + 0.012706303037703037, + 0.011863235384225845, + 0.021617278456687927, + 0.011148808524012566, + 0.008505906909704208, + 0.01821337267756462, + 0.015752604231238365, + 0.010226387530565262, + 0.007445835042744875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.048620712012052536 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_0/index_tensor([9805, 9805])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_0/index_tensor([9805, 9805])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..82457b7d4ed4e9faf37933d1e1aee4d089870158 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_0/index_tensor([9805, 9805])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "fish-eye lens, at an extremely low angle, slowly pans to the right, capturing a distorted view of the interior of an empty circus tent. A single, pale spotlight shines vertically from above, creating a harsh circle of light on the center stage, while the rest of the tent is swallowed by a bottomless darkness. Within the circle of light, a clown with an exaggerated smile paces. His face is covered in thick, slightly cracked white makeup. A large, fixed red smile is painted morbidly from ear to ear. Deep-set black eye sockets contrast eerily with this false cheerfulness, and a shaggy, bright orange wig appears lifeless. He wears a loose, faded clown costume with a mixed pattern (half vertical stripes, half polka dots). The floppy ruffled collar is slightly yellowed. His well-worn, heavy-toed shoes drag heavily across the sawdust-strewn floor, in an uneasy, repetitive wandering. His shoulders slumped slightly, his hands alternately clenched into fists and shoved feebly into his pockets, his posture completely at odds with the perpetual smile on his face. Through the lens of a fisheye lens, the rows of empty wooden benches and the massive tent poles bent inward, like a shrinking, claustrophobic cage. With his anxious steps, his massive, elongated, and distorted shadow danced grotesquely across the ground, casting a cold, eerie, oppressive atmosphere tinged with intense anxiety.", + "seed": 298, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_0/index_9805_seed_298_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.4240390910767019, + "vlm_logprob": -0.4240390910767019, + "vlm_yes_no_margin": 1.875, + "vlm_token_credit": -0.4240390910767019, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with a fish-eye lens effect causing distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a clown wearing a loose, faded costume with mixed vertical stripes and polka dots visible in the center of a spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the clown exhibit an exaggerated smile, thick cracked white makeup, deep-set black eye sockets, and a shaggy bright orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an empty circus tent with rows of empty wooden benches, massive bent tent poles, and a sawdust-strewn floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2357579469680786 + ], + "opposite_logprob_tokens": [ + -0.3607579171657562 + ], + "logprob_reward": -1.2357579469680786, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2357579469680786, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2532300055027008, + 0.10669414699077606, + 0.0716138407588005, + 0.06826893240213394, + 0.04286862164735794, + 0.06080840155482292, + 0.06572918593883514, + 0.14837709069252014, + 0.13437826931476593, + 0.0480315200984478 + ], + "credit_map": [ + 0.03223291411995888, + 0.06600018590688705, + 0.038924962282180786, + 0.03198354318737984, + 0.015946965664625168, + 0.03006826713681221, + 0.014445685781538486, + 0.02362748794257641, + 0.01511430088430643, + 0.01979619637131691, + 0.016523361206054688, + 0.013240637257695198, + 0.009557316079735756, + 0.013767072930932045, + 0.010055230930447578, + 0.008640031330287457, + 0.007155313156545162, + 0.011294246651232243, + 0.018305720761418343, + 0.008788329549133778, + 0.0034974070731550455, + 0.013594364747405052, + 0.0037252313923090696, + 0.005253233946859837, + 0.008412906900048256, + 0.010671396739780903, + 0.016499193385243416, + 0.008535138331353664, + 0.003609014442190528, + 0.012418031692504883, + 0.00262382416985929, + 0.0054994304664433, + 0.006965517532080412, + 0.0034326333552598953, + 0.007165196351706982, + 0.005231416318565607, + 0.003553197719156742, + 0.0070994095876812935, + 0.0034460306633263826, + 0.005975224077701569, + 0.007832220755517483, + 0.007672635372728109, + 0.009652175940573215, + 0.0051153237000107765, + 0.007899654097855091, + 0.009836877696216106, + 0.006333702709525824, + 0.006465810816735029, + 0.015906600281596184, + 0.003732821671292186, + 0.01445370726287365, + 0.003329376457259059, + 0.0055312542244791985, + 0.012715877965092659, + 0.003836921416223049, + 0.006222628057003021, + 0.02782427705824375, + 0.01272538211196661, + 0.020396405830979347, + 0.009985042735934258, + 0.019142385572195053, + 0.02691088430583477, + 0.012985358946025372, + 0.01840735785663128, + 0.017149517312645912, + 0.011539308354258537, + 0.01593843474984169, + 0.008145543746650219, + 0.018200835213065147, + 0.021820493042469025, + 0.020403092727065086, + 0.02118103764951229, + 0.007807544432580471, + 0.0033844145946204662, + 0.0068908329121768475, + 0.002890793140977621, + 0.006285964045673609, + 0.0063308472745120525, + 0.008024374954402447, + 0.006416748743504286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -1.2357579469680786 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with a fish-eye lens effect causing distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a clown wearing a loose, faded costume with mixed vertical stripes and polka dots visible in the center of a spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the clown exhibit an exaggerated smile, thick cracked white makeup, deep-set black eye sockets, and a shaggy bright orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an empty circus tent with rows of empty wooden benches, massive bent tent poles, and a sawdust-strewn floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289668798446655 + ], + "opposite_logprob_tokens": [ + -1.0539668798446655 + ], + "logprob_reward": -0.4289668798446655, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289668798446655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10768561065196991, + 0.10011466592550278, + 0.09278318285942078, + 0.04551480710506439, + 0.042560942471027374, + 0.07751458138227463, + 0.05624055862426758, + 0.09015235304832458, + 0.20024721324443817, + 0.187186136841774 + ], + "credit_map": [ + 0.01226846594363451, + 0.025489255785942078, + 0.016098136082291603, + 0.01350910309702158, + 0.007293982896953821, + 0.016016539186239243, + 0.007414618507027626, + 0.009595498442649841, + 0.0156032033264637, + 0.016934413462877274, + 0.018205299973487854, + 0.009883621707558632, + 0.009948210790753365, + 0.012756112031638622, + 0.009595337323844433, + 0.00718845846131444, + 0.0075067561119794846, + 0.014133957214653492, + 0.02997216209769249, + 0.008056250400841236, + 0.0032749369274824858, + 0.02111012116074562, + 0.0028438875451684, + 0.005885101854801178, + 0.005050828214734793, + 0.008153611794114113, + 0.009741827845573425, + 0.004502122290432453, + 0.003679142566397786, + 0.006924767978489399, + 0.0023805913515388966, + 0.00508191017434001, + 0.006898873019963503, + 0.0034914209973067045, + 0.007899649441242218, + 0.004445526283234358, + 0.0035139559768140316, + 0.00803271122276783, + 0.0029897955246269703, + 0.005289009772241116, + 0.010628382675349712, + 0.008909245021641254, + 0.011973756365478039, + 0.007137712556868792, + 0.010630353353917599, + 0.01143003348261118, + 0.006040363106876612, + 0.010764731094241142, + 0.012547104619443417, + 0.004090403206646442, + 0.013979091309010983, + 0.003745701629668474, + 0.0033402082044631243, + 0.009715637192130089, + 0.003240786259993911, + 0.005581619217991829, + 0.013143415562808514, + 0.008406142704188824, + 0.012297721579670906, + 0.007945348508656025, + 0.012065096758306026, + 0.014712880365550518, + 0.009551617316901684, + 0.01203012652695179, + 0.022836478427052498, + 0.01658414490520954, + 0.018948685377836227, + 0.014721828512847424, + 0.03190173581242561, + 0.030134564265608788, + 0.03899398073554039, + 0.026125794276595116, + 0.02959524653851986, + 0.012547153048217297, + 0.02945931814610958, + 0.00822197925299406, + 0.016614289954304695, + 0.033602215349674225, + 0.017956197261810303, + 0.03918972983956337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.4289668798446655 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with a fish-eye lens effect causing distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a clown wearing a loose, faded costume with mixed vertical stripes and polka dots visible in the center of a spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the clown exhibit an exaggerated smile, thick cracked white makeup, deep-set black eye sockets, and a shaggy bright orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an empty circus tent with rows of empty wooden benches, massive bent tent poles, and a sawdust-strewn floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005960665177553892 + ], + "opposite_logprob_tokens": [ + -5.130960464477539 + ], + "logprob_reward": -0.005960665177553892, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005960665177553892, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20710515975952148, + 0.08513183891773224, + 0.13590674102306366, + 0.10588863492012024, + 0.05364571511745453, + 0.06107497215270996, + 0.07053499668836594, + 0.12320062518119812, + 0.1024429053068161, + 0.055068422108888626 + ], + "credit_map": [ + 0.025665750727057457, + 0.05021098256111145, + 0.025265024974942207, + 0.025573404505848885, + 0.016662966459989548, + 0.025406304746866226, + 0.018886754289269447, + 0.019433969631791115, + 0.009313325397670269, + 0.014539727009832859, + 0.01940399594604969, + 0.008974066935479641, + 0.006033499725162983, + 0.014286741614341736, + 0.006316194776445627, + 0.006264295428991318, + 0.013827917166054249, + 0.026867734268307686, + 0.018473397940397263, + 0.017110180109739304, + 0.013821035623550415, + 0.024155717343091965, + 0.011004777625203133, + 0.010645991191267967, + 0.010784699581563473, + 0.013460177928209305, + 0.025851663202047348, + 0.011972724460065365, + 0.006793825887143612, + 0.02311878465116024, + 0.004911594092845917, + 0.008995160460472107, + 0.0061886548064649105, + 0.003994543105363846, + 0.012017074972391129, + 0.004141591023653746, + 0.004378024954348803, + 0.013661571778357029, + 0.003192795440554619, + 0.006071458570659161, + 0.007987076416611671, + 0.006651066243648529, + 0.01036748569458723, + 0.006593179889023304, + 0.00736757880076766, + 0.00882772821933031, + 0.005817424971610308, + 0.007463430054485798, + 0.014087767340242863, + 0.0073038022965192795, + 0.012762421742081642, + 0.005993665661662817, + 0.00691627524793148, + 0.009961367584764957, + 0.0060402764938771725, + 0.007469422649592161, + 0.016100112348794937, + 0.009094076231122017, + 0.01478142011910677, + 0.013047285377979279, + 0.011601022444665432, + 0.024629555642604828, + 0.012236414477229118, + 0.021710747852921486, + 0.017867915332317352, + 0.004544831812381744, + 0.014488017186522484, + 0.005247167777270079, + 0.014114756137132645, + 0.017257116734981537, + 0.009753904305398464, + 0.019169190898537636, + 0.006723993923515081, + 0.003992024809122086, + 0.006107376888394356, + 0.0032813360448926687, + 0.006417640019208193, + 0.008869748562574387, + 0.009810932911932468, + 0.009865368716418743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.005960665177553892 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with a fish-eye lens effect causing distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a clown wearing a loose, faded costume with mixed vertical stripes and polka dots visible in the center of a spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the clown exhibit an exaggerated smile, thick cracked white makeup, deep-set black eye sockets, and a shaggy bright orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an empty circus tent with rows of empty wooden benches, massive bent tent poles, and a sawdust-strewn floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020589664578437805 + ], + "opposite_logprob_tokens": [ + -3.895589590072632 + ], + "logprob_reward": -0.020589664578437805, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020589664578437805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24928128719329834, + 0.07232684642076492, + 0.10805056244134903, + 0.06301078200340271, + 0.061974816024303436, + 0.051464539021253586, + 0.06408489495515823, + 0.15617726743221283, + 0.1317991316318512, + 0.041829921305179596 + ], + "credit_map": [ + 0.03086402267217636, + 0.06400979310274124, + 0.034129563719034195, + 0.030921807512640953, + 0.018979791551828384, + 0.03400027006864548, + 0.016299111768603325, + 0.020076945424079895, + 0.007996559143066406, + 0.010462052188813686, + 0.018484868109226227, + 0.006353980395942926, + 0.0056455922313034534, + 0.011993343010544777, + 0.005744555499404669, + 0.0056458949111402035, + 0.009479865431785583, + 0.01921194978058338, + 0.01587524823844433, + 0.014463025145232677, + 0.009223450906574726, + 0.024560030549764633, + 0.006889010779559612, + 0.00834798812866211, + 0.006611647084355354, + 0.008555969223380089, + 0.016729528084397316, + 0.005313749425113201, + 0.002205487573519349, + 0.017632870003581047, + 0.0020219460129737854, + 0.003939585294574499, + 0.008248483762145042, + 0.005225740373134613, + 0.011624271981418133, + 0.004009981174021959, + 0.00483583752065897, + 0.017743924632668495, + 0.004075759556144476, + 0.0062108165584504604, + 0.006626694463193417, + 0.005764547269791365, + 0.008931596763432026, + 0.006858807988464832, + 0.005115077365189791, + 0.007351929787546396, + 0.00505254277959466, + 0.005763343069702387, + 0.0122868986800313, + 0.006228952668607235, + 0.012140369042754173, + 0.006826551165431738, + 0.0049786982126533985, + 0.00952066294848919, + 0.005300849676132202, + 0.006801911164075136, + 0.02287604846060276, + 0.01092611625790596, + 0.018563106656074524, + 0.012785926461219788, + 0.015781428664922714, + 0.03498857840895653, + 0.013035089708864689, + 0.027220970019698143, + 0.011759602464735508, + 0.011233238503336906, + 0.010013244114816189, + 0.009112468920648098, + 0.03553735464811325, + 0.018975408747792244, + 0.021383006125688553, + 0.013784810900688171, + 0.004130855202674866, + 0.003807886503636837, + 0.005222885403782129, + 0.0039150891825556755, + 0.00765170669183135, + 0.005877399351447821, + 0.006166412960737944, + 0.005057687405496836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.020589664578437805 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with a fish-eye lens effect causing distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a clown wearing a loose, faded costume with mixed vertical stripes and polka dots visible in the center of a spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the clown exhibit an exaggerated smile, thick cracked white makeup, deep-set black eye sockets, and a shaggy bright orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an empty circus tent with rows of empty wooden benches, massive bent tent poles, and a sawdust-strewn floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42892029881477356 + ], + "opposite_logprob_tokens": [ + -1.0539202690124512 + ], + "logprob_reward": -0.42892029881477356, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42892029881477356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21475310623645782, + 0.08701708167791367, + 0.3351171016693115, + 0.03366187959909439, + 0.05318797752261162, + 0.02590288408100605, + 0.0731915757060051, + 0.06717458367347717, + 0.08900079876184464, + 0.020993096753954887 + ], + "credit_map": [ + 0.0286097452044487, + 0.04691262170672417, + 0.03363684192299843, + 0.026471439749002457, + 0.013880460523068905, + 0.029599186033010483, + 0.014770956709980965, + 0.020871823653578758, + 0.01238885335624218, + 0.015231098979711533, + 0.01869431510567665, + 0.008439465425908566, + 0.006828213110566139, + 0.011419259011745453, + 0.006192252971231937, + 0.007823620922863483, + 0.033775411546230316, + 0.0716535821557045, + 0.07350887358188629, + 0.03367333114147186, + 0.012534409761428833, + 0.07983719557523727, + 0.013027708977460861, + 0.017106592655181885, + 0.0034665532875806093, + 0.004832212347537279, + 0.008137245662510395, + 0.0028978430200368166, + 0.0026141188573092222, + 0.007339172530919313, + 0.001895245397463441, + 0.0024794926866889, + 0.010011802427470684, + 0.004690038040280342, + 0.009739825502038002, + 0.004466046579182148, + 0.003878707066178322, + 0.010356664657592773, + 0.003224335378035903, + 0.006820558104664087, + 0.0036233505234122276, + 0.0029055732302367687, + 0.005155624356120825, + 0.0020774456206709146, + 0.003099125809967518, + 0.003994078375399113, + 0.0019150746520608664, + 0.003132610348984599, + 0.013197825290262699, + 0.008165135979652405, + 0.012033773586153984, + 0.0066492268815636635, + 0.0066724601201713085, + 0.010610437951982021, + 0.007172683719545603, + 0.008690034039318562, + 0.01110184844583273, + 0.00675096083432436, + 0.006914293859153986, + 0.0036898895632475615, + 0.010996740311384201, + 0.011477990075945854, + 0.007017726078629494, + 0.00922513846307993, + 0.016328947618603706, + 0.005451289936900139, + 0.01450850535184145, + 0.004235907457768917, + 0.009035798721015453, + 0.015753991901874542, + 0.008593238890171051, + 0.015093117021024227, + 0.00242245988920331, + 0.0018142623594030738, + 0.0019492245046421885, + 0.0014770255656912923, + 0.0036339934449642897, + 0.0031271532643586397, + 0.003714632010087371, + 0.002854344667866826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.42892029881477356 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fish-eye lens, at an extremely low angle, slowly pans to the right, capturing a distorted view of the interior of an empty circus tent. A single, pale spotlight shines vertically from above, creating a harsh circle of light on the center stage, while the rest of the tent is swallowed by a bottomless darkness. Within the circle of light, a clown with an exaggerated smile paces. His face is covered in thick, slightly cracked white makeup. A large, fixed red smile is painted morbidly from ear to ear. Deep-set black eye sockets contrast eerily with this false cheerfulness, and a shaggy, bright orange wig appears lifeless. He wears a loose, faded clown costume with a mixed pattern (half vertical stripes, half polka dots). The floppy ruffled collar is slightly yellowed. His well-worn, heavy-toed shoes drag heavily across the sawdust-strewn floor, in an uneasy, repetitive wandering. His shoulders slumped slightly, his hands alternately clenched into fists and shoved feebly into his pockets, his posture completely at odds with the perpetual smile on his face. Through the lens of a fisheye lens, the rows of empty wooden benches and the massive tent poles bent inward, like a shrinking, claustrophobic cage. With his anxious steps, his massive, elongated, and distorted shadow danced grotesquely across the ground, casting a cold, eerie, oppressive atmosphere tinged with intense anxiety.", + "seed": 299, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_0/index_9805_seed_299_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1382320615462959, + "vlm_logprob": -0.1382320615462959, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.1382320615462959, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with a fish-eye lens effect causing distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a clown wearing a loose, faded costume with mixed vertical stripes and polka dots visible in the center of a spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the clown exhibit an exaggerated smile, thick cracked white makeup, deep-set black eye sockets, and a shaggy bright orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an empty circus tent with rows of empty wooden benches, massive bent tent poles, and a sawdust-strewn floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4064183235168457 + ], + "opposite_logprob_tokens": [ + -1.1564183235168457 + ], + "logprob_reward": -0.4064183235168457, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.4064183235168457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19875888526439667, + 0.05053206905722618, + 0.029276439920067787, + 0.14360308647155762, + 0.0796007439494133, + 0.10059946030378342, + 0.1072038784623146, + 0.18315808475017548, + 0.042220525443553925, + 0.06504687666893005 + ], + "credit_map": [ + 0.042307738214731216, + 0.04340667277574539, + 0.030689870938658714, + 0.01804056577384472, + 0.015478384681046009, + 0.020722679793834686, + 0.014008238911628723, + 0.01410471461713314, + 0.009696618653833866, + 0.00713234581053257, + 0.007771743927150965, + 0.005604210775345564, + 0.005406467709690332, + 0.00814584456384182, + 0.0036629829555749893, + 0.0031118574552237988, + 0.003038522321730852, + 0.005278776865452528, + 0.004736155737191439, + 0.0025423376355320215, + 0.003744531190022826, + 0.005118942819535732, + 0.0024239534977823496, + 0.0023932200856506824, + 0.031122742220759392, + 0.020150143653154373, + 0.01945367455482483, + 0.010576565749943256, + 0.010715489275753498, + 0.021563515067100525, + 0.012106935493648052, + 0.017914025112986565, + 0.009772335179150105, + 0.008480889722704887, + 0.01201685145497322, + 0.005834236275404692, + 0.010266517288982868, + 0.013093727640807629, + 0.00953537318855524, + 0.010600818321108818, + 0.01265387050807476, + 0.010920626111328602, + 0.01842806115746498, + 0.010464969091117382, + 0.010132228955626488, + 0.014330311678349972, + 0.008242730051279068, + 0.01542666181921959, + 0.011742006987333298, + 0.009888698346912861, + 0.019268274307250977, + 0.009073266759514809, + 0.015497894957661629, + 0.013536782935261726, + 0.013846154324710369, + 0.014350801706314087, + 0.014294860884547234, + 0.005804912652820349, + 0.05983468145132065, + 0.00500712962821126, + 0.006287512835115194, + 0.016192015260457993, + 0.005643559619784355, + 0.07009340077638626, + 0.006408093962818384, + 0.0036294434685260057, + 0.004547504708170891, + 0.0036395429633557796, + 0.005086980760097504, + 0.0075891027227044106, + 0.003902208525687456, + 0.007417649496346712, + 0.007259533274918795, + 0.008725305087864399, + 0.005673492327332497, + 0.00784753542393446, + 0.009523037821054459, + 0.00859295018017292, + 0.007403720170259476, + 0.010021304711699486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.4064183235168457 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with a fish-eye lens effect causing distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a clown wearing a loose, faded costume with mixed vertical stripes and polka dots visible in the center of a spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the clown exhibit an exaggerated smile, thick cracked white makeup, deep-set black eye sockets, and a shaggy bright orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an empty circus tent with rows of empty wooden benches, massive bent tent poles, and a sawdust-strewn floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25211676955223083 + ], + "opposite_logprob_tokens": [ + -1.5021167993545532 + ], + "logprob_reward": -0.25211676955223083, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25211676955223083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15539617836475372, + 0.035741548985242844, + 0.026785383000969887, + 0.0715072974562645, + 0.14404019713401794, + 0.191809743642807, + 0.1098736971616745, + 0.19549201428890228, + 0.016190839931368828, + 0.05316299945116043 + ], + "credit_map": [ + 0.019772930070757866, + 0.030894003808498383, + 0.023353077471256256, + 0.015232187695801258, + 0.01855979487299919, + 0.015307266265153885, + 0.016979144886136055, + 0.015297791920602322, + 0.004903991706669331, + 0.006197973154485226, + 0.0050954995676875114, + 0.003991854842752218, + 0.002771348925307393, + 0.0056532337330281734, + 0.00267465109936893, + 0.0044529978185892105, + 0.0038532400503754616, + 0.00378782139159739, + 0.003678983310237527, + 0.00264887441881001, + 0.0028516526799649, + 0.004241893533617258, + 0.0021928222849965096, + 0.0035300971940159798, + 0.008092984557151794, + 0.008689358830451965, + 0.010311366990208626, + 0.007056823931634426, + 0.007782130502164364, + 0.010854760184884071, + 0.007158386521041393, + 0.011561491526663303, + 0.029799599200487137, + 0.01414214726537466, + 0.02134946547448635, + 0.01139336172491312, + 0.007975196465849876, + 0.023935280740261078, + 0.008137623779475689, + 0.027307529002428055, + 0.02631193771958351, + 0.020824795588850975, + 0.034222327172756195, + 0.01378705631941557, + 0.021745914593338966, + 0.02735638990998268, + 0.01716112345457077, + 0.030400224030017853, + 0.009776389226317406, + 0.007414006162434816, + 0.025948168709874153, + 0.007121805101633072, + 0.012262551113963127, + 0.011594320647418499, + 0.012538233771920204, + 0.02321823313832283, + 0.015650127083063126, + 0.004804814234375954, + 0.06770702451467514, + 0.00786095205694437, + 0.005110797472298145, + 0.012373048812150955, + 0.005661179311573505, + 0.07632409036159515, + 0.0020758642349392176, + 0.001754286582581699, + 0.0016954074380919337, + 0.0018580981995910406, + 0.0018916920525953174, + 0.0017043681582435966, + 0.002897992730140686, + 0.002313130535185337, + 0.004991843830794096, + 0.00647432217374444, + 0.005227718967944384, + 0.006931161042302847, + 0.006744185462594032, + 0.006677241530269384, + 0.006365412846207619, + 0.009751120582222939 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.25211676955223083 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with a fish-eye lens effect causing distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a clown wearing a loose, faded costume with mixed vertical stripes and polka dots visible in the center of a spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the clown exhibit an exaggerated smile, thick cracked white makeup, deep-set black eye sockets, and a shaggy bright orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an empty circus tent with rows of empty wooden benches, massive bent tent poles, and a sawdust-strewn floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001720973290503025 + ], + "opposite_logprob_tokens": [ + -6.376720905303955 + ], + "logprob_reward": -0.001720973290503025, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001720973290503025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10367792099714279, + 0.048340167850255966, + 0.026731101796030998, + 0.07756579667329788, + 0.0971284881234169, + 0.07052449882030487, + 0.056100521236658096, + 0.4189077913761139, + 0.017525754868984222, + 0.08349793404340744 + ], + "credit_map": [ + 0.017217187210917473, + 0.01900898478925228, + 0.01697145402431488, + 0.013227996416389942, + 0.008008836768567562, + 0.011338743381202221, + 0.010257084853947163, + 0.007647632621228695, + 0.007285775151103735, + 0.007590534631162882, + 0.007779810577630997, + 0.004996385890990496, + 0.003604383673518896, + 0.0089625995606184, + 0.0033392750192433596, + 0.0047814082354307175, + 0.003318200819194317, + 0.0035939484369009733, + 0.003937484230846167, + 0.002447462175041437, + 0.0036345915868878365, + 0.0036084118764847517, + 0.003077833680436015, + 0.0031131699215620756, + 0.013416170142591, + 0.012074315920472145, + 0.00900905579328537, + 0.006571020465344191, + 0.0072358595207333565, + 0.008979390375316143, + 0.01039031334221363, + 0.009889676235616207, + 0.02072514407336712, + 0.009184313006699085, + 0.013733291998505592, + 0.007592161651700735, + 0.0063704149797558784, + 0.014745009131729603, + 0.009187246672809124, + 0.015590911731123924, + 0.009598269127309322, + 0.008463441394269466, + 0.010136034339666367, + 0.006446850951761007, + 0.007467768620699644, + 0.011962389573454857, + 0.0066391294822096825, + 0.009810615330934525, + 0.0092386519536376, + 0.006754010915756226, + 0.011975696310400963, + 0.0028463939670473337, + 0.0041510616429150105, + 0.007122879847884178, + 0.0037352086510509253, + 0.010276627726852894, + 0.033662088215351105, + 0.00968180038034916, + 0.14527612924575806, + 0.014904706738889217, + 0.009388981387019157, + 0.027858903631567955, + 0.010551666840910912, + 0.16758352518081665, + 0.002156627830117941, + 0.0018866212340071797, + 0.0019279135158285499, + 0.0018592308042570949, + 0.002192443236708641, + 0.002263830043375492, + 0.0019043181091547012, + 0.0033347716089338064, + 0.01392200868576765, + 0.0053895763121545315, + 0.00800018198788166, + 0.004410198424011469, + 0.01000388152897358, + 0.022333970293402672, + 0.00504314387217164, + 0.014394988305866718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.001720973290503025 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with a fish-eye lens effect causing distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a clown wearing a loose, faded costume with mixed vertical stripes and polka dots visible in the center of a spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the clown exhibit an exaggerated smile, thick cracked white makeup, deep-set black eye sockets, and a shaggy bright orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an empty circus tent with rows of empty wooden benches, massive bent tent poles, and a sawdust-strewn floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00763084227219224 + ], + "opposite_logprob_tokens": [ + -4.882630825042725 + ], + "logprob_reward": -0.00763084227219224, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00763084227219224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13380268216133118, + 0.08280612528324127, + 0.03840980678796768, + 0.12873631715774536, + 0.056631918996572495, + 0.270175576210022, + 0.06596185266971588, + 0.16124586760997772, + 0.020817529410123825, + 0.04141233488917351 + ], + "credit_map": [ + 0.018839217722415924, + 0.019294830039143562, + 0.02385047636926174, + 0.015210705809295177, + 0.014145837165415287, + 0.0129806874319911, + 0.018785186111927032, + 0.010695736855268478, + 0.009814501740038395, + 0.010086044669151306, + 0.013683749362826347, + 0.009526057168841362, + 0.010745072737336159, + 0.013964693062007427, + 0.008136219345033169, + 0.006849787663668394, + 0.007067841477692127, + 0.0063652521930634975, + 0.005631642881780863, + 0.0026240090373903513, + 0.0043154130689799786, + 0.006973074283450842, + 0.002679710742086172, + 0.002752861240878701, + 0.027903562411665916, + 0.013555598445236683, + 0.017222370952367783, + 0.008613616228103638, + 0.010850466787815094, + 0.020770853385329247, + 0.011660881340503693, + 0.01815895549952984, + 0.011649241670966148, + 0.006409935653209686, + 0.0076459841802716255, + 0.005212553311139345, + 0.005377238150686026, + 0.007892757654190063, + 0.0041689202189445496, + 0.008275291882455349, + 0.03954407572746277, + 0.026341311633586884, + 0.04709408059716225, + 0.019670264795422554, + 0.03090718202292919, + 0.04074706882238388, + 0.02475009858608246, + 0.041121479123830795, + 0.008815469220280647, + 0.006425558123737574, + 0.015046986751258373, + 0.00469342153519392, + 0.0054099904373288155, + 0.006114867981523275, + 0.007107526063919067, + 0.01234803069382906, + 0.01670670509338379, + 0.0049196104519069195, + 0.050287071615457535, + 0.005710262805223465, + 0.004640900529921055, + 0.012170562520623207, + 0.004845003597438335, + 0.061965763568878174, + 0.0032428365666419268, + 0.0017530826153233647, + 0.0023697882425040007, + 0.0019406569190323353, + 0.0021342174150049686, + 0.003688455792143941, + 0.0024811874609440565, + 0.003207304049283266, + 0.005676206201314926, + 0.003930698148906231, + 0.004152792040258646, + 0.0046147918328642845, + 0.0052257077768445015, + 0.006328721530735493, + 0.004670920316129923, + 0.006812496576458216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.00763084227219224 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with a fish-eye lens effect causing distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a clown wearing a loose, faded costume with mixed vertical stripes and polka dots visible in the center of a spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the clown exhibit an exaggerated smile, thick cracked white makeup, deep-set black eye sockets, and a shaggy bright orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an empty circus tent with rows of empty wooden benches, massive bent tent poles, and a sawdust-strewn floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023273399099707603 + ], + "opposite_logprob_tokens": [ + -3.773273468017578 + ], + "logprob_reward": -0.023273399099707603, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023273399099707603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11588220298290253, + 0.06979884207248688, + 0.02585318684577942, + 0.07069794833660126, + 0.026271017268300056, + 0.24153175950050354, + 0.06634096801280975, + 0.20150065422058105, + 0.026976220309734344, + 0.15514716506004333 + ], + "credit_map": [ + 0.02523728273808956, + 0.024528637528419495, + 0.019862191751599312, + 0.012423740699887276, + 0.007031713146716356, + 0.008153527975082397, + 0.0074334461241960526, + 0.011211659759283066, + 0.011470277793705463, + 0.010828699916601181, + 0.011537912301719189, + 0.00725275045260787, + 0.003833866212517023, + 0.014788161031901836, + 0.003139827400445938, + 0.0069473423063755035, + 0.00403659837320447, + 0.00415424769744277, + 0.0036304560489952564, + 0.0022016719449311495, + 0.0023154905065894127, + 0.004654984455555677, + 0.0016472504939883947, + 0.0032124866265803576, + 0.01486087404191494, + 0.01101738028228283, + 0.008886532858014107, + 0.0064593954011797905, + 0.005724526010453701, + 0.009400082752108574, + 0.006603822112083435, + 0.007745330687612295, + 0.003960971720516682, + 0.0032371939159929752, + 0.0036437753587961197, + 0.002127352636307478, + 0.001824545906856656, + 0.00574338436126709, + 0.0014555681264027953, + 0.0042782253585755825, + 0.03174266964197159, + 0.024791236966848373, + 0.044726379215717316, + 0.018571246415376663, + 0.023485438898205757, + 0.03790389001369476, + 0.019070731475949287, + 0.04124016314744949, + 0.00963430292904377, + 0.00703820213675499, + 0.013554248958826065, + 0.003881333628669381, + 0.004992250353097916, + 0.010075868107378483, + 0.0036092596128582954, + 0.013555504381656647, + 0.016931386664509773, + 0.004467811901122332, + 0.06837962567806244, + 0.0060245078057050705, + 0.005763502325862646, + 0.015214580111205578, + 0.004491065628826618, + 0.08022817969322205, + 0.003972074948251247, + 0.0017415935872122645, + 0.0033653504215180874, + 0.0017281617037951946, + 0.003726840717718005, + 0.005898313131183386, + 0.002523195231333375, + 0.004020692314952612, + 0.02705051563680172, + 0.007663560099899769, + 0.014703108929097652, + 0.0077409083023667336, + 0.019382409751415253, + 0.043627675622701645, + 0.008814300410449505, + 0.026164676994085312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 9 + ], + "selected_reward": -0.023273399099707603 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fish-eye lens, at an extremely low angle, slowly pans to the right, capturing a distorted view of the interior of an empty circus tent. A single, pale spotlight shines vertically from above, creating a harsh circle of light on the center stage, while the rest of the tent is swallowed by a bottomless darkness. Within the circle of light, a clown with an exaggerated smile paces. His face is covered in thick, slightly cracked white makeup. A large, fixed red smile is painted morbidly from ear to ear. Deep-set black eye sockets contrast eerily with this false cheerfulness, and a shaggy, bright orange wig appears lifeless. He wears a loose, faded clown costume with a mixed pattern (half vertical stripes, half polka dots). The floppy ruffled collar is slightly yellowed. His well-worn, heavy-toed shoes drag heavily across the sawdust-strewn floor, in an uneasy, repetitive wandering. His shoulders slumped slightly, his hands alternately clenched into fists and shoved feebly into his pockets, his posture completely at odds with the perpetual smile on his face. Through the lens of a fisheye lens, the rows of empty wooden benches and the massive tent poles bent inward, like a shrinking, claustrophobic cage. With his anxious steps, his massive, elongated, and distorted shadow danced grotesquely across the ground, casting a cold, eerie, oppressive atmosphere tinged with intense anxiety.", + "seed": 300, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_0/index_9805_seed_300_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.19136623716913165, + "vlm_logprob": -0.19136623716913165, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.19136623716913165, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with a fish-eye lens effect causing distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a clown wearing a loose, faded costume with mixed vertical stripes and polka dots visible in the center of a spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the clown exhibit an exaggerated smile, thick cracked white makeup, deep-set black eye sockets, and a shaggy bright orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an empty circus tent with rows of empty wooden benches, massive bent tent poles, and a sawdust-strewn floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7067386507987976 + ], + "opposite_logprob_tokens": [ + -0.7067386507987976 + ], + "logprob_reward": -0.7067386507987976, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.7067386507987976, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.29484668374061584, + 0.08789093047380447, + 0.056490834802389145, + 0.06927908211946487, + 0.09419471770524979, + 0.12261482328176498, + 0.05086313560605049, + 0.05234623700380325, + 0.11587236076593399, + 0.05560113117098808 + ], + "credit_map": [ + 0.030943943187594414, + 0.10556663572788239, + 0.035138923674821854, + 0.04270080476999283, + 0.019884716719388962, + 0.0295343529433012, + 0.012743324972689152, + 0.018334003165364265, + 0.007149691227823496, + 0.015265564434230328, + 0.012345102615654469, + 0.013757931999862194, + 0.006073149386793375, + 0.01094523910433054, + 0.008585668168962002, + 0.013768591918051243, + 0.004564976319670677, + 0.011231011711061, + 0.007298536133021116, + 0.006966490298509598, + 0.007058888673782349, + 0.006359125021845102, + 0.007013728842139244, + 0.005998080130666494, + 0.007829271256923676, + 0.017070556059479713, + 0.011212260462343693, + 0.006116670556366444, + 0.003716044593602419, + 0.013813324272632599, + 0.004088818561285734, + 0.0054321372881531715, + 0.012363191694021225, + 0.027744542807340622, + 0.009676468558609486, + 0.007458320818841457, + 0.008927878923714161, + 0.013145805336534977, + 0.0066143060103058815, + 0.008264211006462574, + 0.01525107491761446, + 0.028308263048529625, + 0.01688561402261257, + 0.012044604867696762, + 0.012104859575629234, + 0.015397932380437851, + 0.010249199345707893, + 0.012373288162052631, + 0.006539538502693176, + 0.008996191434562206, + 0.006555912550538778, + 0.0049599152989685535, + 0.0061971209943294525, + 0.006705272011458874, + 0.004870256409049034, + 0.006038933526724577, + 0.006252487190067768, + 0.008657712489366531, + 0.005755356978625059, + 0.004311584867537022, + 0.004872388206422329, + 0.008276509121060371, + 0.005938328802585602, + 0.00828187819570303, + 0.009501976892352104, + 0.010309305042028427, + 0.007603125646710396, + 0.006670563016086817, + 0.004830242600291967, + 0.04005526378750801, + 0.006838865578174591, + 0.030063029378652573, + 0.004477354697883129, + 0.0030680964700877666, + 0.0023431985173374414, + 0.002124964725226164, + 0.002856101607903838, + 0.020916886627674103, + 0.003445050213485956, + 0.016369478777050972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.7067386507987976 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with a fish-eye lens effect causing distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a clown wearing a loose, faded costume with mixed vertical stripes and polka dots visible in the center of a spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the clown exhibit an exaggerated smile, thick cracked white makeup, deep-set black eye sockets, and a shaggy bright orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an empty circus tent with rows of empty wooden benches, massive bent tent poles, and a sawdust-strewn floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16039887070655823 + ], + "opposite_logprob_tokens": [ + -1.9103988409042358 + ], + "logprob_reward": -0.16039887070655823, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16039887070655823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08858504891395569, + 0.07715995609760284, + 0.03731732815504074, + 0.06919028609991074, + 0.09323250502347946, + 0.04682816565036774, + 0.04092796891927719, + 0.11474176496267319, + 0.26511189341545105, + 0.16690514981746674 + ], + "credit_map": [ + 0.008080235682427883, + 0.017336662858724594, + 0.011969994753599167, + 0.017810910940170288, + 0.009947774931788445, + 0.00899107102304697, + 0.007116778753697872, + 0.007331619504839182, + 0.006642555817961693, + 0.010414653457701206, + 0.01602167822420597, + 0.016499018296599388, + 0.004699926357716322, + 0.007319215219467878, + 0.005714446771889925, + 0.009848465211689472, + 0.0036487593315541744, + 0.007584350183606148, + 0.005418673157691956, + 0.005789455957710743, + 0.004419505596160889, + 0.00344525882974267, + 0.003549812128767371, + 0.0034615122713148594, + 0.009501595981419086, + 0.015889307484030724, + 0.008137664757668972, + 0.00743472296744585, + 0.005684617906808853, + 0.008904636837542057, + 0.006189352832734585, + 0.0074483794160187244, + 0.010972565039992332, + 0.024733083322644234, + 0.010825078003108501, + 0.011022308841347694, + 0.007370103616267443, + 0.012713956646621227, + 0.0060828085988759995, + 0.009512601420283318, + 0.005709120538085699, + 0.008023201487958431, + 0.00602152943611145, + 0.005583911668509245, + 0.0048950365744531155, + 0.006886118557304144, + 0.003982120659202337, + 0.005727124400436878, + 0.004645761102437973, + 0.00800030305981636, + 0.004926601890474558, + 0.003956059459596872, + 0.004276207182556391, + 0.005259371362626553, + 0.005420978181064129, + 0.004442686215043068, + 0.01268787868320942, + 0.018263565376400948, + 0.011518371291458607, + 0.00901973806321621, + 0.013977388851344585, + 0.01849737949669361, + 0.015341068617999554, + 0.015436382032930851, + 0.018383435904979706, + 0.02416137605905533, + 0.013618594966828823, + 0.015351451933383942, + 0.0090154018253088, + 0.09174410253763199, + 0.015021156519651413, + 0.07781637459993362, + 0.01072978600859642, + 0.007116805296391249, + 0.006588007789105177, + 0.0041345758363604546, + 0.011163903400301933, + 0.06151385232806206, + 0.013902935199439526, + 0.05175528675317764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.16039887070655823 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with a fish-eye lens effect causing distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a clown wearing a loose, faded costume with mixed vertical stripes and polka dots visible in the center of a spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the clown exhibit an exaggerated smile, thick cracked white makeup, deep-set black eye sockets, and a shaggy bright orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an empty circus tent with rows of empty wooden benches, massive bent tent poles, and a sawdust-strewn floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036239444743841887 + ], + "opposite_logprob_tokens": [ + -5.628623962402344 + ], + "logprob_reward": -0.0036239444743841887, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036239444743841887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12774676084518433, + 0.13954895734786987, + 0.036364294588565826, + 0.04385450854897499, + 0.08706291019916534, + 0.04921533539891243, + 0.046175409108400345, + 0.12885643541812897, + 0.32040250301361084, + 0.020772915333509445 + ], + "credit_map": [ + 0.014215185306966305, + 0.046504419296979904, + 0.016896001994609833, + 0.019365038722753525, + 0.006661649793386459, + 0.012538451701402664, + 0.005787355359643698, + 0.005778641905635595, + 0.011964605189859867, + 0.021483082324266434, + 0.025588838383555412, + 0.031227990984916687, + 0.007325541693717241, + 0.013522005639970303, + 0.010160946287214756, + 0.01827593520283699, + 0.0028828855138272047, + 0.006659143604338169, + 0.005905402358621359, + 0.003958690445870161, + 0.004548043943941593, + 0.004349039401859045, + 0.0038837261963635683, + 0.004177358467131853, + 0.004850096069276333, + 0.010606067255139351, + 0.007013948168605566, + 0.005885844584554434, + 0.003581637516617775, + 0.004637174773961306, + 0.0033914176747202873, + 0.003888320177793503, + 0.012010267935693264, + 0.025290662422776222, + 0.009850142523646355, + 0.0083839800208807, + 0.006179386284202337, + 0.011113828048110008, + 0.004783291835337877, + 0.009451343677937984, + 0.008899496868252754, + 0.010024551302194595, + 0.007921493612229824, + 0.004410736262798309, + 0.0029132405761629343, + 0.00650803791359067, + 0.0026204341556876898, + 0.005917341448366642, + 0.006382178049534559, + 0.007954816333949566, + 0.004307136405259371, + 0.004522669594734907, + 0.004134820774197578, + 0.006615952588617802, + 0.0066133360378444195, + 0.0056444983929395676, + 0.014187503606081009, + 0.01852281205356121, + 0.012927839532494545, + 0.010365513153374195, + 0.014624174684286118, + 0.021820126101374626, + 0.017522942274808884, + 0.01888551004230976, + 0.021644841879606247, + 0.030446039512753487, + 0.017253555357456207, + 0.020608719438314438, + 0.01138890627771616, + 0.1075940951704979, + 0.018706392496824265, + 0.09275991469621658, + 0.0016792701790109277, + 0.0019962929654866457, + 0.0015170603292062879, + 0.0013356845593079925, + 0.0019919162150472403, + 0.00539385387673974, + 0.0017359189223498106, + 0.005122917238622904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 7 + ], + "selected_reward": -0.0036239444743841887 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with a fish-eye lens effect causing distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a clown wearing a loose, faded costume with mixed vertical stripes and polka dots visible in the center of a spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the clown exhibit an exaggerated smile, thick cracked white makeup, deep-set black eye sockets, and a shaggy bright orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an empty circus tent with rows of empty wooden benches, massive bent tent poles, and a sawdust-strewn floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016072357073426247 + ], + "opposite_logprob_tokens": [ + -4.1410722732543945 + ], + "logprob_reward": -0.016072357073426247, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016072357073426247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10720554739236832, + 0.09465896338224411, + 0.05162739008665085, + 0.06015018746256828, + 0.09386910498142242, + 0.055122122168540955, + 0.05331723392009735, + 0.12591788172721863, + 0.30536311864852905, + 0.052768487483263016 + ], + "credit_map": [ + 0.01096162386238575, + 0.021947629749774933, + 0.015171392820775509, + 0.02053910866379738, + 0.009016221389174461, + 0.012574988417327404, + 0.008275859989225864, + 0.008718724362552166, + 0.00904612522572279, + 0.01681758277118206, + 0.014158808626234531, + 0.02082033082842827, + 0.006119459867477417, + 0.009311373345553875, + 0.007241460494697094, + 0.01114381942898035, + 0.003888794220983982, + 0.009113899432122707, + 0.007851321250200272, + 0.005052372347563505, + 0.007674755994230509, + 0.005852073431015015, + 0.0060293590649962425, + 0.006164814345538616, + 0.0072380355559289455, + 0.012921412475407124, + 0.010676072910428047, + 0.006873597856611013, + 0.004224452655762434, + 0.008523293770849705, + 0.003952388651669025, + 0.005740932654589415, + 0.012958837673068047, + 0.024946661666035652, + 0.011216934770345688, + 0.011277168989181519, + 0.006475904956459999, + 0.01206634659320116, + 0.004655186552554369, + 0.010272066108882427, + 0.009840420447289944, + 0.013213707134127617, + 0.007072653621435165, + 0.004176566377282143, + 0.003481107298284769, + 0.007097129710018635, + 0.003875037655234337, + 0.006365499459207058, + 0.0072092595510184765, + 0.008634560741484165, + 0.004743111785501242, + 0.0050660097040236, + 0.007157726678997278, + 0.006545847747474909, + 0.00786880124360323, + 0.0060919178649783134, + 0.01434260979294777, + 0.017684126272797585, + 0.012824099510908127, + 0.010380340740084648, + 0.013556662015616894, + 0.020011989399790764, + 0.016909604892134666, + 0.020208435133099556, + 0.02044975571334362, + 0.029314134269952774, + 0.016577478498220444, + 0.020458664745092392, + 0.011108285747468472, + 0.10137538611888885, + 0.018336854875087738, + 0.08774254471063614, + 0.004164160694926977, + 0.004097753670066595, + 0.0036742158699780703, + 0.0024716367479413748, + 0.002941632177680731, + 0.016396580263972282, + 0.0036194720305502415, + 0.01540303323417902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.016072357073426247 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with a fish-eye lens effect causing distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a clown wearing a loose, faded costume with mixed vertical stripes and polka dots visible in the center of a spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the clown exhibit an exaggerated smile, thick cracked white makeup, deep-set black eye sockets, and a shaggy bright orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an empty circus tent with rows of empty wooden benches, massive bent tent poles, and a sawdust-strewn floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999736279249191 + ], + "opposite_logprob_tokens": [ + -2.6949973106384277 + ], + "logprob_reward": -0.06999736279249191, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999736279249191, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1546332836151123, + 0.13543199002742767, + 0.03473629057407379, + 0.046055156737565994, + 0.07747218012809753, + 0.0542454719543457, + 0.04052667319774628, + 0.13242179155349731, + 0.310867041349411, + 0.01361018419265747 + ], + "credit_map": [ + 0.017141131684184074, + 0.05797171965241432, + 0.02076868526637554, + 0.025776231661438942, + 0.006898702122271061, + 0.014370608143508434, + 0.005933993496000767, + 0.005772185046225786, + 0.011831951327621937, + 0.02354811504483223, + 0.02253497764468193, + 0.030508074909448624, + 0.007024018093943596, + 0.013090329244732857, + 0.009488645009696484, + 0.01740586943924427, + 0.002725389087572694, + 0.006227496080100536, + 0.005204787477850914, + 0.003804620122537017, + 0.005096632055938244, + 0.003785362234339118, + 0.004292519297450781, + 0.0035994781646877527, + 0.005473915487527847, + 0.010384989902377129, + 0.007109668105840683, + 0.005767942871898413, + 0.0032796987798064947, + 0.0061749960295856, + 0.0031095086596906185, + 0.004754429683089256, + 0.009910427033901215, + 0.022570529952645302, + 0.008998158387839794, + 0.007480870001018047, + 0.0056512062437832355, + 0.010301655158400536, + 0.004632017575204372, + 0.007927307859063148, + 0.010065609589219093, + 0.011249033734202385, + 0.008946680463850498, + 0.004657289013266563, + 0.0026346456725150347, + 0.007258885074406862, + 0.002719608834013343, + 0.006713710725307465, + 0.005610054824501276, + 0.007476442493498325, + 0.0037882181350141764, + 0.0036407383158802986, + 0.004163355100899935, + 0.005272173788398504, + 0.005673677194863558, + 0.004902008920907974, + 0.014092442579567432, + 0.01931493915617466, + 0.013167057186365128, + 0.011024662293493748, + 0.015243548899888992, + 0.022966861724853516, + 0.018481168895959854, + 0.01813109591603279, + 0.02098304219543934, + 0.02983231469988823, + 0.016969475895166397, + 0.02011709287762642, + 0.011253353208303452, + 0.10378145426511765, + 0.018213234841823578, + 0.0897170826792717, + 0.001231760368682444, + 0.0018082668539136648, + 0.001220444799400866, + 0.001225442043505609, + 0.0012439641868695617, + 0.002934858901426196, + 0.0012049772776663303, + 0.0027404676657170057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.06999736279249191 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fish-eye lens, at an extremely low angle, slowly pans to the right, capturing a distorted view of the interior of an empty circus tent. A single, pale spotlight shines vertically from above, creating a harsh circle of light on the center stage, while the rest of the tent is swallowed by a bottomless darkness. Within the circle of light, a clown with an exaggerated smile paces. His face is covered in thick, slightly cracked white makeup. A large, fixed red smile is painted morbidly from ear to ear. Deep-set black eye sockets contrast eerily with this false cheerfulness, and a shaggy, bright orange wig appears lifeless. He wears a loose, faded clown costume with a mixed pattern (half vertical stripes, half polka dots). The floppy ruffled collar is slightly yellowed. His well-worn, heavy-toed shoes drag heavily across the sawdust-strewn floor, in an uneasy, repetitive wandering. His shoulders slumped slightly, his hands alternately clenched into fists and shoved feebly into his pockets, his posture completely at odds with the perpetual smile on his face. Through the lens of a fisheye lens, the rows of empty wooden benches and the massive tent poles bent inward, like a shrinking, claustrophobic cage. With his anxious steps, his massive, elongated, and distorted shadow danced grotesquely across the ground, casting a cold, eerie, oppressive atmosphere tinged with intense anxiety.", + "seed": 301, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_0/index_9805_seed_301_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.33180777411907914, + "vlm_logprob": -0.33180777411907914, + "vlm_yes_no_margin": 1.95, + "vlm_token_credit": -0.33180777411907914, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with a fish-eye lens effect causing distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a clown wearing a loose, faded costume with mixed vertical stripes and polka dots visible in the center of a spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the clown exhibit an exaggerated smile, thick cracked white makeup, deep-set black eye sockets, and a shaggy bright orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an empty circus tent with rows of empty wooden benches, massive bent tent poles, and a sawdust-strewn floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9156321883201599 + ], + "opposite_logprob_tokens": [ + -0.5406321883201599 + ], + "logprob_reward": -0.9156321883201599, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.9156321883201599, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05964486673474312, + 0.0459393709897995, + 0.051649563014507294, + 0.04337656497955322, + 0.02308827079832554, + 0.0405050665140152, + 0.3426027297973633, + 0.07246290892362595, + 0.2761377990245819, + 0.044592782855033875 + ], + "credit_map": [ + 0.006767852231860161, + 0.008785913698375225, + 0.007958837784826756, + 0.008162559010088444, + 0.008915350772440434, + 0.009325588122010231, + 0.00451535452157259, + 0.0052134147845208645, + 0.006291023921221495, + 0.0057234084233641624, + 0.006938731297850609, + 0.00406094454228878, + 0.0061798421666026115, + 0.007255183067172766, + 0.004786415491253138, + 0.004703820217400789, + 0.007912795059382915, + 0.0042165908962488174, + 0.00646365899592638, + 0.0033282481599599123, + 0.0069514112547039986, + 0.01291780173778534, + 0.004726052284240723, + 0.005133005324751139, + 0.007765676826238632, + 0.002865628805011511, + 0.0035593612119555473, + 0.0034775175154209137, + 0.003935931250452995, + 0.012786091305315495, + 0.003951115533709526, + 0.005035246256738901, + 0.003209573682397604, + 0.0022807582281529903, + 0.0014215183909982443, + 0.0023596007376909256, + 0.002697580959647894, + 0.005852550733834505, + 0.0019432975677773356, + 0.0033233934082090855, + 0.004058291204273701, + 0.005755266640335321, + 0.00350091140717268, + 0.004099161829799414, + 0.006509617902338505, + 0.006281046662479639, + 0.004423868842422962, + 0.005876902956515551, + 0.015126654878258705, + 0.007014730479568243, + 0.008947746828198433, + 0.0060343449003994465, + 0.03888135030865669, + 0.1380019187927246, + 0.010104808025062084, + 0.11849122494459152, + 0.005541152786463499, + 0.004713483154773712, + 0.003232336835935712, + 0.003139035776257515, + 0.009277396835386753, + 0.02394099161028862, + 0.003841183613985777, + 0.018777335062623024, + 0.027440300211310387, + 0.019161226227879524, + 0.013777271844446659, + 0.016068527474999428, + 0.023477815091609955, + 0.07784871011972427, + 0.016026044264435768, + 0.08233793824911118, + 0.004972325172275305, + 0.0033119372092187405, + 0.003146749222651124, + 0.002949020592495799, + 0.009068664163351059, + 0.006120253354310989, + 0.0062796431593596935, + 0.0087441923096776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.9156321883201599 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with a fish-eye lens effect causing distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a clown wearing a loose, faded costume with mixed vertical stripes and polka dots visible in the center of a spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the clown exhibit an exaggerated smile, thick cracked white makeup, deep-set black eye sockets, and a shaggy bright orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an empty circus tent with rows of empty wooden benches, massive bent tent poles, and a sawdust-strewn floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813621163368225 + ], + "opposite_logprob_tokens": [ + -1.4063620567321777 + ], + "logprob_reward": -0.2813621163368225, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813621163368225, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.041753336787223816, + 0.041260577738285065, + 0.025321194902062416, + 0.03261565417051315, + 0.0092336880043149, + 0.05942727252840996, + 0.214992493391037, + 0.1443420946598053, + 0.36255088448524475, + 0.0685027614235878 + ], + "credit_map": [ + 0.0054903277195990086, + 0.006820967886596918, + 0.004621836822479963, + 0.0054993885569274426, + 0.0075086853466928005, + 0.005351914092898369, + 0.0029565116856247187, + 0.0035037025809288025, + 0.004519780166447163, + 0.006203059572726488, + 0.007400836795568466, + 0.005497903097420931, + 0.004551362711936235, + 0.006109159905463457, + 0.003031087340787053, + 0.003947383724153042, + 0.0037507370579987764, + 0.0033781779929995537, + 0.0035479015205055475, + 0.0025018537417054176, + 0.0036775562912225723, + 0.00332728517241776, + 0.002274711849167943, + 0.00286297220736742, + 0.0043658241629600525, + 0.00363938114605844, + 0.0028978935442864895, + 0.004541883245110512, + 0.004361064173281193, + 0.0053598275408148766, + 0.0038349430542439222, + 0.0036148347426205873, + 0.0011259332532063127, + 0.0012335266219452024, + 0.0011337706819176674, + 0.0011279521277174354, + 0.0011751885758712888, + 0.0014501195400953293, + 0.0007899258052930236, + 0.00119727139826864, + 0.005799143109470606, + 0.005532780662178993, + 0.0044174110516905785, + 0.0046956827864050865, + 0.00896152202039957, + 0.01086543407291174, + 0.005362814757972956, + 0.013792483136057854, + 0.008095230907201767, + 0.006157405208796263, + 0.005509120412170887, + 0.0037304158322513103, + 0.02208457887172699, + 0.08667927980422974, + 0.006039211060851812, + 0.07669723778963089, + 0.016079353168606758, + 0.01438923366367817, + 0.00921718031167984, + 0.008588322438299656, + 0.022284861654043198, + 0.0351036936044693, + 0.012219506315886974, + 0.02645994909107685, + 0.033183109015226364, + 0.010841680690646172, + 0.015931006520986557, + 0.008961382322013378, + 0.023800866678357124, + 0.1258089542388916, + 0.016193632036447525, + 0.12783025205135345, + 0.005932217463850975, + 0.006375445518642664, + 0.0056468164548277855, + 0.006213993765413761, + 0.012118430808186531, + 0.009196166880428791, + 0.01299765519797802, + 0.010022036731243134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.2813621163368225 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with a fish-eye lens effect causing distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a clown wearing a loose, faded costume with mixed vertical stripes and polka dots visible in the center of a spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the clown exhibit an exaggerated smile, thick cracked white makeup, deep-set black eye sockets, and a shaggy bright orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an empty circus tent with rows of empty wooden benches, massive bent tent poles, and a sawdust-strewn floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009793330915272236 + ], + "opposite_logprob_tokens": [ + -4.634793281555176 + ], + "logprob_reward": -0.009793330915272236, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009793330915272236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04921264573931694, + 0.02769298292696476, + 0.026482420042157173, + 0.04030578210949898, + 0.023093121126294136, + 0.08428999036550522, + 0.32379150390625, + 0.09217996150255203, + 0.3110077679157257, + 0.021943775936961174 + ], + "credit_map": [ + 0.006589208729565144, + 0.01090274564921856, + 0.005548587068915367, + 0.008680788800120354, + 0.005418702028691769, + 0.005854214541614056, + 0.00276165665127337, + 0.0034567497204989195, + 0.005146537441760302, + 0.005636643618345261, + 0.003535281866788864, + 0.00306489085778594, + 0.0029618353582918644, + 0.003636626061052084, + 0.0016184492269530892, + 0.0020927214063704014, + 0.00431625172495842, + 0.004776304587721825, + 0.0036629093810915947, + 0.0025897470768541098, + 0.0027187089435756207, + 0.003339907852932811, + 0.0022696650121361017, + 0.0028089270927011967, + 0.00670800032094121, + 0.005044689867645502, + 0.003476503072306514, + 0.004743259400129318, + 0.0029882648959755898, + 0.011315296404063702, + 0.0019528652774170041, + 0.004076903685927391, + 0.004023393616080284, + 0.002793602179735899, + 0.002222844399511814, + 0.0018858788534998894, + 0.002597374841570854, + 0.004434856586158276, + 0.001762045663781464, + 0.0033731269650161266, + 0.007579780649393797, + 0.010879122652113438, + 0.005806670058518648, + 0.00847114622592926, + 0.01444660872220993, + 0.012399622239172459, + 0.006458280608057976, + 0.01824876479804516, + 0.013777071610093117, + 0.005821170285344124, + 0.00899184588342905, + 0.006903698667883873, + 0.03612809628248215, + 0.12959158420562744, + 0.009611982852220535, + 0.11296606063842773, + 0.007413046900182962, + 0.00646628113463521, + 0.004626798443496227, + 0.005072799511253834, + 0.012999004684388638, + 0.02712419629096985, + 0.005762617569416761, + 0.022715218365192413, + 0.028169462457299232, + 0.012388195842504501, + 0.013359235599637032, + 0.010538801550865173, + 0.021072134375572205, + 0.1053512915968895, + 0.014463203959167004, + 0.10566546767950058, + 0.0029549410101026297, + 0.001773976138792932, + 0.0014926455914974213, + 0.0011188513599336147, + 0.0037452098913490772, + 0.004063093569129705, + 0.0020216181874275208, + 0.00477343937382102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.009793330915272236 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with a fish-eye lens effect causing distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a clown wearing a loose, faded costume with mixed vertical stripes and polka dots visible in the center of a spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the clown exhibit an exaggerated smile, thick cracked white makeup, deep-set black eye sockets, and a shaggy bright orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an empty circus tent with rows of empty wooden benches, massive bent tent poles, and a sawdust-strewn floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023300418630242348 + ], + "opposite_logprob_tokens": [ + -3.7733004093170166 + ], + "logprob_reward": -0.023300418630242348, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023300418630242348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03813333809375763, + 0.02254054881632328, + 0.03520090878009796, + 0.011595248244702816, + 0.01531895063817501, + 0.07228297740221024, + 0.46180498600006104, + 0.046911273151636124, + 0.2842627465724945, + 0.011949014849960804 + ], + "credit_map": [ + 0.004987642634660006, + 0.006214594468474388, + 0.005707733798772097, + 0.005427264142781496, + 0.005482629872858524, + 0.004794890061020851, + 0.0022711502388119698, + 0.003247430082410574, + 0.0037046477664262056, + 0.003910992294549942, + 0.0031378392595797777, + 0.0031001302413642406, + 0.0023423265665769577, + 0.002671765862032771, + 0.0013836203143000603, + 0.0022892258130013943, + 0.004439144860953093, + 0.0059617795050144196, + 0.004385848063975573, + 0.003660953603684902, + 0.004392160102725029, + 0.005146448500454426, + 0.003569891443476081, + 0.0036446822341531515, + 0.0016203357372432947, + 0.00159012945368886, + 0.0011868519941344857, + 0.0010828461963683367, + 0.0015320151578634977, + 0.002161390380933881, + 0.0011236793361604214, + 0.0012979995226487517, + 0.002872121287509799, + 0.0018780744867399335, + 0.0015912392409518361, + 0.0010131731396541, + 0.0021478303242474794, + 0.0026064980775117874, + 0.0010976219782605767, + 0.002112392336130142, + 0.006367829162627459, + 0.009089927189052105, + 0.005032737739384174, + 0.0074697220697999, + 0.011492845602333546, + 0.011032036505639553, + 0.005709565710276365, + 0.016088316217064857, + 0.018534239381551743, + 0.006145960185676813, + 0.01076495461165905, + 0.007019446697086096, + 0.048915475606918335, + 0.19594137370586395, + 0.01127713918685913, + 0.16320641338825226, + 0.0022493035066872835, + 0.0021556958090513945, + 0.0017159662675112486, + 0.0018570699030533433, + 0.003975669853389263, + 0.01724943146109581, + 0.0024091817904263735, + 0.015298952348530293, + 0.028571760281920433, + 0.012576070614159107, + 0.01403332781046629, + 0.010507393628358841, + 0.020762523636221886, + 0.08960743248462677, + 0.015023639425635338, + 0.0931805968284607, + 0.0013254748191684484, + 0.0010427541565150023, + 0.0009072861284948885, + 0.0006813418003730476, + 0.002496463479474187, + 0.0020036697387695312, + 0.0017214956460520625, + 0.0017705296631902456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 5 + ], + "selected_reward": -0.023300418630242348 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with a fish-eye lens effect causing distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a clown wearing a loose, faded costume with mixed vertical stripes and polka dots visible in the center of a spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the clown exhibit an exaggerated smile, thick cracked white makeup, deep-set black eye sockets, and a shaggy bright orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an empty circus tent with rows of empty wooden benches, massive bent tent poles, and a sawdust-strewn floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42895081639289856 + ], + "opposite_logprob_tokens": [ + -1.0539507865905762 + ], + "logprob_reward": -0.42895081639289856, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42895081639289856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02073751762509346, + 0.013551468029618263, + 0.04248056560754776, + 0.014713169075548649, + 0.014791158959269524, + 0.12873806059360504, + 0.5899641513824463, + 0.12470752000808716, + 0.03819732740521431, + 0.012119043618440628 + ], + "credit_map": [ + 0.0032240201253443956, + 0.003129919059574604, + 0.0027344990521669388, + 0.002932970644906163, + 0.002037644851952791, + 0.0032095061615109444, + 0.0013597613433375955, + 0.002109197434037924, + 0.0020804840605705976, + 0.0022879205644130707, + 0.0022071590647101402, + 0.0014542166609317064, + 0.0015462146839126945, + 0.001760203274898231, + 0.001136788516305387, + 0.0010784828336909413, + 0.006279794033616781, + 0.005114715546369553, + 0.00535902613773942, + 0.003348131664097309, + 0.00590362586081028, + 0.007371475454419851, + 0.004594063851982355, + 0.004509736318141222, + 0.0020082485862076283, + 0.0021028807386755943, + 0.001974937040358782, + 0.0021748007275164127, + 0.001547034946270287, + 0.002262459835037589, + 0.0010189668973907828, + 0.00162384076975286, + 0.0026087507139891386, + 0.0017631524242460728, + 0.0014521526172757149, + 0.001237367745488882, + 0.0019705411978065968, + 0.0023022876121103764, + 0.001156821264885366, + 0.00230008689686656, + 0.011484033428132534, + 0.015685269609093666, + 0.00810512900352478, + 0.01200677640736103, + 0.018851131200790405, + 0.022370651364326477, + 0.009190890938043594, + 0.03104417771100998, + 0.023359406739473343, + 0.008245225064456463, + 0.014521509408950806, + 0.010248713195323944, + 0.06162704899907112, + 0.2470029592514038, + 0.014970368705689907, + 0.20998895168304443, + 0.006780113093554974, + 0.0055391076020896435, + 0.003671279875561595, + 0.004255702719092369, + 0.011698008514940739, + 0.04791204631328583, + 0.00542037608101964, + 0.0394308976829052, + 0.0037288181483745575, + 0.0016798444557935, + 0.0020610725041478872, + 0.0012734660413116217, + 0.0024747916031628847, + 0.012333355844020844, + 0.0019931066781282425, + 0.012652876786887646, + 0.001826273975893855, + 0.0011095863301306963, + 0.0012041322188451886, + 0.0008104870212264359, + 0.001356859807856381, + 0.00206992169842124, + 0.0013656170340254903, + 0.0023761671036481857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.42895081639289856 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_12/index_tensor([9555, 9555])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_12/index_tensor([9555, 9555])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e82d046989cc6045372b60b8ff5264c38df8128c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_12/index_tensor([9555, 9555])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, slightly low-angled, extreme close-up depicts the face of a massive, weathered stone statue, carved directly into a sheer rock face. Warm, golden morning light streaks across the surface, creating a dramatic play of light and shadow, emphasizing the rough, pitted texture of the ancient gray granite and the deep, wrinkled cracks that crisscross its surface. The frame is dominated by a massive, tightly closed eye. The heavy stone eyelid begins to lift, an almost imperceptible, grinding movement that speaks of immense weight and ancient power. As the lid's lower edge separates from the cheek, a cloud of fine dust and tiny pebbles is expelled, dripping steadily along the richly contoured, weathered surface of the stone cheek like a gritty tear, each particle gleaming in the dawn sunlight. The slow, relentless movement reveals a profound darkness beneath the lid, creating a solemn and awe-inspiring atmosphere of a mysterious power awakening from its slumber.", + "seed": 25195, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_12/index_9555_seed_25195_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04394114427268505, + "vlm_logprob": -0.04394114427268505, + "vlm_yes_no_margin": 3.524999976158142, + "vlm_token_credit": -0.04394114427268505, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, weathered stone statue carved into a sheer rock face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue made of ancient gray granite with a rough, pitted texture and deep wrinkled cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frame dominated by a massive, tightly closed eye that begins to lift slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"As the eyelid lifts, is a cloud of fine dust and tiny pebbles expelled that drips along the stone cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, creating a dramatic play of light and shadow on the stone surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02583436854183674 + ], + "opposite_logprob_tokens": [ + -3.900834321975708 + ], + "logprob_reward": -0.02583436854183674, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02583436854183674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3638901710510254, + 0.08147234469652176, + 0.0847875326871872, + 0.04600777104496956, + 0.07675991207361221, + 0.13840550184249878, + 0.05037744343280792, + 0.023170314729213715, + 0.11317454278469086, + 0.021954456344246864 + ], + "credit_map": [ + 0.0305921733379364, + 0.04287266731262207, + 0.05391860753297806, + 0.09965137392282486, + 0.03551825135946274, + 0.02808871865272522, + 0.03195533528923988, + 0.041293032467365265, + 0.0073567102663218975, + 0.011998191475868225, + 0.012365798465907574, + 0.014604914002120495, + 0.009191375225782394, + 0.006866361014544964, + 0.007607687264680862, + 0.011481312103569508, + 0.009009596891701221, + 0.011304768733680248, + 0.008692036382853985, + 0.012232882902026176, + 0.008896857500076294, + 0.008940279483795166, + 0.01012004166841507, + 0.015591070987284184, + 0.004545075353235006, + 0.004985044244676828, + 0.004857142921537161, + 0.00803459994494915, + 0.004389454144984484, + 0.00538055133074522, + 0.006147316657006741, + 0.007668584585189819, + 0.007128710858523846, + 0.011078973300755024, + 0.00885025691241026, + 0.016456197947263718, + 0.009741401299834251, + 0.008068043738603592, + 0.00871436856687069, + 0.006721958518028259, + 0.013414395973086357, + 0.018984006717801094, + 0.017004895955324173, + 0.028229961171746254, + 0.016620315611362457, + 0.014368581585586071, + 0.017523743212223053, + 0.012259591370821, + 0.0033934132661670446, + 0.006554473657160997, + 0.005665536038577557, + 0.015087682753801346, + 0.004915193654596806, + 0.0038554968778043985, + 0.006418507546186447, + 0.004487138707190752, + 0.0022630314342677593, + 0.002552565187215805, + 0.0026786313392221928, + 0.0034854449331760406, + 0.002799043897539377, + 0.002493241336196661, + 0.0036354796029627323, + 0.0032628781627863646, + 0.008065025322139263, + 0.013133538886904716, + 0.013051006942987442, + 0.0227337758988142, + 0.016524070873856544, + 0.010239500552415848, + 0.019196201115846634, + 0.010231416672468185, + 0.0020995463710278273, + 0.0028020855970680714, + 0.0024260745849460363, + 0.004417558666318655, + 0.00267553748562932, + 0.0019625378772616386, + 0.0026560171972960234, + 0.002915098564699292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.02583436854183674 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, weathered stone statue carved into a sheer rock face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue made of ancient gray granite with a rough, pitted texture and deep wrinkled cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frame dominated by a massive, tightly closed eye that begins to lift slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"As the eyelid lifts, is a cloud of fine dust and tiny pebbles expelled that drips along the stone cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, creating a dramatic play of light and shadow on the stone surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023313229903578758 + ], + "opposite_logprob_tokens": [ + -3.773313283920288 + ], + "logprob_reward": -0.023313229903578758, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023313229903578758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2532070279121399, + 0.07564819604158401, + 0.08768461644649506, + 0.07638002932071686, + 0.08110649883747101, + 0.18651996552944183, + 0.07953102886676788, + 0.03792287036776543, + 0.08966042101383209, + 0.03233928605914116 + ], + "credit_map": [ + 0.02306278981268406, + 0.03173767775297165, + 0.03914369270205498, + 0.05521245300769806, + 0.02319774031639099, + 0.02717098966240883, + 0.022876916453242302, + 0.030804777517914772, + 0.005205205641686916, + 0.00896427035331726, + 0.013571958988904953, + 0.017459044232964516, + 0.006838548928499222, + 0.0060746693052351475, + 0.010705668479204178, + 0.006828836165368557, + 0.008243577554821968, + 0.01088379230350256, + 0.007880829274654388, + 0.013942833058536053, + 0.01011855248361826, + 0.00902636256068945, + 0.012223757803440094, + 0.015364918857812881, + 0.008122150786221027, + 0.007212188094854355, + 0.008545184507966042, + 0.013071713037788868, + 0.007248524576425552, + 0.01009721215814352, + 0.009677043184638023, + 0.012406017631292343, + 0.00862329825758934, + 0.01268568355590105, + 0.008903419598937035, + 0.01351475715637207, + 0.011791468597948551, + 0.007713868282735348, + 0.008692584931850433, + 0.009181425906717777, + 0.020368380472064018, + 0.02517475187778473, + 0.02433566376566887, + 0.03658299520611763, + 0.022247949615120888, + 0.017958620563149452, + 0.023143457248806953, + 0.01670815423130989, + 0.007247790694236755, + 0.00990391056984663, + 0.010153534822165966, + 0.01924481801688671, + 0.007341408636420965, + 0.006293816491961479, + 0.00996024627238512, + 0.009385505691170692, + 0.0041433474980294704, + 0.004346823785454035, + 0.004875151440501213, + 0.0061541772447526455, + 0.0033744655083864927, + 0.003387963864952326, + 0.005112491548061371, + 0.006528456229716539, + 0.006272465456277132, + 0.009113219566643238, + 0.008971462026238441, + 0.02342524565756321, + 0.010923628695309162, + 0.0074326093308627605, + 0.014643093571066856, + 0.008878706954419613, + 0.0031345731113106012, + 0.005313991568982601, + 0.0034421661403030157, + 0.005168300122022629, + 0.004478129092603922, + 0.002931360388174653, + 0.003835878800600767, + 0.004034890327602625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.023313229903578758 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, weathered stone statue carved into a sheer rock face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue made of ancient gray granite with a rough, pitted texture and deep wrinkled cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frame dominated by a massive, tightly closed eye that begins to lift slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"As the eyelid lifts, is a cloud of fine dust and tiny pebbles expelled that drips along the stone cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, creating a dramatic play of light and shadow on the stone surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371492773294449 + ], + "opposite_logprob_tokens": [ + -3.408715009689331 + ], + "logprob_reward": -0.03371492773294449, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371492773294449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3074694871902466, + 0.06753779202699661, + 0.04309602826833725, + 0.06953902542591095, + 0.09872899949550629, + 0.05873529240489006, + 0.03589087724685669, + 0.12830011546611786, + 0.1174129843711853, + 0.07328944653272629 + ], + "credit_map": [ + 0.029707856476306915, + 0.04303237423300743, + 0.0357159785926342, + 0.05095910280942917, + 0.03492264077067375, + 0.030211519449949265, + 0.03819318860769272, + 0.04472682252526283, + 0.00515039311721921, + 0.009199646301567554, + 0.006218522321432829, + 0.013466211967170238, + 0.008375461213290691, + 0.0070892940275371075, + 0.008567480370402336, + 0.00947077851742506, + 0.004452247638255358, + 0.004163588862866163, + 0.00502318749204278, + 0.004983820486813784, + 0.004842760972678661, + 0.006330183707177639, + 0.005600512959063053, + 0.007699726615101099, + 0.0077804746106266975, + 0.010137555189430714, + 0.00794197991490364, + 0.009968207217752934, + 0.009624849073588848, + 0.007081532385200262, + 0.008541825227439404, + 0.008462600409984589, + 0.008521846495568752, + 0.013361859135329723, + 0.010952871292829514, + 0.016812311485409737, + 0.01236066222190857, + 0.01163146086037159, + 0.012867609038949013, + 0.012220388278365135, + 0.005962484981864691, + 0.006647780537605286, + 0.006137315183877945, + 0.008250043727457523, + 0.006400043144822121, + 0.006452915724366903, + 0.008316894061863422, + 0.01056781504303217, + 0.003622959600761533, + 0.004435562063008547, + 0.00394150335341692, + 0.006535680033266544, + 0.0042616231366992, + 0.0036965745966881514, + 0.004479134920984507, + 0.00491784093901515, + 0.011514185927808285, + 0.016465209424495697, + 0.014377294108271599, + 0.0272656437009573, + 0.014268649742007256, + 0.011092453263700008, + 0.016217738389968872, + 0.017098939046263695, + 0.0071657984517514706, + 0.013228416442871094, + 0.013730499893426895, + 0.028750773519277573, + 0.014079967513680458, + 0.010605269111692905, + 0.017297789454460144, + 0.0125544723123312, + 0.0063188825733959675, + 0.008667788468301296, + 0.009486952796578407, + 0.017032425850629807, + 0.008104542270302773, + 0.0061960360035300255, + 0.008711442351341248, + 0.008771384134888649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.03371492773294449 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, weathered stone statue carved into a sheer rock face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue made of ancient gray granite with a rough, pitted texture and deep wrinkled cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frame dominated by a massive, tightly closed eye that begins to lift slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"As the eyelid lifts, is a cloud of fine dust and tiny pebbles expelled that drips along the stone cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, creating a dramatic play of light and shadow on the stone surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705907225608826 + ], + "opposite_logprob_tokens": [ + -2.127058982849121 + ], + "logprob_reward": -0.12705907225608826, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12705907225608826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47672080993652344, + 0.06882821023464203, + 0.05363529548048973, + 0.029068684205412865, + 0.06142415106296539, + 0.06832624971866608, + 0.06098110228776932, + 0.03663243353366852, + 0.09084605425596237, + 0.05353698506951332 + ], + "credit_map": [ + 0.030639005824923515, + 0.0550018772482872, + 0.09121934324502945, + 0.14342601597309113, + 0.042386628687381744, + 0.0328265056014061, + 0.037357181310653687, + 0.043864283710718155, + 0.004646523855626583, + 0.01049143634736538, + 0.011400444433093071, + 0.013641081750392914, + 0.006510433275252581, + 0.005703386850655079, + 0.009559721685945988, + 0.0068751852959394455, + 0.004404406063258648, + 0.005899351555854082, + 0.005368239711970091, + 0.006651036906987429, + 0.007111797109246254, + 0.007749543059617281, + 0.0071299090050160885, + 0.00932101346552372, + 0.002879407722502947, + 0.004046095069497824, + 0.0031571388244628906, + 0.003540778998285532, + 0.003897671354934573, + 0.003547149943187833, + 0.00432404363527894, + 0.0036763984244316816, + 0.005745249334722757, + 0.00862609501928091, + 0.005931123159825802, + 0.009815865196287632, + 0.008420558646321297, + 0.0056026410311460495, + 0.008271221071481705, + 0.009011396206915379, + 0.006505630910396576, + 0.008943038992583752, + 0.00797175895422697, + 0.014982854016125202, + 0.007655665278434753, + 0.006637056358158588, + 0.00733334431424737, + 0.008296899497509003, + 0.005073091480880976, + 0.008773595094680786, + 0.006245959084481001, + 0.014933429658412933, + 0.006116979755461216, + 0.004975303076207638, + 0.008059751242399216, + 0.006802990101277828, + 0.004395400173962116, + 0.004066563211381435, + 0.004159342031925917, + 0.006933193653821945, + 0.0033583627082407475, + 0.0031663174740970135, + 0.0052407183684408665, + 0.005312537774443626, + 0.006429052446037531, + 0.009717447683215141, + 0.01179084088653326, + 0.017935724928975105, + 0.011381048709154129, + 0.008889510296285152, + 0.01426983717828989, + 0.010432594455778599, + 0.00526769645512104, + 0.006841003429144621, + 0.007013323251157999, + 0.010809516534209251, + 0.0060350410640239716, + 0.00494473148137331, + 0.005878771655261517, + 0.006746900267899036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.12705907225608826 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, weathered stone statue carved into a sheer rock face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue made of ancient gray granite with a rough, pitted texture and deep wrinkled cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frame dominated by a massive, tightly closed eye that begins to lift slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"As the eyelid lifts, is a cloud of fine dust and tiny pebbles expelled that drips along the stone cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, creating a dramatic play of light and shadow on the stone surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784122928977013 + ], + "opposite_logprob_tokens": [ + -4.63478422164917 + ], + "logprob_reward": -0.009784122928977013, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784122928977013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7004906535148621, + 0.049922578036785126, + 0.037867456674575806, + 0.02260061912238598, + 0.03433832898736, + 0.031958721578121185, + 0.04442564770579338, + 0.0242011658847332, + 0.019077632576227188, + 0.03511729836463928 + ], + "credit_map": [ + 0.04742734879255295, + 0.05732012167572975, + 0.14094278216362, + 0.24628663063049316, + 0.06257525831460953, + 0.04641784727573395, + 0.04910120740532875, + 0.050419386476278305, + 0.004773889202624559, + 0.006693239789456129, + 0.006189335137605667, + 0.006295635364949703, + 0.005748216528445482, + 0.006428995635360479, + 0.0055887093767523766, + 0.008204549551010132, + 0.0029855992179363966, + 0.0037645543925464153, + 0.004435023292899132, + 0.004671065602451563, + 0.005190287251025438, + 0.0038163168355822563, + 0.00582517497241497, + 0.00717942975461483, + 0.0018571879481896758, + 0.003548330394551158, + 0.0023778630420565605, + 0.0036135977134108543, + 0.00327674625441432, + 0.0020141522400081158, + 0.0032387059181928635, + 0.0026740331668406725, + 0.0028576599434018135, + 0.0040840692818164825, + 0.004277057480067015, + 0.004335924983024597, + 0.004186698235571384, + 0.004888346418738365, + 0.004173568915575743, + 0.0055350037291646, + 0.0027989002410322428, + 0.0031288170721381903, + 0.0037144527304917574, + 0.004527367185801268, + 0.0035995354410260916, + 0.0038775200955569744, + 0.0037071709521114826, + 0.006604958325624466, + 0.003447952214628458, + 0.005435210652649403, + 0.004933007061481476, + 0.012243054807186127, + 0.004354214761406183, + 0.003606378333643079, + 0.005592621862888336, + 0.004813209641724825, + 0.002123582875356078, + 0.002647367073222995, + 0.00261109322309494, + 0.004701470490545034, + 0.0025159958750009537, + 0.0029357033781707287, + 0.00361772533506155, + 0.003048224840313196, + 0.0015471171354874969, + 0.0019016691949218512, + 0.0021862229332327843, + 0.00259422417730093, + 0.0022757253609597683, + 0.0024385314900428057, + 0.0031029684469103813, + 0.0030311704613268375, + 0.004009525291621685, + 0.0039791627787053585, + 0.004417350050061941, + 0.006933014839887619, + 0.0037736317608505487, + 0.0033812925685197115, + 0.0038612380158156157, + 0.004762080032378435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.009784122928977013 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly low-angled, extreme close-up depicts the face of a massive, weathered stone statue, carved directly into a sheer rock face. Warm, golden morning light streaks across the surface, creating a dramatic play of light and shadow, emphasizing the rough, pitted texture of the ancient gray granite and the deep, wrinkled cracks that crisscross its surface. The frame is dominated by a massive, tightly closed eye. The heavy stone eyelid begins to lift, an almost imperceptible, grinding movement that speaks of immense weight and ancient power. As the lid's lower edge separates from the cheek, a cloud of fine dust and tiny pebbles is expelled, dripping steadily along the richly contoured, weathered surface of the stone cheek like a gritty tear, each particle gleaming in the dawn sunlight. The slow, relentless movement reveals a profound darkness beneath the lid, creating a solemn and awe-inspiring atmosphere of a mysterious power awakening from its slumber.", + "seed": 25196, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_12/index_9555_seed_25196_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.022316603874787688, + "vlm_logprob": -0.022316603874787688, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.022316603874787688, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, weathered stone statue carved into a sheer rock face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue made of ancient gray granite with a rough, pitted texture and deep wrinkled cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frame dominated by a massive, tightly closed eye that begins to lift slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"As the eyelid lifts, is a cloud of fine dust and tiny pebbles expelled that drips along the stone cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, creating a dramatic play of light and shadow on the stone surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010147410444915295 + ], + "opposite_logprob_tokens": [ + -4.885147571563721 + ], + "logprob_reward": -0.010147410444915295, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.010147410444915295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05836196616292, + 0.06273707002401352, + 0.040502533316612244, + 0.22351312637329102, + 0.03246547654271126, + 0.37453100085258484, + 0.03752906993031502, + 0.06753648817539215, + 0.07569091767072678, + 0.02713240310549736 + ], + "credit_map": [ + 0.0089554563164711, + 0.00537317106500268, + 0.012111157178878784, + 0.005266956519335508, + 0.009083171375095844, + 0.004167856182903051, + 0.008399485610425472, + 0.005004708655178547, + 0.0075817410834133625, + 0.019931476563215256, + 0.006006497889757156, + 0.009662071242928505, + 0.007869839668273926, + 0.003415721468627453, + 0.00504349684342742, + 0.003226224100217223, + 0.0043111564591526985, + 0.006115010008215904, + 0.005910037085413933, + 0.005854737013578415, + 0.005391169339418411, + 0.004043914843350649, + 0.005141958594322205, + 0.0037345520686358213, + 0.020694520324468613, + 0.048725664615631104, + 0.07406898587942123, + 0.02980552241206169, + 0.010826360434293747, + 0.015208710916340351, + 0.011098370887339115, + 0.01308497879654169, + 0.006065633613616228, + 0.0047492519952356815, + 0.003988281823694706, + 0.006030275020748377, + 0.002902095438912511, + 0.002355600008741021, + 0.002903008833527565, + 0.00347132608294487, + 0.06377916038036346, + 0.04468711465597153, + 0.05395065248012543, + 0.08971209079027176, + 0.04217724874615669, + 0.020670274272561073, + 0.038514599204063416, + 0.021039849147200584, + 0.006689788773655891, + 0.003960356581956148, + 0.009053781628608704, + 0.004759862553328276, + 0.002015122212469578, + 0.0039120810106396675, + 0.004250232595950365, + 0.002887845505028963, + 0.01087906677275896, + 0.006231823470443487, + 0.010916121304035187, + 0.011806740425527096, + 0.004280238877981901, + 0.005852525122463703, + 0.010320884175598621, + 0.007249090354889631, + 0.023613551631569862, + 0.0039118011482059956, + 0.02847660332918167, + 0.0057307519018650055, + 0.001879549352452159, + 0.004452432040125132, + 0.0035376993473619223, + 0.004088522866368294, + 0.0037894845008850098, + 0.0021735995542258024, + 0.003567719366401434, + 0.004174238070845604, + 0.002392381662502885, + 0.003995159640908241, + 0.0026302156038582325, + 0.00440960330888629 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 8 + ], + "selected_reward": -0.010147410444915295 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, weathered stone statue carved into a sheer rock face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue made of ancient gray granite with a rough, pitted texture and deep wrinkled cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frame dominated by a massive, tightly closed eye that begins to lift slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"As the eyelid lifts, is a cloud of fine dust and tiny pebbles expelled that drips along the stone cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, creating a dramatic play of light and shadow on the stone surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02637499012053013 + ], + "opposite_logprob_tokens": [ + -3.6513750553131104 + ], + "logprob_reward": -0.02637499012053013, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02637499012053013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039211075752973557, + 0.0748777836561203, + 0.05852542072534561, + 0.4250792860984802, + 0.02569732256233692, + 0.23150275647640228, + 0.03963341563940048, + 0.05380234494805336, + 0.033960457891225815, + 0.017710134387016296 + ], + "credit_map": [ + 0.005614167544990778, + 0.0032645908650010824, + 0.010762863792479038, + 0.00408829003572464, + 0.004258552100509405, + 0.0022287534084171057, + 0.004380489699542522, + 0.004613369703292847, + 0.01048431545495987, + 0.017013175413012505, + 0.007113149389624596, + 0.011649779975414276, + 0.008096008561551571, + 0.0049254270270466805, + 0.007271634414792061, + 0.008324295282363892, + 0.007029808592051268, + 0.00843112450093031, + 0.00840973760932684, + 0.008532494306564331, + 0.008335193619132042, + 0.005339942406862974, + 0.006973484996706247, + 0.005473635159432888, + 0.04054705798625946, + 0.08994212001562119, + 0.141536682844162, + 0.0550401397049427, + 0.020680047571659088, + 0.029120588675141335, + 0.023346994072198868, + 0.024865666404366493, + 0.0051297880709171295, + 0.003761395812034607, + 0.003234974341467023, + 0.0031534279696643353, + 0.002511192113161087, + 0.0018100960878655314, + 0.0025178154464811087, + 0.003578632138669491, + 0.039851944893598557, + 0.02790110558271408, + 0.03377585485577583, + 0.05375037342309952, + 0.02585987001657486, + 0.013126119039952755, + 0.023465877398848534, + 0.013771622441709042, + 0.006314351689070463, + 0.005141906905919313, + 0.00814809463918209, + 0.005295730661600828, + 0.0035802554339170456, + 0.004333272110670805, + 0.003827163949608803, + 0.002992636989802122, + 0.011828298680484295, + 0.003790148301050067, + 0.014291317202150822, + 0.009229100309312344, + 0.002258031629025936, + 0.0029168156906962395, + 0.004532113671302795, + 0.004956520162522793, + 0.008542552590370178, + 0.0038114571943879128, + 0.010039236396551132, + 0.003135395236313343, + 0.0014432057505473495, + 0.0021762740798294544, + 0.003046703524887562, + 0.0017656326526775956, + 0.0024014883674681187, + 0.002718480536714196, + 0.0019420166499912739, + 0.0032803439535200596, + 0.0014441517414525151, + 0.0015794545179232955, + 0.0024503853637725115, + 0.0018938126740977168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 1 + ], + "selected_reward": -0.02637499012053013 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, weathered stone statue carved into a sheer rock face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue made of ancient gray granite with a rough, pitted texture and deep wrinkled cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frame dominated by a massive, tightly closed eye that begins to lift slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"As the eyelid lifts, is a cloud of fine dust and tiny pebbles expelled that drips along the stone cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, creating a dramatic play of light and shadow on the stone surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002203776268288493 + ], + "opposite_logprob_tokens": [ + -6.127203941345215 + ], + "logprob_reward": -0.002203776268288493, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002203776268288493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09428096562623978, + 0.15950320661067963, + 0.11993733048439026, + 0.09370780736207962, + 0.046266213059425354, + 0.2399284392595291, + 0.04155828431248665, + 0.09298315644264221, + 0.07698998600244522, + 0.03484465181827545 + ], + "credit_map": [ + 0.0144187081605196, + 0.010103507898747921, + 0.022710369899868965, + 0.010049157775938511, + 0.012573206797242165, + 0.005482138600200415, + 0.010351575911045074, + 0.008592303842306137, + 0.02220344729721546, + 0.05891691520810127, + 0.006510991603136063, + 0.022708620876073837, + 0.018320156261324883, + 0.006245992612093687, + 0.010475932620465755, + 0.014121158048510551, + 0.014202023856341839, + 0.01687166839838028, + 0.02069583162665367, + 0.01854858547449112, + 0.012419137172400951, + 0.009546027518808842, + 0.015749633312225342, + 0.011904417537152767, + 0.009426604025065899, + 0.017644500359892845, + 0.028606772422790527, + 0.01224387064576149, + 0.005832274444401264, + 0.006269392557442188, + 0.007459711283445358, + 0.006224684417247772, + 0.007121321279555559, + 0.004799420014023781, + 0.007965420372784138, + 0.007280781865119934, + 0.002923044376075268, + 0.004423384089022875, + 0.004673205781728029, + 0.007079636678099632, + 0.04139624163508415, + 0.02918951027095318, + 0.033424876630306244, + 0.05745464190840721, + 0.027498550713062286, + 0.012570519000291824, + 0.02462035045027733, + 0.013773730956017971, + 0.008169658482074738, + 0.00629243953153491, + 0.008491013199090958, + 0.005560848396271467, + 0.0030141142196953297, + 0.0026775163132697344, + 0.003823457518592477, + 0.0035292357206344604, + 0.01626054383814335, + 0.011632218956947327, + 0.016605248674750328, + 0.013249185867607594, + 0.006463984958827496, + 0.007883206009864807, + 0.014139923267066479, + 0.006748850457370281, + 0.020666901022195816, + 0.005930235609412193, + 0.021071024239063263, + 0.006263882387429476, + 0.004134701564908028, + 0.005033857189118862, + 0.007013088092207909, + 0.0068762945011258125, + 0.005209042225033045, + 0.004136866424232721, + 0.003835066454485059, + 0.0069684823974967, + 0.0036402044352144003, + 0.003979082684963942, + 0.0034381479490548372, + 0.0036377599462866783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 2 + ], + "selected_reward": -0.002203776268288493 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, weathered stone statue carved into a sheer rock face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue made of ancient gray granite with a rough, pitted texture and deep wrinkled cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frame dominated by a massive, tightly closed eye that begins to lift slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"As the eyelid lifts, is a cloud of fine dust and tiny pebbles expelled that drips along the stone cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, creating a dramatic play of light and shadow on the stone surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003437727689743 + ], + "opposite_logprob_tokens": [ + -2.6950342655181885 + ], + "logprob_reward": -0.07003437727689743, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003437727689743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03879712149500847, + 0.04709922894835472, + 0.09241297096014023, + 0.23712849617004395, + 0.06865472346544266, + 0.18762533366680145, + 0.042217087000608444, + 0.06290694326162338, + 0.18223272264003754, + 0.040925368666648865 + ], + "credit_map": [ + 0.005314158275723457, + 0.003253808245062828, + 0.009792085736989975, + 0.004597744904458523, + 0.004649298265576363, + 0.002506933407858014, + 0.005677941255271435, + 0.003005153266713023, + 0.005737435072660446, + 0.013800490647554398, + 0.003315167734399438, + 0.007846158929169178, + 0.005169565789401531, + 0.0020121184643357992, + 0.004089712165296078, + 0.00512858247384429, + 0.010458240285515785, + 0.01297386921942234, + 0.015451764687895775, + 0.012091400101780891, + 0.013241362757980824, + 0.006593671627342701, + 0.01178784016519785, + 0.009814828634262085, + 0.020617235451936722, + 0.053655955940485, + 0.07649514079093933, + 0.03342236578464508, + 0.01165033783763647, + 0.015981242060661316, + 0.01163236889988184, + 0.013673844747245312, + 0.012261536903679371, + 0.013513846322894096, + 0.00848182663321495, + 0.011583867482841015, + 0.007445411290973425, + 0.004728719592094421, + 0.005461594555526972, + 0.0051779188215732574, + 0.032157786190509796, + 0.021945856511592865, + 0.02866535075008869, + 0.04205366224050522, + 0.020693957805633545, + 0.011450810357928276, + 0.018613073974847794, + 0.012044835835695267, + 0.007487429305911064, + 0.004561327863484621, + 0.010000430047512054, + 0.005789902061223984, + 0.0026806641835719347, + 0.004339974373579025, + 0.004224935080856085, + 0.0031324231531471014, + 0.010094969533383846, + 0.005750374402850866, + 0.00925277266651392, + 0.010365123860538006, + 0.0045145126059651375, + 0.0062287189066410065, + 0.00884198397397995, + 0.007858489640057087, + 0.058024995028972626, + 0.011534641496837139, + 0.06868410855531693, + 0.011401467956602573, + 0.005533407907932997, + 0.0075475177727639675, + 0.010173290967941284, + 0.0093332938849926, + 0.00593842938542366, + 0.004844601731747389, + 0.004387300927191973, + 0.004864135291427374, + 0.0034670718014240265, + 0.004456883762031794, + 0.003528980305418372, + 0.009437964297831059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 8 + ], + "selected_reward": -0.07003437727689743 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, weathered stone statue carved into a sheer rock face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue made of ancient gray granite with a rough, pitted texture and deep wrinkled cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frame dominated by a massive, tightly closed eye that begins to lift slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"As the eyelid lifts, is a cloud of fine dust and tiny pebbles expelled that drips along the stone cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, creating a dramatic play of light and shadow on the stone surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028224652633070946 + ], + "opposite_logprob_tokens": [ + -5.877822399139404 + ], + "logprob_reward": -0.0028224652633070946, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028224652633070946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09466684609651566, + 0.030411964282393456, + 0.021311407908797264, + 0.2185930460691452, + 0.05212658271193504, + 0.31607434153556824, + 0.03468260541558266, + 0.05018847808241844, + 0.11885003000497818, + 0.06309474259614944 + ], + "credit_map": [ + 0.01495407149195671, + 0.009451543912291527, + 0.026983510702848434, + 0.010653309524059296, + 0.01159185916185379, + 0.00588958477601409, + 0.006712203845381737, + 0.008430752903223038, + 0.003639147849753499, + 0.007145638577640057, + 0.0028504456859081984, + 0.0041115847416222095, + 0.0030832532793283463, + 0.0021101655438542366, + 0.002991188084706664, + 0.0044805374927818775, + 0.0028512058779597282, + 0.0036654833238571882, + 0.002060254104435444, + 0.0029422699008136988, + 0.0026373881846666336, + 0.0022720731794834137, + 0.0028564443346112967, + 0.002026286441832781, + 0.01682501845061779, + 0.05239411070942879, + 0.0677216649055481, + 0.03387394919991493, + 0.012553781270980835, + 0.014594321139156818, + 0.009883313439786434, + 0.010746863670647144, + 0.008399446494877338, + 0.007958757691085339, + 0.006579751614481211, + 0.00970783643424511, + 0.004746199119836092, + 0.0039366488344967365, + 0.004539016634225845, + 0.00625892449170351, + 0.05341188609600067, + 0.037055376917123795, + 0.04729563370347023, + 0.07262600213289261, + 0.03499539569020271, + 0.019218649715185165, + 0.03233439475297928, + 0.019136987626552582, + 0.007090503815561533, + 0.0038982636760920286, + 0.009132100269198418, + 0.004574629478156567, + 0.0019550020806491375, + 0.0026231599040329456, + 0.0024538009893149137, + 0.0029551424086093903, + 0.009734198451042175, + 0.004035281017422676, + 0.011676710098981857, + 0.00929301604628563, + 0.002652649534866214, + 0.004214416723698378, + 0.0035903146490454674, + 0.004991884343326092, + 0.03902902454137802, + 0.006355063058435917, + 0.04933539032936096, + 0.007631254848092794, + 0.002533405087888241, + 0.00493614049628377, + 0.0045087141916155815, + 0.00452103465795517, + 0.009227820672094822, + 0.004463486839085817, + 0.010281389579176903, + 0.009245829656720161, + 0.005047861952334642, + 0.007709070108830929, + 0.006606662645936012, + 0.010512619279325008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 8 + ], + "selected_reward": -0.0028224652633070946 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly low-angled, extreme close-up depicts the face of a massive, weathered stone statue, carved directly into a sheer rock face. Warm, golden morning light streaks across the surface, creating a dramatic play of light and shadow, emphasizing the rough, pitted texture of the ancient gray granite and the deep, wrinkled cracks that crisscross its surface. The frame is dominated by a massive, tightly closed eye. The heavy stone eyelid begins to lift, an almost imperceptible, grinding movement that speaks of immense weight and ancient power. As the lid's lower edge separates from the cheek, a cloud of fine dust and tiny pebbles is expelled, dripping steadily along the richly contoured, weathered surface of the stone cheek like a gritty tear, each particle gleaming in the dawn sunlight. The slow, relentless movement reveals a profound darkness beneath the lid, creating a solemn and awe-inspiring atmosphere of a mysterious power awakening from its slumber.", + "seed": 25197, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_12/index_9555_seed_25197_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.014496297296136617, + "vlm_logprob": -0.014496297296136617, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.014496297296136617, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, weathered stone statue carved into a sheer rock face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue made of ancient gray granite with a rough, pitted texture and deep wrinkled cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frame dominated by a massive, tightly closed eye that begins to lift slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"As the eyelid lifts, is a cloud of fine dust and tiny pebbles expelled that drips along the stone cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, creating a dramatic play of light and shadow on the stone surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01122127566486597 + ], + "opposite_logprob_tokens": [ + -5.136221408843994 + ], + "logprob_reward": -0.01122127566486597, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.01122127566486597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.613695502281189, + 0.015985816717147827, + 0.019531553611159325, + 0.06760319322347641, + 0.017144307494163513, + 0.12344741076231003, + 0.06818491220474243, + 0.021484404802322388, + 0.00535673787817359, + 0.04756619408726692 + ], + "credit_map": [ + 0.19487698376178741, + 0.0879533663392067, + 0.026755137369036674, + 0.045517537742853165, + 0.021656658500432968, + 0.04019200801849365, + 0.0434083417057991, + 0.15333543717861176, + 0.0024711673613637686, + 0.004153535235673189, + 0.0013896236196160316, + 0.0019917981699109077, + 0.002044427441433072, + 0.0015260097570717335, + 0.0012493127724155784, + 0.0011599416611716151, + 0.003986278548836708, + 0.00497607234865427, + 0.001007014769129455, + 0.002298190025612712, + 0.0026958908420056105, + 0.0015264638932421803, + 0.0014291142579168081, + 0.0016125275287777185, + 0.02432614006102085, + 0.01198635809123516, + 0.003172132186591625, + 0.003572326386347413, + 0.00246537197381258, + 0.003917081281542778, + 0.00420428067445755, + 0.013959492556750774, + 0.004984219092875719, + 0.002792170038446784, + 0.0007636688533239067, + 0.0018821669509634376, + 0.0010528353741392493, + 0.0009507918148301542, + 0.0020942622795701027, + 0.002624191576614976, + 0.029289133846759796, + 0.04859944060444832, + 0.005677275825291872, + 0.008542764000594616, + 0.006117288023233414, + 0.004876673687249422, + 0.004956287331879139, + 0.015388536266982555, + 0.021036086603999138, + 0.029656074941158295, + 0.0032577146776020527, + 0.005370719823986292, + 0.003184888046234846, + 0.0010810814565047622, + 0.0022195714991539717, + 0.0023787717800587416, + 0.002124304650351405, + 0.003470292314887047, + 0.002597960876300931, + 0.0041434625163674355, + 0.00223563308827579, + 0.0017589378403499722, + 0.0030473745428025723, + 0.0021064365282654762, + 0.0008819387876428664, + 0.0009903406025841832, + 0.0006732360925525427, + 0.0007771378732286394, + 0.0004722652374766767, + 0.0003846564795821905, + 0.0006983797065913677, + 0.00047878289478830993, + 0.00733261089771986, + 0.012270450592041016, + 0.003132403129711747, + 0.00651400163769722, + 0.0037032715044915676, + 0.0035918857902288437, + 0.00671899551525712, + 0.00430257199332118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.01122127566486597 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, weathered stone statue carved into a sheer rock face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue made of ancient gray granite with a rough, pitted texture and deep wrinkled cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frame dominated by a massive, tightly closed eye that begins to lift slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"As the eyelid lifts, is a cloud of fine dust and tiny pebbles expelled that drips along the stone cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, creating a dramatic play of light and shadow on the stone surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012556307949125767 + ], + "opposite_logprob_tokens": [ + -4.387556076049805 + ], + "logprob_reward": -0.012556307949125767, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012556307949125767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0685708224773407, + 0.028973577544093132, + 0.024151530116796494, + 0.03167301043868065, + 0.06105301156640053, + 0.4252661466598511, + 0.03295052796602249, + 0.013415620662271976, + 0.035397469997406006, + 0.2785482704639435 + ], + "credit_map": [ + 0.010081569664180279, + 0.01749921403825283, + 0.0055731297470629215, + 0.009444445371627808, + 0.010787825100123882, + 0.004081853665411472, + 0.005927224177867174, + 0.005175561644136906, + 0.0046864645555615425, + 0.007090471219271421, + 0.0019315163372084498, + 0.0035721305757761, + 0.004351629875600338, + 0.002553808270022273, + 0.0022446231450885534, + 0.0025429334491491318, + 0.004712461028248072, + 0.004503814037889242, + 0.001920871203765273, + 0.002671766560524702, + 0.0025849875528365374, + 0.0020125731825828552, + 0.001895033405162394, + 0.003850020933896303, + 0.010948956944048405, + 0.005265283398330212, + 0.001418689964339137, + 0.0018880626885220408, + 0.0011615592520684004, + 0.0018566657090559602, + 0.002087445231154561, + 0.007046346087008715, + 0.020143942907452583, + 0.01195173617452383, + 0.00237479736097157, + 0.0052754199132323265, + 0.003784895408898592, + 0.0032972083427011967, + 0.004713971167802811, + 0.009511039592325687, + 0.10340957343578339, + 0.17138229310512543, + 0.019465485587716103, + 0.02881530672311783, + 0.021237816661596298, + 0.01652579754590988, + 0.012624509632587433, + 0.05180538073182106, + 0.010574527084827423, + 0.013064791448414326, + 0.0012584851356223226, + 0.0025451614055782557, + 0.0014110348420217633, + 0.000877475889865309, + 0.0011073556961491704, + 0.0021116991993039846, + 0.002438818570226431, + 0.002847955096513033, + 0.0014647453790530562, + 0.0018660255009308457, + 0.0011096763191744685, + 0.0008101588464342058, + 0.0016365480842068791, + 0.0012416928075253963, + 0.005495935212820768, + 0.007151450030505657, + 0.003337758593261242, + 0.005533732008188963, + 0.00362762832082808, + 0.0019365512998774648, + 0.005940258502960205, + 0.002374157076701522, + 0.056410323828458786, + 0.09766586124897003, + 0.0175313763320446, + 0.03197475150227547, + 0.012705535627901554, + 0.011676529422402382, + 0.024799833074212074, + 0.025784064084291458 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.012556307949125767 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, weathered stone statue carved into a sheer rock face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue made of ancient gray granite with a rough, pitted texture and deep wrinkled cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frame dominated by a massive, tightly closed eye that begins to lift slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"As the eyelid lifts, is a cloud of fine dust and tiny pebbles expelled that drips along the stone cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, creating a dramatic play of light and shadow on the stone surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005956517532467842 + ], + "opposite_logprob_tokens": [ + -5.130956649780273 + ], + "logprob_reward": -0.005956517532467842, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005956517532467842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44438841938972473, + 0.01703750714659691, + 0.01992502436041832, + 0.015142099000513554, + 0.07494373619556427, + 0.04298356547951698, + 0.15016813576221466, + 0.04483971372246742, + 0.06454204022884369, + 0.12602970004081726 + ], + "credit_map": [ + 0.14157500863075256, + 0.0640210434794426, + 0.017819944769144058, + 0.03285254165530205, + 0.01642504706978798, + 0.029544295743107796, + 0.03156333789229393, + 0.1105872094631195, + 0.0033653320278972387, + 0.002455509500578046, + 0.001594739151187241, + 0.00196227990090847, + 0.0014894083142280579, + 0.001749225426465273, + 0.0014251660322770476, + 0.0029958460945636034, + 0.005109016317874193, + 0.00392352556809783, + 0.0016153083415701985, + 0.0016722691943868995, + 0.00186475133523345, + 0.0016714093508198857, + 0.0016739164711907506, + 0.002394826849922538, + 0.0037872723769396544, + 0.0020586138125509024, + 0.0011166189797222614, + 0.0015029885107651353, + 0.001259718555957079, + 0.0016059476183727384, + 0.0013169209705665708, + 0.0024940180592238903, + 0.025958511978387833, + 0.013820154592394829, + 0.003488681046292186, + 0.004811722319573164, + 0.004595622420310974, + 0.004082970786839724, + 0.004397592972964048, + 0.013788484036922455, + 0.0068108271807432175, + 0.010812658816576004, + 0.0029319648165255785, + 0.006027769763022661, + 0.004212724044919014, + 0.00259418785572052, + 0.00567134702578187, + 0.0039220815524458885, + 0.03659413382411003, + 0.017464399337768555, + 0.009220315143465996, + 0.011324248276650906, + 0.0075953006744384766, + 0.013289256952702999, + 0.012555507011711597, + 0.042124975472688675, + 0.00438112486153841, + 0.00762294651940465, + 0.005668341182172298, + 0.009249875321984291, + 0.004936134908348322, + 0.0029134219512343407, + 0.006028235424309969, + 0.004039631690829992, + 0.009582613594830036, + 0.013043136335909367, + 0.0046699158847332, + 0.011478692293167114, + 0.006826155818998814, + 0.00287007843144238, + 0.012060360983014107, + 0.004011082928627729, + 0.026824776083230972, + 0.04730027914047241, + 0.007136612664908171, + 0.011540460400283337, + 0.004815066233277321, + 0.004785646218806505, + 0.009372896514832973, + 0.014253969304263592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.005956517532467842 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, weathered stone statue carved into a sheer rock face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue made of ancient gray granite with a rough, pitted texture and deep wrinkled cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frame dominated by a massive, tightly closed eye that begins to lift slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"As the eyelid lifts, is a cloud of fine dust and tiny pebbles expelled that drips along the stone cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, creating a dramatic play of light and shadow on the stone surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381057932972908 + ], + "opposite_logprob_tokens": [ + -3.2881057262420654 + ], + "logprob_reward": -0.0381057932972908, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381057932972908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11008832603693008, + 0.02657795511186123, + 0.02357192151248455, + 0.05328047648072243, + 0.04187329113483429, + 0.23815111815929413, + 0.18807640671730042, + 0.08972794562578201, + 0.10437499731779099, + 0.12427759915590286 + ], + "credit_map": [ + 0.027041049674153328, + 0.020503487437963486, + 0.006967714987695217, + 0.01039019227027893, + 0.012132525444030762, + 0.0065987021662294865, + 0.007936901412904263, + 0.018517756834626198, + 0.005695089232176542, + 0.0046097282320261, + 0.0018574120476841927, + 0.00309916608966887, + 0.002634516218677163, + 0.0017895103665068746, + 0.002609716961160302, + 0.004282817244529724, + 0.004648696631193161, + 0.006041902583092451, + 0.0019195417407900095, + 0.002573843812569976, + 0.002602057298645377, + 0.0022632903419435024, + 0.0020319221075624228, + 0.0014906662981957197, + 0.01692306436598301, + 0.008103922009468079, + 0.0028689114842563868, + 0.003672427963465452, + 0.0026820397470146418, + 0.0038368734531104565, + 0.004000452347099781, + 0.011192785575985909, + 0.00856416579335928, + 0.005363930948078632, + 0.0025896946899592876, + 0.0047351461835205555, + 0.0041313786059618, + 0.0037725132424384356, + 0.006461439188569784, + 0.006255021318793297, + 0.0504278801381588, + 0.08556991070508957, + 0.0116255609318614, + 0.02034236490726471, + 0.014972821809351444, + 0.009670845232903957, + 0.015300750732421875, + 0.030240988358855247, + 0.051680415868759155, + 0.038802579045295715, + 0.008622086606919765, + 0.015867168083786964, + 0.010247541591525078, + 0.012390824034810066, + 0.014471338130533695, + 0.03599444776773453, + 0.011324114166200161, + 0.021049046888947487, + 0.009790993295609951, + 0.018494876101613045, + 0.00768651906400919, + 0.003971862606704235, + 0.010341531597077847, + 0.007069005165249109, + 0.01568084955215454, + 0.021150147542357445, + 0.007725001312792301, + 0.018040288239717484, + 0.010692366398870945, + 0.004840394482016563, + 0.01951390877366066, + 0.0067320396192371845, + 0.026745520532131195, + 0.047450561076402664, + 0.006946673151105642, + 0.010707719251513481, + 0.005522624123841524, + 0.005030745640397072, + 0.008260573260486126, + 0.013613175600767136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 9 + ], + "selected_reward": -0.0381057932972908 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, weathered stone statue carved into a sheer rock face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue made of ancient gray granite with a rough, pitted texture and deep wrinkled cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frame dominated by a massive, tightly closed eye that begins to lift slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"As the eyelid lifts, is a cloud of fine dust and tiny pebbles expelled that drips along the stone cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, creating a dramatic play of light and shadow on the stone surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004641592036932707 + ], + "opposite_logprob_tokens": [ + -5.379641532897949 + ], + "logprob_reward": -0.004641592036932707, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004641592036932707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0886339396238327, + 0.007665129378437996, + 0.010284238494932652, + 0.07158923894166946, + 0.04162231832742691, + 0.06043725833296776, + 0.07610537856817245, + 0.05484172701835632, + 0.05130501836538315, + 0.5375157594680786 + ], + "credit_map": [ + 0.019416309893131256, + 0.01914696767926216, + 0.005070084240287542, + 0.009779421612620354, + 0.010677178390324116, + 0.005289809312671423, + 0.005839399062097073, + 0.013414768502116203, + 0.001122611458413303, + 0.0019303930457681417, + 0.0006335704238153994, + 0.0010408954694867134, + 0.0009382821153849363, + 0.0007367409416474402, + 0.0007378474110737443, + 0.0005247886874713004, + 0.0029761875048279762, + 0.0017192559316754341, + 0.0008987724431790411, + 0.0007841609767638147, + 0.0006761124241165817, + 0.0007645783480256796, + 0.0006664874963462353, + 0.0017986837774515152, + 0.025669053196907043, + 0.012196855619549751, + 0.003341058501973748, + 0.003939561080187559, + 0.00236942688934505, + 0.004111698362976313, + 0.004578843247145414, + 0.015382740646600723, + 0.0145652424544096, + 0.008380177430808544, + 0.001704263617284596, + 0.0029317543376237154, + 0.002351714763790369, + 0.0021657480392605066, + 0.0024259730707854033, + 0.007097442168742418, + 0.015251320786774158, + 0.023562772199511528, + 0.0030367346480488777, + 0.004446194972842932, + 0.00314595689997077, + 0.002372664399445057, + 0.002240084810182452, + 0.006381527986377478, + 0.02326590195298195, + 0.03353022411465645, + 0.003630411811172962, + 0.005854177754372358, + 0.0035271968226879835, + 0.0010985995177179575, + 0.0023177885450422764, + 0.0028810780495405197, + 0.008714418858289719, + 0.014410625211894512, + 0.003781590610742569, + 0.009956077672541142, + 0.004383739084005356, + 0.0022355043329298496, + 0.00610682088881731, + 0.005252948962152004, + 0.006552955135703087, + 0.008685938082635403, + 0.004126892425119877, + 0.009831252507865429, + 0.005878719035536051, + 0.0027794225607067347, + 0.010134967043995857, + 0.0033148692455142736, + 0.11105266958475113, + 0.19500704109668732, + 0.03198739513754845, + 0.05762670934200287, + 0.022761382162570953, + 0.021762127056717873, + 0.04341818764805794, + 0.05390023812651634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.004641592036932707 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly low-angled, extreme close-up depicts the face of a massive, weathered stone statue, carved directly into a sheer rock face. Warm, golden morning light streaks across the surface, creating a dramatic play of light and shadow, emphasizing the rough, pitted texture of the ancient gray granite and the deep, wrinkled cracks that crisscross its surface. The frame is dominated by a massive, tightly closed eye. The heavy stone eyelid begins to lift, an almost imperceptible, grinding movement that speaks of immense weight and ancient power. As the lid's lower edge separates from the cheek, a cloud of fine dust and tiny pebbles is expelled, dripping steadily along the richly contoured, weathered surface of the stone cheek like a gritty tear, each particle gleaming in the dawn sunlight. The slow, relentless movement reveals a profound darkness beneath the lid, creating a solemn and awe-inspiring atmosphere of a mysterious power awakening from its slumber.", + "seed": 25198, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_12/index_9555_seed_25198_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.013841739762574435, + "vlm_logprob": -0.013841739762574435, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.013841739762574435, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, weathered stone statue carved into a sheer rock face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue made of ancient gray granite with a rough, pitted texture and deep wrinkled cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frame dominated by a massive, tightly closed eye that begins to lift slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"As the eyelid lifts, is a cloud of fine dust and tiny pebbles expelled that drips along the stone cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, creating a dramatic play of light and shadow on the stone surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007728555705398321 + ], + "opposite_logprob_tokens": [ + -5.632728576660156 + ], + "logprob_reward": -0.007728555705398321, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.007728555705398321, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47499698400497437, + 0.0801672413945198, + 0.0737665519118309, + 0.1733463704586029, + 0.039130520075559616, + 0.04711586609482765, + 0.029056018218398094, + 0.02633850648999214, + 0.0170906949788332, + 0.038991302251815796 + ], + "credit_map": [ + 0.05157974362373352, + 0.023657388985157013, + 0.1751096546649933, + 0.06618573516607285, + 0.015643389895558357, + 0.026126373559236526, + 0.0336153618991375, + 0.08307932317256927, + 0.010601984336972237, + 0.006688820663839579, + 0.019340278580784798, + 0.011379890143871307, + 0.005268361885100603, + 0.004674610681831837, + 0.009856162592768669, + 0.012357134371995926, + 0.008657936938107014, + 0.0036292527802288532, + 0.02490442432463169, + 0.008180355653166771, + 0.0026486278511583805, + 0.0038277714047580957, + 0.005409973673522472, + 0.016508208587765694, + 0.018516000360250473, + 0.026319453492760658, + 0.02896885573863983, + 0.035677261650562286, + 0.014159691520035267, + 0.008376683108508587, + 0.025330254808068275, + 0.015998177230358124, + 0.004545547533780336, + 0.005618961062282324, + 0.006507745012640953, + 0.007547428365796804, + 0.0033498797565698624, + 0.0024217618629336357, + 0.005674567073583603, + 0.00346462894231081, + 0.006245477590709925, + 0.007206210400909185, + 0.006398733705282211, + 0.006884782575070858, + 0.006040356121957302, + 0.0028466617222875357, + 0.00727046187967062, + 0.004223182797431946, + 0.004448309540748596, + 0.0029427313711494207, + 0.004779948852956295, + 0.0034784653689712286, + 0.002197477500885725, + 0.0024654921144247055, + 0.002766447374597192, + 0.00597714725881815, + 0.0038209559861570597, + 0.003648907644674182, + 0.003005316946655512, + 0.00420894380658865, + 0.0021836981177330017, + 0.0017782081849873066, + 0.004962919279932976, + 0.0027295572217553854, + 0.0036229677498340607, + 0.0018616882152855396, + 0.0018831094494089484, + 0.0021865512244403362, + 0.0014182506129145622, + 0.0011193761602044106, + 0.0030718143098056316, + 0.0019269366748631, + 0.006405853200703859, + 0.005652965046465397, + 0.0038157361559569836, + 0.006872260943055153, + 0.003387156408280134, + 0.003325254423543811, + 0.006125821731984615, + 0.00340625480748713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.007728555705398321 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, weathered stone statue carved into a sheer rock face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue made of ancient gray granite with a rough, pitted texture and deep wrinkled cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frame dominated by a massive, tightly closed eye that begins to lift slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"As the eyelid lifts, is a cloud of fine dust and tiny pebbles expelled that drips along the stone cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, creating a dramatic play of light and shadow on the stone surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00764184445142746 + ], + "opposite_logprob_tokens": [ + -4.882641792297363 + ], + "logprob_reward": -0.00764184445142746, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00764184445142746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2661849856376648, + 0.15465237200260162, + 0.2752877175807953, + 0.040582310408353806, + 0.14926151931285858, + 0.014192234724760056, + 0.015173714607954025, + 0.017213232815265656, + 0.034251756966114044, + 0.03320016711950302 + ], + "credit_map": [ + 0.03365125507116318, + 0.015604560263454914, + 0.07775340229272842, + 0.0330205038189888, + 0.010478648357093334, + 0.01429703552275896, + 0.02161993458867073, + 0.05975964665412903, + 0.02185048535466194, + 0.01209343783557415, + 0.049832601100206375, + 0.01919075846672058, + 0.007863590493798256, + 0.00835107360035181, + 0.012914804741740227, + 0.02255561761558056, + 0.033992402255535126, + 0.014078117907047272, + 0.10212627798318863, + 0.03167348355054855, + 0.007783603388816118, + 0.013880783692002296, + 0.017885299399495125, + 0.053867753595113754, + 0.0046929544769227505, + 0.00576021708548069, + 0.007174135185778141, + 0.008638674393296242, + 0.003307810053229332, + 0.0030441717244684696, + 0.005767315160483122, + 0.002197032095864415, + 0.017403719946742058, + 0.019071467220783234, + 0.026377340778708458, + 0.02730829082429409, + 0.014492377638816833, + 0.007258095312863588, + 0.025785379111766815, + 0.011564860120415688, + 0.0026919154915958643, + 0.0017987835453823209, + 0.0016013936838135123, + 0.002594791818410158, + 0.001527966931462288, + 0.0008743127691559494, + 0.0017345939995720983, + 0.0013684757286682725, + 0.002867063507437706, + 0.0018819315591827035, + 0.0018470862414687872, + 0.0021507979836314917, + 0.0010864693904295564, + 0.0012946269707754254, + 0.0013496048050001264, + 0.0026961348485201597, + 0.0033723292872309685, + 0.0016739409184083343, + 0.001802927115932107, + 0.002776217879727483, + 0.0012899155262857676, + 0.0014451697934418917, + 0.00258666486479342, + 0.002266067313030362, + 0.004438101779669523, + 0.005765622481703758, + 0.004982588347047567, + 0.0049110823310911655, + 0.004066997207701206, + 0.0029402337968349457, + 0.004446826409548521, + 0.0027003083378076553, + 0.005319019313901663, + 0.004008691757917404, + 0.006426527164876461, + 0.004269759636372328, + 0.00296604773029685, + 0.0026230921503156424, + 0.004488762933760881, + 0.003098264103755355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.00764184445142746 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, weathered stone statue carved into a sheer rock face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue made of ancient gray granite with a rough, pitted texture and deep wrinkled cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frame dominated by a massive, tightly closed eye that begins to lift slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"As the eyelid lifts, is a cloud of fine dust and tiny pebbles expelled that drips along the stone cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, creating a dramatic play of light and shadow on the stone surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005954147316515446 + ], + "opposite_logprob_tokens": [ + -5.130954265594482 + ], + "logprob_reward": -0.005954147316515446, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005954147316515446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18118873238563538, + 0.05550004541873932, + 0.059804655611515045, + 0.0884995386004448, + 0.1534668356180191, + 0.034055180847644806, + 0.01570351980626583, + 0.16154494881629944, + 0.2419840544462204, + 0.008252464234828949 + ], + "credit_map": [ + 0.02209727093577385, + 0.008882150053977966, + 0.05499565973877907, + 0.0235869400203228, + 0.006032666191458702, + 0.010132969357073307, + 0.01400880515575409, + 0.041452277451753616, + 0.005712639540433884, + 0.004473009612411261, + 0.012262998148798943, + 0.007626185659319162, + 0.004300291649997234, + 0.003999951295554638, + 0.00809911172837019, + 0.009025857783854008, + 0.007791837211698294, + 0.005120246671140194, + 0.017290523275732994, + 0.007350961212068796, + 0.0036052053328603506, + 0.004172645974904299, + 0.005754455458372831, + 0.008718777447938919, + 0.009123503230512142, + 0.010553575120866299, + 0.021710075438022614, + 0.01374789047986269, + 0.007134950719773769, + 0.005015296395868063, + 0.01264215912669897, + 0.00857208576053381, + 0.017837930470705032, + 0.022249776870012283, + 0.03229114040732384, + 0.029584679752588272, + 0.013538224622607231, + 0.007775881327688694, + 0.01610635593533516, + 0.014082850888371468, + 0.004423612263053656, + 0.003880534553900361, + 0.0029180075507611036, + 0.005904958117753267, + 0.0035135895013809204, + 0.00316784274764359, + 0.006390703842043877, + 0.0038559313397854567, + 0.0019664892461150885, + 0.0026566151063889265, + 0.0015874962555244565, + 0.002184569835662842, + 0.002070881426334381, + 0.001444404129870236, + 0.001995331607758999, + 0.001797730685211718, + 0.0194723978638649, + 0.02487054094672203, + 0.016760343685746193, + 0.024885602295398712, + 0.018630018457770348, + 0.011309221386909485, + 0.027227435261011124, + 0.01838938519358635, + 0.030214298516511917, + 0.03709573671221733, + 0.0391794815659523, + 0.03610476106405258, + 0.02858901210129261, + 0.01976020820438862, + 0.03225329890847206, + 0.018787255510687828, + 0.0011030867462977767, + 0.0013059466145932674, + 0.0013194247148931026, + 0.0011344898957759142, + 0.0007090437575243413, + 0.0006153934518806636, + 0.0010749580105766654, + 0.000990121508948505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.005954147316515446 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, weathered stone statue carved into a sheer rock face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue made of ancient gray granite with a rough, pitted texture and deep wrinkled cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frame dominated by a massive, tightly closed eye that begins to lift slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"As the eyelid lifts, is a cloud of fine dust and tiny pebbles expelled that drips along the stone cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, creating a dramatic play of light and shadow on the stone surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810097277164459 + ], + "opposite_logprob_tokens": [ + -3.2881009578704834 + ], + "logprob_reward": -0.03810097277164459, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810097277164459, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34638097882270813, + 0.042337894439697266, + 0.03890790417790413, + 0.05348442867398262, + 0.027003837749361992, + 0.03455813229084015, + 0.02278940938413143, + 0.1429232954978943, + 0.27801159024238586, + 0.013602609746158123 + ], + "credit_map": [ + 0.041299451142549515, + 0.014069801196455956, + 0.12762999534606934, + 0.04375097528100014, + 0.00900205411016941, + 0.019370971247553825, + 0.020693333819508553, + 0.07056436687707901, + 0.006692785769701004, + 0.0045339795760810375, + 0.004429685417562723, + 0.007076984737068415, + 0.003778330748900771, + 0.003696154337376356, + 0.008034277707338333, + 0.004095693118870258, + 0.004651340190321207, + 0.0041500274091959, + 0.005980565212666988, + 0.005771231837570667, + 0.0034173757303506136, + 0.003129486460238695, + 0.005677938461303711, + 0.006129934918135405, + 0.008167804218828678, + 0.0057990532368421555, + 0.01021844707429409, + 0.009363644756376743, + 0.004703945945948362, + 0.004182593896985054, + 0.006341787986457348, + 0.004707145504653454, + 0.003640187671408057, + 0.0032683108001947403, + 0.002819981426000595, + 0.004214086569845676, + 0.0030890549533069134, + 0.002361508784815669, + 0.005407191812992096, + 0.0022035129368305206, + 0.005519709084182978, + 0.005170612595975399, + 0.003690798068419099, + 0.004952786490321159, + 0.004100738558918238, + 0.0027811150066554546, + 0.004886402282863855, + 0.0034559674095362425, + 0.0025526429526507854, + 0.0029848236590623856, + 0.0022753875236958265, + 0.00358423194848001, + 0.0023289311211556196, + 0.001483227126300335, + 0.005476719234138727, + 0.002103443956002593, + 0.017929833382368088, + 0.02279067039489746, + 0.013964347541332245, + 0.02252899296581745, + 0.016717005521059036, + 0.009713999927043915, + 0.02362932451069355, + 0.015649104490876198, + 0.034060779958963394, + 0.04306682199239731, + 0.04405554011464119, + 0.04159672185778618, + 0.03321536257863045, + 0.022393831983208656, + 0.03802679106593132, + 0.02159569412469864, + 0.0015984554775059223, + 0.002066361252218485, + 0.0016087910626083612, + 0.0019739835988730192, + 0.001156798331066966, + 0.0012955882120877504, + 0.002367907203733921, + 0.0015347239095717669 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.03810097277164459 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, weathered stone statue carved into a sheer rock face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue made of ancient gray granite with a rough, pitted texture and deep wrinkled cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frame dominated by a massive, tightly closed eye that begins to lift slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"As the eyelid lifts, is a cloud of fine dust and tiny pebbles expelled that drips along the stone cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, creating a dramatic play of light and shadow on the stone surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783178567886353 + ], + "opposite_logprob_tokens": [ + -4.6347832679748535 + ], + "logprob_reward": -0.009783178567886353, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783178567886353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38675129413604736, + 0.11374861001968384, + 0.11548975110054016, + 0.05557341128587723, + 0.04224557802081108, + 0.04615376144647598, + 0.05033348500728607, + 0.033072467893362045, + 0.0685746967792511, + 0.08805691450834274 + ], + "credit_map": [ + 0.035015299916267395, + 0.03184914216399193, + 0.09134647995233536, + 0.12156794965267181, + 0.01558765210211277, + 0.018606018275022507, + 0.04436872527003288, + 0.028410032391548157, + 0.026837751269340515, + 0.013078360818326473, + 0.01033848337829113, + 0.019840214401483536, + 0.009352142922580242, + 0.007356989197432995, + 0.015431753359735012, + 0.011512917466461658, + 0.019663743674755096, + 0.012207483872771263, + 0.02870304509997368, + 0.019095001742243767, + 0.006273151375353336, + 0.008349956013262272, + 0.010875030420720577, + 0.010322344489395618, + 0.006494841072708368, + 0.008794918656349182, + 0.007806428242474794, + 0.011870790272951126, + 0.0036156063433736563, + 0.0026757442392408848, + 0.0077472347766160965, + 0.006567845121026039, + 0.0048017194494605064, + 0.0053557478822767735, + 0.003968861885368824, + 0.007894338108599186, + 0.004380705766379833, + 0.002619908656924963, + 0.01007779873907566, + 0.0031464993953704834, + 0.007114346604794264, + 0.0057706986553967, + 0.005670573096722364, + 0.006940755061805248, + 0.004494033753871918, + 0.004562194924801588, + 0.00651328731328249, + 0.005087875295430422, + 0.008780629374086857, + 0.00776748638600111, + 0.0047044833190739155, + 0.007440527901053429, + 0.004779557231813669, + 0.002538901986554265, + 0.007661737035959959, + 0.006660159211605787, + 0.004467744845896959, + 0.004955301992595196, + 0.0031917737796902657, + 0.0049756355583667755, + 0.0037488427478820086, + 0.0028894217684865, + 0.004891072399914265, + 0.003952674567699432, + 0.008927756920456886, + 0.01063492614775896, + 0.010468634776771069, + 0.010366604663431644, + 0.008155034855008125, + 0.005565396510064602, + 0.00927822943776846, + 0.00517811207100749, + 0.017482753843069077, + 0.00968071911484003, + 0.024266406893730164, + 0.009688537567853928, + 0.004878530278801918, + 0.0062578413635492325, + 0.00731853349134326, + 0.008483586832880974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.009783178567886353 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_15/index_tensor([3145, 3145])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_15/index_tensor([3145, 3145])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7350f7378fbd022eee143b89e0309dfefbd0d12b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_15/index_tensor([3145, 3145])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium shot captures a vivid scene on a sun-drenched schoolyard. Four young boys, about eight to ten years old, huddled on cracked concrete, huddled joyfully in a tight, happy circle, engrossed in a game. A boy with messy blond hair and a bright red hoodie was laughing, his hands clenched into tight, rock-like fists. Across from him, a boy in a blue striped T-shirt countered with an open, paper-like hand. They were chattering when their playful bubble suddenly burst. Two much taller teenage boys walked directly through their circle, forcing the younger boys to spread out slightly. The first intruder, wearing an oversized black hoodie with the hood pulled down and baggy jeans, didn't break stride. His companion, with short hair, followed closely behind. They paused a few feet away from the group and, with practiced indifference, posed against a weathered brick wall in the background. The first man leaned back, shoving his thumbs in his pockets, while the second stood with his arms crossed, a stony expression on his face. They immediately launched into their own lively conversation, speaking and gesturing only to each other, their body language creating a clear boundary. The four young boys, forgetting their game, now sat down to watch the duo, their expressions a mix of silent annoyance and curiosity. The warm golden light of the late afternoon cast long shadows, creating an atmosphere of palpable social dynamics.", + "seed": 27104, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_15/index_3145_seed_27104_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08227782137691975, + "vlm_logprob": -0.08227782137691975, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.08227782137691975, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched schoolyard with cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four young boys (approx. 8-10 years old) initially huddled in a tight circle playing a game?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do two much taller teenage boys walk through the circle and pose against a weathered brick wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the teenage boys wearing an oversized black hoodie with the hood pulled down and baggy jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with warm golden late afternoon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2781350314617157 + ], + "opposite_logprob_tokens": [ + -1.528135061264038 + ], + "logprob_reward": -0.2781350314617157, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2781350314617157, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05853485316038132, + 0.010121884755790234, + 0.01524637546390295, + 0.019716542214155197, + 0.08004385232925415, + 0.6693746447563171, + 0.08544716238975525, + 0.031020119786262512, + 0.01528707705438137, + 0.015207460150122643 + ], + "credit_map": [ + 0.00787056889384985, + 0.006115150637924671, + 0.00364045612514019, + 0.0045408401638269424, + 0.004076804965734482, + 0.013506563380360603, + 0.003768298076465726, + 0.015016172081232071, + 0.0012846678728237748, + 0.0012967726215720177, + 0.0008101051207631826, + 0.0010959217324852943, + 0.0013732638908550143, + 0.0013919237535446882, + 0.0013638394884765148, + 0.0015053902752697468, + 0.001851315493695438, + 0.0026607278268784285, + 0.001878015580587089, + 0.0017908393638208508, + 0.001461175037547946, + 0.0023411919828504324, + 0.0014545174781233072, + 0.0018085924675688148, + 0.002866279100999236, + 0.0038847264368087053, + 0.0020616387482732534, + 0.0027566507924348116, + 0.0023997703101485968, + 0.002591244876384735, + 0.0017221953021362424, + 0.0014340378111228347, + 0.010928656905889511, + 0.0063592055812478065, + 0.00813704077154398, + 0.010138662531971931, + 0.009304237551987171, + 0.010043177753686905, + 0.0056717307306826115, + 0.019461145624518394, + 0.030204080045223236, + 0.0149042634293437, + 0.011369108222424984, + 0.007784247398376465, + 0.2921539545059204, + 0.2822280526161194, + 0.013085528276860714, + 0.01764542981982231, + 0.012675081379711628, + 0.005872632842510939, + 0.011849630624055862, + 0.004460480995476246, + 0.008651693351566792, + 0.009218588471412659, + 0.00574757345020771, + 0.0269714817404747, + 0.005662469193339348, + 0.0025810939259827137, + 0.004026379436254501, + 0.0035516354255378246, + 0.003373006358742714, + 0.003951325546950102, + 0.0029348740354180336, + 0.004939338658004999, + 0.002876079175621271, + 0.001601073076017201, + 0.0016415168065577745, + 0.0015657561598345637, + 0.0020935109350830317, + 0.002388788154348731, + 0.001412860001437366, + 0.001707493676804006, + 0.0021909044589847326, + 0.0018109787488356233, + 0.001568890642374754, + 0.0012306353310123086, + 0.0027186565566807985, + 0.002256173174828291, + 0.0012872248189523816, + 0.0021439974661916494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.2781350314617157 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched schoolyard with cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four young boys (approx. 8-10 years old) initially huddled in a tight circle playing a game?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do two much taller teenage boys walk through the circle and pose against a weathered brick wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the teenage boys wearing an oversized black hoodie with the hood pulled down and baggy jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with warm golden late afternoon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373878449201584 + ], + "opposite_logprob_tokens": [ + -3.408738851547241 + ], + "logprob_reward": -0.03373878449201584, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373878449201584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14618897438049316, + 0.07652018219232559, + 0.09131784737110138, + 0.03436857834458351, + 0.09380849450826645, + 0.2384374737739563, + 0.21883240342140198, + 0.05984976142644882, + 0.016372518613934517, + 0.024303792044520378 + ], + "credit_map": [ + 0.017786575481295586, + 0.021831706166267395, + 0.011526048183441162, + 0.014457070268690586, + 0.020258871838450432, + 0.032316189259290695, + 0.010056613944470882, + 0.017955893650650978, + 0.01182617899030447, + 0.014397719874978065, + 0.006301711779087782, + 0.0074620479717850685, + 0.013618163764476776, + 0.010531828738749027, + 0.0064992946572601795, + 0.0058832368813455105, + 0.009351318702101707, + 0.02591842971742153, + 0.007148598786443472, + 0.012494589202105999, + 0.010921775363385677, + 0.014309428632259369, + 0.0064581772312521935, + 0.0047155325300991535, + 0.004787679295986891, + 0.005130085162818432, + 0.0030844779685139656, + 0.0038646580651402473, + 0.0067200083285570145, + 0.004817538894712925, + 0.0032576746307313442, + 0.0027064539026468992, + 0.012199156917631626, + 0.011800872161984444, + 0.006819701287895441, + 0.00946279987692833, + 0.01216970570385456, + 0.012694284319877625, + 0.005824120715260506, + 0.022837860509753227, + 0.01311399694532156, + 0.010602042078971863, + 0.0070663513615727425, + 0.006388911046087742, + 0.0928601399064064, + 0.08889677375555038, + 0.008577648550271988, + 0.010931595228612423, + 0.032972365617752075, + 0.015508746728301048, + 0.027674997225403786, + 0.015283636748790741, + 0.015321549028158188, + 0.025952234864234924, + 0.014079378917813301, + 0.07203948497772217, + 0.0071227471344172955, + 0.00598089350387454, + 0.006677720695734024, + 0.004801533650606871, + 0.010413430631160736, + 0.008628813549876213, + 0.0057601663284003735, + 0.010464455932378769, + 0.002707999898120761, + 0.002283340087160468, + 0.0018072902457788587, + 0.002205284545198083, + 0.0022828702349215746, + 0.002238847780972719, + 0.001466658664867282, + 0.001380226225592196, + 0.0043020001612603664, + 0.0025880830362439156, + 0.002341729588806629, + 0.0022921720519661903, + 0.0033441749401390553, + 0.0030861536506563425, + 0.0025003920309245586, + 0.003849085420370102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.03373878449201584 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched schoolyard with cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four young boys (approx. 8-10 years old) initially huddled in a tight circle playing a game?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do two much taller teenage boys walk through the circle and pose against a weathered brick wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the teenage boys wearing an oversized black hoodie with the hood pulled down and baggy jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with warm golden late afternoon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038113709539175034 + ], + "opposite_logprob_tokens": [ + -3.288113594055176 + ], + "logprob_reward": -0.038113709539175034, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038113709539175034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02221817709505558, + 0.015912098810076714, + 0.016378654167056084, + 0.012653221376240253, + 0.04690930247306824, + 0.49612048268318176, + 0.32537370920181274, + 0.02817927673459053, + 0.02598118782043457, + 0.010273879393935204 + ], + "credit_map": [ + 0.0034668913576751947, + 0.0032354164868593216, + 0.0016847619554027915, + 0.003258763113990426, + 0.002497411798685789, + 0.0034667239524424076, + 0.0019999819342046976, + 0.0026082259137183428, + 0.0026508918963372707, + 0.002861221320927143, + 0.0013429663376882672, + 0.002041590167209506, + 0.002158733084797859, + 0.0022449903190135956, + 0.0011021988466382027, + 0.0015095070702955127, + 0.0022114766761660576, + 0.0025787949562072754, + 0.0021286443807184696, + 0.0020000108052045107, + 0.0019899695180356503, + 0.0026361383497714996, + 0.0015766442520543933, + 0.0012569744139909744, + 0.001614757813513279, + 0.001693811733275652, + 0.0009966864017769694, + 0.0018791294423863292, + 0.002628196496516466, + 0.0015990369720384479, + 0.0014640172012150288, + 0.0007775851991027594, + 0.006596533581614494, + 0.005648401100188494, + 0.005099646747112274, + 0.00782946776598692, + 0.006239146459847689, + 0.00605058204382658, + 0.0031624038238078356, + 0.006283119786530733, + 0.020672155544161797, + 0.012415597215294838, + 0.008158029988408089, + 0.006340006832033396, + 0.21485726535320282, + 0.21025198698043823, + 0.01055245753377676, + 0.012872976250946522, + 0.04788154736161232, + 0.02747814729809761, + 0.040721576660871506, + 0.019975006580352783, + 0.028948992490768433, + 0.0419311448931694, + 0.017299775034189224, + 0.10113753378391266, + 0.004598279483616352, + 0.0037920051254332066, + 0.002539952052757144, + 0.002351805567741394, + 0.005101450253278017, + 0.0047216336242854595, + 0.0027227383106946945, + 0.0023514123167842627, + 0.003467177739366889, + 0.0022371362429112196, + 0.0029826785903424025, + 0.002021484775468707, + 0.003715184750035405, + 0.0035085801500827074, + 0.003037596819922328, + 0.005011350382119417, + 0.001789699774235487, + 0.0012100210878998041, + 0.001203079940751195, + 0.0009296369971707463, + 0.001483594416640699, + 0.001634605578146875, + 0.000797464803326875, + 0.0012257765047252178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.038113709539175034 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched schoolyard with cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four young boys (approx. 8-10 years old) initially huddled in a tight circle playing a game?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do two much taller teenage boys walk through the circle and pose against a weathered brick wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the teenage boys wearing an oversized black hoodie with the hood pulled down and baggy jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with warm golden late afternoon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023297740146517754 + ], + "opposite_logprob_tokens": [ + -3.7732977867126465 + ], + "logprob_reward": -0.023297740146517754, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023297740146517754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04109629616141319, + 0.02218933217227459, + 0.030372055247426033, + 0.03293951600790024, + 0.08526221662759781, + 0.16845695674419403, + 0.21446342766284943, + 0.2826869785785675, + 0.050851527601480484, + 0.07168176025152206 + ], + "credit_map": [ + 0.005117654800415039, + 0.005099561996757984, + 0.003730514319613576, + 0.004368928261101246, + 0.005551867187023163, + 0.005345322657376528, + 0.0032279014121741056, + 0.008654541335999966, + 0.003069463651627302, + 0.0038436304312199354, + 0.0018920137081295252, + 0.0028885062783956528, + 0.003820145968347788, + 0.002785751363262534, + 0.00222454103641212, + 0.0016652767080813646, + 0.004055190831422806, + 0.004762720782309771, + 0.003099987516179681, + 0.0036419250536710024, + 0.005188010632991791, + 0.005139138083904982, + 0.002499863738194108, + 0.00198521395213902, + 0.005087529309093952, + 0.00532962242141366, + 0.003118048654869199, + 0.0038458523340523243, + 0.004775378853082657, + 0.004194324836134911, + 0.002458972157910466, + 0.004129780922085047, + 0.013330288231372833, + 0.008490251377224922, + 0.009835697710514069, + 0.01126015279442072, + 0.010677202604711056, + 0.011021694168448448, + 0.00607904139906168, + 0.014567875303328037, + 0.013255054131150246, + 0.01403126958757639, + 0.006796085741370916, + 0.008578217588365078, + 0.05745203420519829, + 0.050845447927713394, + 0.008111996576189995, + 0.009386831894516945, + 0.03261201083660126, + 0.024072589352726936, + 0.02015911042690277, + 0.014491698704659939, + 0.02949085272848606, + 0.03171173110604286, + 0.015139468014240265, + 0.04678595811128616, + 0.029111994430422783, + 0.03365205228328705, + 0.0393034964799881, + 0.017002422362565994, + 0.048919036984443665, + 0.038268011063337326, + 0.021991046145558357, + 0.05443890765309334, + 0.008073847740888596, + 0.004527551122009754, + 0.005138998385518789, + 0.0044925701804459095, + 0.006348541472107172, + 0.008001483045518398, + 0.007356687448918819, + 0.006911842152476311, + 0.009520670399069786, + 0.008246337994933128, + 0.0075196861289441586, + 0.007563294842839241, + 0.011752353981137276, + 0.008724677376449108, + 0.00765692163258791, + 0.0106978053227067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.023297740146517754 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched schoolyard with cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four young boys (approx. 8-10 years old) initially huddled in a tight circle playing a game?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do two much taller teenage boys walk through the circle and pose against a weathered brick wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the teenage boys wearing an oversized black hoodie with the hood pulled down and baggy jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with warm golden late afternoon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810384124517441 + ], + "opposite_logprob_tokens": [ + -3.2881038188934326 + ], + "logprob_reward": -0.03810384124517441, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810384124517441, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10965904593467712, + 0.02466859482228756, + 0.033388059586286545, + 0.036474306136369705, + 0.08076994121074677, + 0.41135188937187195, + 0.1890045702457428, + 0.04903586581349373, + 0.025343026965856552, + 0.040304604917764664 + ], + "credit_map": [ + 0.012338845990598202, + 0.01585235446691513, + 0.008075522258877754, + 0.008381113409996033, + 0.010774374939501286, + 0.019805630668997765, + 0.006987881381064653, + 0.027443327009677887, + 0.0031922957859933376, + 0.0036245817318558693, + 0.0022917669266462326, + 0.002245886018499732, + 0.004797967150807381, + 0.003656401764601469, + 0.0022868518717586994, + 0.002572843339294195, + 0.005480308551341295, + 0.004181572701781988, + 0.002882901579141617, + 0.0037148960400372744, + 0.004741765558719635, + 0.005510337185114622, + 0.0031168588902801275, + 0.0037594211753457785, + 0.007071933709084988, + 0.006448820233345032, + 0.002555851824581623, + 0.004142443183809519, + 0.0050320103764534, + 0.005539668723940849, + 0.0027658198960125446, + 0.0029177593532949686, + 0.012579337693750858, + 0.010817066766321659, + 0.0073116314597427845, + 0.008320028893649578, + 0.01390080712735653, + 0.010435739532113075, + 0.008458035998046398, + 0.008947288617491722, + 0.01978243701159954, + 0.017013069242239, + 0.0120624341070652, + 0.010899928398430347, + 0.15086908638477325, + 0.15186968445777893, + 0.027755698189139366, + 0.021099576726555824, + 0.02768506482243538, + 0.02073286846280098, + 0.02003791555762291, + 0.014737063087522984, + 0.019956547766923904, + 0.022438300773501396, + 0.014604780822992325, + 0.048812031745910645, + 0.007927712984383106, + 0.006021938752382994, + 0.0052231536246836185, + 0.005942741874605417, + 0.006618591491132975, + 0.006049351301044226, + 0.004817577078938484, + 0.0064347973093390465, + 0.00355698773637414, + 0.003029943211004138, + 0.0027581783942878246, + 0.002403146354481578, + 0.004087966866791248, + 0.0037242488469928503, + 0.003558747936040163, + 0.002223807852715254, + 0.0048888856545090675, + 0.006029088515788317, + 0.002824027556926012, + 0.0030745144467800856, + 0.01062299869954586, + 0.005349178332835436, + 0.0035836640745401382, + 0.0039322455413639545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.03810384124517441 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a vivid scene on a sun-drenched schoolyard. Four young boys, about eight to ten years old, huddled on cracked concrete, huddled joyfully in a tight, happy circle, engrossed in a game. A boy with messy blond hair and a bright red hoodie was laughing, his hands clenched into tight, rock-like fists. Across from him, a boy in a blue striped T-shirt countered with an open, paper-like hand. They were chattering when their playful bubble suddenly burst. Two much taller teenage boys walked directly through their circle, forcing the younger boys to spread out slightly. The first intruder, wearing an oversized black hoodie with the hood pulled down and baggy jeans, didn't break stride. His companion, with short hair, followed closely behind. They paused a few feet away from the group and, with practiced indifference, posed against a weathered brick wall in the background. The first man leaned back, shoving his thumbs in his pockets, while the second stood with his arms crossed, a stony expression on his face. They immediately launched into their own lively conversation, speaking and gesturing only to each other, their body language creating a clear boundary. The four young boys, forgetting their game, now sat down to watch the duo, their expressions a mix of silent annoyance and curiosity. The warm golden light of the late afternoon cast long shadows, creating an atmosphere of palpable social dynamics.", + "seed": 27105, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_15/index_3145_seed_27105_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.040366864297538996, + "vlm_logprob": -0.040366864297538996, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.040366864297538996, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched schoolyard with cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four young boys (approx. 8-10 years old) initially huddled in a tight circle playing a game?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do two much taller teenage boys walk through the circle and pose against a weathered brick wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the teenage boys wearing an oversized black hoodie with the hood pulled down and baggy jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with warm golden late afternoon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07847852259874344 + ], + "opposite_logprob_tokens": [ + -2.7034785747528076 + ], + "logprob_reward": -0.07847852259874344, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07847852259874344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044108010828495026, + 0.01342721190303564, + 0.01119727548211813, + 0.0827215313911438, + 0.05848760902881622, + 0.039829321205616, + 0.04978341981768608, + 0.056529171764850616, + 0.09729954600334167, + 0.5466168522834778 + ], + "credit_map": [ + 0.007288086228072643, + 0.007950077764689922, + 0.00809725746512413, + 0.0056153046898543835, + 0.006275912281125784, + 0.0031877944711595774, + 0.003109554061666131, + 0.0025840250309556723, + 0.0019907860551029444, + 0.0018415035447105765, + 0.0014975592494010925, + 0.0012356217484921217, + 0.0011703216005116701, + 0.003621380776166916, + 0.0013303508749231696, + 0.0007396895671263337, + 0.0021565300412476063, + 0.0023276768624782562, + 0.0013467323733493686, + 0.001019177958369255, + 0.0014406605623662472, + 0.001152808079496026, + 0.0008500419207848608, + 0.0009036491392180324, + 0.016976067796349525, + 0.006378813646733761, + 0.007980669848620892, + 0.022090641781687737, + 0.007045247126370668, + 0.00789328757673502, + 0.008143965154886246, + 0.0062128398567438126, + 0.005577376578003168, + 0.0054114628583192825, + 0.013933186419308186, + 0.00705709308385849, + 0.005144280847162008, + 0.010593011975288391, + 0.0044798231683671474, + 0.006291376426815987, + 0.005209166556596756, + 0.004093914292752743, + 0.004810214042663574, + 0.0031338296830654144, + 0.0028623538091778755, + 0.00824966561049223, + 0.004088943358510733, + 0.007381238043308258, + 0.004905288107693195, + 0.004467986989766359, + 0.0053458744660019875, + 0.0038368969690054655, + 0.004677849356085062, + 0.007400774396955967, + 0.004540632013231516, + 0.014608122408390045, + 0.008555309846997261, + 0.005619924981147051, + 0.006051936186850071, + 0.004378518555313349, + 0.004075914621353149, + 0.008383254520595074, + 0.005254080519080162, + 0.014210233464837074, + 0.013641447760164738, + 0.01175934448838234, + 0.018785880878567696, + 0.011572971008718014, + 0.008266206830739975, + 0.009714089334011078, + 0.007795154582709074, + 0.01576446183025837, + 0.025575153529644012, + 0.07123167812824249, + 0.00956061389297247, + 0.016392450779676437, + 0.013295738026499748, + 0.24972540140151978, + 0.008653411641716957, + 0.1521824598312378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.07847852259874344 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched schoolyard with cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four young boys (approx. 8-10 years old) initially huddled in a tight circle playing a game?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do two much taller teenage boys walk through the circle and pose against a weathered brick wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the teenage boys wearing an oversized black hoodie with the hood pulled down and baggy jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with warm golden late afternoon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029844488948583603 + ], + "opposite_logprob_tokens": [ + -3.5298445224761963 + ], + "logprob_reward": -0.029844488948583603, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029844488948583603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12327253073453903, + 0.04227187857031822, + 0.04327543079853058, + 0.22361305356025696, + 0.04633164778351784, + 0.061005331575870514, + 0.030688036233186722, + 0.04003089293837547, + 0.04411052539944649, + 0.3454006612300873 + ], + "credit_map": [ + 0.015493996441364288, + 0.018905034288764, + 0.02690257504582405, + 0.021363697946071625, + 0.014829782769083977, + 0.008031741715967655, + 0.00995866023004055, + 0.0077870567329227924, + 0.007168392650783062, + 0.006851566024124622, + 0.0065234643407166, + 0.004286196082830429, + 0.005749842617660761, + 0.004599594511091709, + 0.00395811814814806, + 0.003134707920253277, + 0.00709947757422924, + 0.004792208317667246, + 0.005106155760586262, + 0.003218360710889101, + 0.007079039700329304, + 0.0051765842363238335, + 0.006997484248131514, + 0.00380612607114017, + 0.07585979998111725, + 0.01301587838679552, + 0.02057173289358616, + 0.06935475021600723, + 0.011030133813619614, + 0.013165566138923168, + 0.011026144959032536, + 0.009589063934981823, + 0.004426359664648771, + 0.006373206619173288, + 0.005350865423679352, + 0.005076171364635229, + 0.009051017463207245, + 0.006495311856269836, + 0.004789955914020538, + 0.0047687659971416, + 0.005883925128728151, + 0.005768803413957357, + 0.00672304630279541, + 0.005977028515189886, + 0.006425234954804182, + 0.010777837596833706, + 0.010115851648151875, + 0.009333614259958267, + 0.00276472931727767, + 0.0023454150650650263, + 0.004029484931379557, + 0.0029524697456508875, + 0.0027739135548472404, + 0.004603849723935127, + 0.002543196314945817, + 0.008674981072545052, + 0.004653716925531626, + 0.005732951685786247, + 0.004707846790552139, + 0.004327806178480387, + 0.004992273636162281, + 0.004884056281298399, + 0.004419734701514244, + 0.006312512326985598, + 0.006814727559685707, + 0.004314678255468607, + 0.0063256495632231236, + 0.005213635042309761, + 0.0031302389688789845, + 0.00395392719656229, + 0.003217787714675069, + 0.01113988272845745, + 0.01962018385529518, + 0.02322942204773426, + 0.007981636561453342, + 0.010603884235024452, + 0.009792398661375046, + 0.13757841289043427, + 0.00694085331633687, + 0.1296539008617401 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.029844488948583603 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched schoolyard with cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four young boys (approx. 8-10 years old) initially huddled in a tight circle playing a game?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do two much taller teenage boys walk through the circle and pose against a weathered brick wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the teenage boys wearing an oversized black hoodie with the hood pulled down and baggy jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with warm golden late afternoon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067427014000713825 + ], + "opposite_logprob_tokens": [ + -5.006742477416992 + ], + "logprob_reward": -0.0067427014000713825, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067427014000713825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0355629026889801, + 0.01001279428601265, + 0.011990096420049667, + 0.020626381039619446, + 0.03598490357398987, + 0.03733116760849953, + 0.022709310054779053, + 0.042624332010746, + 0.032602693885564804, + 0.7505553960800171 + ], + "credit_map": [ + 0.008037101477384567, + 0.0035546207800507545, + 0.008334551006555557, + 0.0049352324567735195, + 0.003300128970295191, + 0.0021414002403616905, + 0.0025817370042204857, + 0.0026781298220157623, + 0.0017911200411617756, + 0.0018083470640704036, + 0.0016390627715736628, + 0.0010631842305883765, + 0.001016647438518703, + 0.0010776749113574624, + 0.000959482102189213, + 0.0006572753190994263, + 0.0016268085455521941, + 0.001301703043282032, + 0.0015197412576526403, + 0.0009044359321705997, + 0.001838546129874885, + 0.001513287890702486, + 0.001690923934802413, + 0.0015946493949741125, + 0.0029898537322878838, + 0.002450711326673627, + 0.0029349082615226507, + 0.003278807271271944, + 0.0035052429884672165, + 0.0021808319725096226, + 0.0015218844637274742, + 0.0017641405574977398, + 0.003277956508100033, + 0.003801774699240923, + 0.00819319672882557, + 0.004832076840102673, + 0.004423382226377726, + 0.00483336579054594, + 0.0025368703063577414, + 0.00408628024160862, + 0.0031945689115673304, + 0.0031336485408246517, + 0.0035379789769649506, + 0.0032548634335398674, + 0.003350059036165476, + 0.007929785177111626, + 0.0066638062708079815, + 0.006266456563025713, + 0.002559787593781948, + 0.0020135052036494017, + 0.0027536561246961355, + 0.0019056439632549882, + 0.0025303857401013374, + 0.003041632240638137, + 0.002590565010905266, + 0.005314136389642954, + 0.005941037088632584, + 0.004261028487235308, + 0.004355237819254398, + 0.0032181148417294025, + 0.0038597004022449255, + 0.006385317072272301, + 0.003719606902450323, + 0.010884289629757404, + 0.004471378866583109, + 0.004230228252708912, + 0.005579977761954069, + 0.004148791544139385, + 0.002808595774695277, + 0.003505638102069497, + 0.0027506090700626373, + 0.005107472650706768, + 0.03276704251766205, + 0.12982597947120667, + 0.012661917135119438, + 0.02282160148024559, + 0.017693333327770233, + 0.3410625755786896, + 0.011071398854255676, + 0.182651549577713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.0067427014000713825 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched schoolyard with cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four young boys (approx. 8-10 years old) initially huddled in a tight circle playing a game?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do two much taller teenage boys walk through the circle and pose against a weathered brick wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the teenage boys wearing an oversized black hoodie with the hood pulled down and baggy jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with warm golden late afternoon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038103725761175156 + ], + "opposite_logprob_tokens": [ + -3.2881038188934326 + ], + "logprob_reward": -0.038103725761175156, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038103725761175156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.009668258018791676, + 0.0026662300806492567, + 0.002755307126790285, + 0.022899748757481575, + 0.02417561784386635, + 0.02350637875497341, + 0.004588642623275518, + 0.006840053014457226, + 0.012431426905095577, + 0.8904682993888855 + ], + "credit_map": [ + 0.0018729658331722021, + 0.0012936325510963798, + 0.002017332473769784, + 0.0012461880687624216, + 0.00109540403354913, + 0.0007070977007970214, + 0.0006903472822159529, + 0.0007452911231666803, + 0.00043862327584065497, + 0.00036565581103786826, + 0.00034268651506863534, + 0.0002924729196820408, + 0.00038408918771892786, + 0.00034730613697320223, + 0.00026712522958405316, + 0.00022827151406090707, + 0.0003792050702031702, + 0.0007704595336690545, + 0.0002680277975741774, + 0.0002892195770982653, + 0.0002704639337025583, + 0.00039913805085234344, + 0.00018265271501149982, + 0.00019614084158092737, + 0.008315897546708584, + 0.0013410046231001616, + 0.002261729445308447, + 0.0072141289710998535, + 0.0006197137990966439, + 0.0011931973276659846, + 0.0009112302213907242, + 0.001042849151417613, + 0.0017474401975050569, + 0.0025995951145887375, + 0.003305845893919468, + 0.0027765168342739344, + 0.004106348380446434, + 0.005829861853271723, + 0.0017060438403859735, + 0.0021039689891040325, + 0.003303660312667489, + 0.0024962821044027805, + 0.002037263475358486, + 0.0016380376182496548, + 0.0017410649452358484, + 0.0052788578905165195, + 0.0027042627334594727, + 0.004306953400373459, + 0.00043000790174119174, + 0.0004381100879982114, + 0.0005751906428486109, + 0.0005407401476986706, + 0.00047656288370490074, + 0.0006742916884832084, + 0.0004977028584107757, + 0.0009560368489474058, + 0.001221509650349617, + 0.0009799058316275477, + 0.0007391549297608435, + 0.0007884899387136102, + 0.0006109693204052746, + 0.000855116406455636, + 0.00047999870730564, + 0.0011649089865386486, + 0.0020857886411249638, + 0.0010916779283434153, + 0.0027532954700291157, + 0.0013375460403040051, + 0.0007874189177528024, + 0.001134373713284731, + 0.000832211400847882, + 0.002409116132184863, + 0.03997873514890671, + 0.13560114800930023, + 0.015281349420547485, + 0.02693704329431057, + 0.020971333608031273, + 0.4059947729110718, + 0.013560831546783447, + 0.23214319348335266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 5 + ], + "selected_reward": -0.038103725761175156 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched schoolyard with cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four young boys (approx. 8-10 years old) initially huddled in a tight circle playing a game?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do two much taller teenage boys walk through the circle and pose against a weathered brick wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the teenage boys wearing an oversized black hoodie with the hood pulled down and baggy jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with warm golden late afternoon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0486648827791214 + ], + "opposite_logprob_tokens": [ + -3.0486648082733154 + ], + "logprob_reward": -0.0486648827791214, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0486648827791214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0289236381649971, + 0.012868565507233143, + 0.0188132431358099, + 0.0935506671667099, + 0.04053695127367973, + 0.03426876291632652, + 0.022937733680009842, + 0.015779396519064903, + 0.032903458923101425, + 0.6994175910949707 + ], + "credit_map": [ + 0.004770483355969191, + 0.004730704706162214, + 0.005104809999465942, + 0.004413678776472807, + 0.0036907633766531944, + 0.0020740062464028597, + 0.0021114444825798273, + 0.0020277462899684906, + 0.002283754525706172, + 0.0012827091850340366, + 0.0012057742569595575, + 0.0011907604057341814, + 0.0020387819968163967, + 0.0027723161038011312, + 0.001361851580440998, + 0.0007326161139644682, + 0.002261875197291374, + 0.001948477467522025, + 0.0019306542817503214, + 0.0012128567323088646, + 0.003957914188504219, + 0.002299083862453699, + 0.0022494548466056585, + 0.0029529265593737364, + 0.031766846776008606, + 0.0056867050006985664, + 0.009585492312908173, + 0.027027223259210587, + 0.0041907248087227345, + 0.006183215416967869, + 0.004656911361962557, + 0.004453550558537245, + 0.004042340908199549, + 0.005036910995841026, + 0.00656080711632967, + 0.005813124123960733, + 0.005298134405165911, + 0.005862376652657986, + 0.0032673156820237637, + 0.004655941855162382, + 0.0021397757809609175, + 0.0026761849876493216, + 0.003899031551554799, + 0.0018415608210489154, + 0.0023134490475058556, + 0.008845080621540546, + 0.0032926525454968214, + 0.009261024184525013, + 0.0023407696280628443, + 0.002914074342697859, + 0.0027883013244718313, + 0.001759184175170958, + 0.001956075197085738, + 0.003167342394590378, + 0.0018709150608628988, + 0.006141073536127806, + 0.002265159972012043, + 0.0022289163898676634, + 0.0015791032928973436, + 0.0015082653844729066, + 0.0022304931189864874, + 0.0015366649022325873, + 0.001814626855775714, + 0.0026161668356508017, + 0.005087058991193771, + 0.0035341570619493723, + 0.008280275389552116, + 0.0037803847808390856, + 0.0024360944516956806, + 0.002719530602917075, + 0.002388221677392721, + 0.004677736200392246, + 0.03353814408183098, + 0.08052530139684677, + 0.012530950829386711, + 0.02113843895494938, + 0.0165393203496933, + 0.31994864344596863, + 0.011408355087041855, + 0.2037884145975113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 4 + ], + "selected_reward": -0.0486648827791214 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a vivid scene on a sun-drenched schoolyard. Four young boys, about eight to ten years old, huddled on cracked concrete, huddled joyfully in a tight, happy circle, engrossed in a game. A boy with messy blond hair and a bright red hoodie was laughing, his hands clenched into tight, rock-like fists. Across from him, a boy in a blue striped T-shirt countered with an open, paper-like hand. They were chattering when their playful bubble suddenly burst. Two much taller teenage boys walked directly through their circle, forcing the younger boys to spread out slightly. The first intruder, wearing an oversized black hoodie with the hood pulled down and baggy jeans, didn't break stride. His companion, with short hair, followed closely behind. They paused a few feet away from the group and, with practiced indifference, posed against a weathered brick wall in the background. The first man leaned back, shoving his thumbs in his pockets, while the second stood with his arms crossed, a stony expression on his face. They immediately launched into their own lively conversation, speaking and gesturing only to each other, their body language creating a clear boundary. The four young boys, forgetting their game, now sat down to watch the duo, their expressions a mix of silent annoyance and curiosity. The warm golden light of the late afternoon cast long shadows, creating an atmosphere of palpable social dynamics.", + "seed": 27106, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_15/index_3145_seed_27106_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.15064211711287498, + "vlm_logprob": -0.15064211711287498, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.15064211711287498, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched schoolyard with cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four young boys (approx. 8-10 years old) initially huddled in a tight circle playing a game?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do two much taller teenage boys walk through the circle and pose against a weathered brick wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the teenage boys wearing an oversized black hoodie with the hood pulled down and baggy jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with warm golden late afternoon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6591609120368958 + ], + "opposite_logprob_tokens": [ + -0.7841609120368958 + ], + "logprob_reward": -0.6591609120368958, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6591609120368958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20966199040412903, + 0.06275762617588043, + 0.07002314925193787, + 0.07404030114412308, + 0.262489914894104, + 0.04172272980213165, + 0.05309447646141052, + 0.07388090342283249, + 0.08510245382785797, + 0.06722643226385117 + ], + "credit_map": [ + 0.03020235151052475, + 0.02705598808825016, + 0.015558524988591671, + 0.02791476435959339, + 0.03929617628455162, + 0.030964551493525505, + 0.022022489458322525, + 0.01664716750383377, + 0.006767451297491789, + 0.00911715254187584, + 0.005747533403337002, + 0.004571160301566124, + 0.00822892040014267, + 0.011771110817790031, + 0.005997312255203724, + 0.010556989349424839, + 0.005185096990317106, + 0.009419816546142101, + 0.0063796802423894405, + 0.007715444080531597, + 0.0113422442227602, + 0.010145113803446293, + 0.00932618323713541, + 0.010509577579796314, + 0.005962836090475321, + 0.00743061164394021, + 0.005448983516544104, + 0.005299355834722519, + 0.008467827923595905, + 0.01763942465186119, + 0.013024508021771908, + 0.01076675858348608, + 0.025992823764681816, + 0.031361643224954605, + 0.03759223222732544, + 0.023983540013432503, + 0.03913659602403641, + 0.03378177061676979, + 0.036168478429317474, + 0.03447284176945686, + 0.005204536486417055, + 0.006982653867453337, + 0.0038750513922423124, + 0.005075742024928331, + 0.0045951055362820625, + 0.004351106937974691, + 0.007045252248644829, + 0.004593283403664827, + 0.005986811127513647, + 0.01007023360580206, + 0.0030246207024902105, + 0.0033262143842875957, + 0.006223747972398996, + 0.011107223108410835, + 0.006125095300376415, + 0.007230537943542004, + 0.019970376044511795, + 0.014387709088623524, + 0.005599732045084238, + 0.005975519306957722, + 0.006037736311554909, + 0.007636986672878265, + 0.008890940807759762, + 0.005381908733397722, + 0.023734629154205322, + 0.026296468451619148, + 0.006477962713688612, + 0.005973802413791418, + 0.005054153501987457, + 0.005400113761425018, + 0.004773038439452648, + 0.007392294239252806, + 0.017492758110165596, + 0.015312529169023037, + 0.00792261678725481, + 0.0048366268165409565, + 0.0046493192203342915, + 0.004565346520394087, + 0.0060112494975328445, + 0.006435994058847427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.6591609120368958 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched schoolyard with cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four young boys (approx. 8-10 years old) initially huddled in a tight circle playing a game?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do two much taller teenage boys walk through the circle and pose against a weathered brick wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the teenage boys wearing an oversized black hoodie with the hood pulled down and baggy jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with warm golden late afternoon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02985050529241562 + ], + "opposite_logprob_tokens": [ + -3.529850482940674 + ], + "logprob_reward": -0.02985050529241562, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02985050529241562, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2253163754940033, + 0.2010391354560852, + 0.07247447967529297, + 0.04989190772175789, + 0.045223038643598557, + 0.041982896625995636, + 0.08615760505199432, + 0.13332852721214294, + 0.10545291751623154, + 0.039133068174123764 + ], + "credit_map": [ + 0.026841480284929276, + 0.030399193987250328, + 0.018496790900826454, + 0.025171687826514244, + 0.044832732528448105, + 0.031330008059740067, + 0.021103402599692345, + 0.027141083031892776, + 0.02284170873463154, + 0.031082170084118843, + 0.01807471364736557, + 0.016656825318932533, + 0.028217744082212448, + 0.0337759293615818, + 0.020120088011026382, + 0.030269959941506386, + 0.00706472247838974, + 0.007110368926078081, + 0.007522296160459518, + 0.007263862527906895, + 0.011137453839182854, + 0.010759910568594933, + 0.011510432697832584, + 0.010105435736477375, + 0.004026542417705059, + 0.0048063043504953384, + 0.006152001675218344, + 0.004292449913918972, + 0.008253020234405994, + 0.00839284248650074, + 0.006861668545752764, + 0.007107077166438103, + 0.00456586480140686, + 0.006285098847001791, + 0.005007052794098854, + 0.003824521554633975, + 0.00509606720879674, + 0.00871586799621582, + 0.0044860211201012135, + 0.007242545485496521, + 0.004797386471182108, + 0.006128635723143816, + 0.00731402775272727, + 0.0054526920430362225, + 0.005015746224671602, + 0.003874348010867834, + 0.004869774915277958, + 0.004530284088104963, + 0.01097613014280796, + 0.010672781616449356, + 0.004707951098680496, + 0.006613481789827347, + 0.011523261666297913, + 0.014470075257122517, + 0.013499919325113297, + 0.013694011606276035, + 0.02128089778125286, + 0.026644406840205193, + 0.010236191563308239, + 0.013130499981343746, + 0.015883896499872208, + 0.020001526921987534, + 0.015905404463410378, + 0.010245705023407936, + 0.03160090371966362, + 0.03226728364825249, + 0.006057956255972385, + 0.007429194170981646, + 0.007391155697405338, + 0.006491422653198242, + 0.007114429492503405, + 0.00710056908428669, + 0.004428158514201641, + 0.006323144305497408, + 0.005129633005708456, + 0.004088833462446928, + 0.004485644865781069, + 0.004280128516256809, + 0.004933613818138838, + 0.005463911220431328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.02985050529241562 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched schoolyard with cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four young boys (approx. 8-10 years old) initially huddled in a tight circle playing a game?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do two much taller teenage boys walk through the circle and pose against a weathered brick wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the teenage boys wearing an oversized black hoodie with the hood pulled down and baggy jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with warm golden late afternoon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011087356135249138 + ], + "opposite_logprob_tokens": [ + -4.511087417602539 + ], + "logprob_reward": -0.011087356135249138, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011087356135249138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13686659932136536, + 0.06313581764698029, + 0.036600809544324875, + 0.04969465732574463, + 0.02572227083146572, + 0.030453449115157127, + 0.283903568983078, + 0.16347052156925201, + 0.11092275381088257, + 0.0992295891046524 + ], + "credit_map": [ + 0.020858876407146454, + 0.021028248593211174, + 0.010554569773375988, + 0.016888465732336044, + 0.02707538567483425, + 0.01518466230481863, + 0.01514729205518961, + 0.010129082016646862, + 0.00844335462898016, + 0.012293031439185143, + 0.006259421352297068, + 0.005127561744302511, + 0.007711366284638643, + 0.010569454170763493, + 0.005547648761421442, + 0.007183969486504793, + 0.003921388182789087, + 0.005763476714491844, + 0.004026373382657766, + 0.004851360805332661, + 0.00416829576715827, + 0.005005642306059599, + 0.0036357773933559656, + 0.005228492431342602, + 0.00399413425475359, + 0.005669199861586094, + 0.004867469426244497, + 0.005492415279150009, + 0.005006655119359493, + 0.008765341714024544, + 0.0063390424475073814, + 0.009560396894812584, + 0.0027772251050919294, + 0.002808845601975918, + 0.0031130125280469656, + 0.0022152098827064037, + 0.0026587173342704773, + 0.004753855522722006, + 0.00301849702373147, + 0.004376905970275402, + 0.0035848470870405436, + 0.005015178117901087, + 0.0034554305020719767, + 0.0044462126679718494, + 0.003286470426246524, + 0.0041200704872608185, + 0.0030750008299946785, + 0.0034702373668551445, + 0.030212169513106346, + 0.04632417485117912, + 0.011311320587992668, + 0.016429705545306206, + 0.04353971406817436, + 0.06736168265342712, + 0.035902488976716995, + 0.032822269946336746, + 0.040776077657938004, + 0.01928381621837616, + 0.015914101153612137, + 0.014644589275121689, + 0.012878107838332653, + 0.021653104573488235, + 0.018000269308686256, + 0.02032044343650341, + 0.021811824291944504, + 0.029105225577950478, + 0.007341776508837938, + 0.007568489294499159, + 0.014892050996422768, + 0.009841479361057281, + 0.011152826249599457, + 0.009209079667925835, + 0.012793309986591339, + 0.014407520182430744, + 0.012476754374802113, + 0.010638417676091194, + 0.015118459239602089, + 0.01280033029615879, + 0.010957160033285618, + 0.01003762986510992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.011087356135249138 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched schoolyard with cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four young boys (approx. 8-10 years old) initially huddled in a tight circle playing a game?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do two much taller teenage boys walk through the circle and pose against a weathered brick wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the teenage boys wearing an oversized black hoodie with the hood pulled down and baggy jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with warm golden late afternoon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329995296895504 + ], + "opposite_logprob_tokens": [ + -3.7732999324798584 + ], + "logprob_reward": -0.02329995296895504, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329995296895504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12430749833583832, + 0.03577353432774544, + 0.034338437020778656, + 0.031251825392246246, + 0.05949990823864937, + 0.10449974238872528, + 0.08347859978675842, + 0.14155563712120056, + 0.2652008533477783, + 0.12009397894144058 + ], + "credit_map": [ + 0.024940580129623413, + 0.013287520967423916, + 0.009956256486475468, + 0.010416876524686813, + 0.02259361930191517, + 0.02075956203043461, + 0.01335318572819233, + 0.008999885059893131, + 0.0030687025282531977, + 0.005000426899641752, + 0.003458105493336916, + 0.0036635526921600103, + 0.0048773931339383125, + 0.006184029392898083, + 0.0038308529183268547, + 0.005690467078238726, + 0.00535363145172596, + 0.006074294913560152, + 0.0031742232386022806, + 0.004126450978219509, + 0.0046947491355240345, + 0.0035973633639514446, + 0.0035362597554922104, + 0.003781461389735341, + 0.0030863506253808737, + 0.00329843838699162, + 0.0037916514556854963, + 0.0026591140776872635, + 0.004310589283704758, + 0.00646832725033164, + 0.0036400381941348314, + 0.003997312858700752, + 0.005751516204327345, + 0.00677506672218442, + 0.006630449555814266, + 0.005496491678059101, + 0.005386726930737495, + 0.011190387420356274, + 0.00854867696762085, + 0.009720588102936745, + 0.012779311276972294, + 0.02590813860297203, + 0.010705800727009773, + 0.010960932821035385, + 0.012326727621257305, + 0.01000798586755991, + 0.01033566053956747, + 0.01147517655044794, + 0.0114003149792552, + 0.00971797201782465, + 0.008186501450836658, + 0.0068351514637470245, + 0.006371558643877506, + 0.01795312948524952, + 0.007770165801048279, + 0.01524379849433899, + 0.028616497293114662, + 0.02072918973863125, + 0.010783825069665909, + 0.013778037391602993, + 0.018780989572405815, + 0.016582274809479713, + 0.019922910258173943, + 0.012361898086965084, + 0.042382869869470596, + 0.03333250805735588, + 0.025478044524788857, + 0.022469934076070786, + 0.0412464402616024, + 0.021531451493501663, + 0.03739951178431511, + 0.04136005416512489, + 0.032835498452186584, + 0.04074813425540924, + 0.006026957184076309, + 0.007909276522696018, + 0.01221571583300829, + 0.006870134733617306, + 0.006097166333347559, + 0.007391090039163828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.02329995296895504 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched schoolyard with cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four young boys (approx. 8-10 years old) initially huddled in a tight circle playing a game?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do two much taller teenage boys walk through the circle and pose against a weathered brick wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the teenage boys wearing an oversized black hoodie with the hood pulled down and baggy jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with warm golden late afternoon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029811859130859375 + ], + "opposite_logprob_tokens": [ + -3.5298118591308594 + ], + "logprob_reward": -0.029811859130859375, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029811859130859375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14884574711322784, + 0.03610004857182503, + 0.028640270233154297, + 0.03862202167510986, + 0.04194312542676926, + 0.09440477192401886, + 0.08146718144416809, + 0.12988834083080292, + 0.23020510375499725, + 0.16988348960876465 + ], + "credit_map": [ + 0.01617397554218769, + 0.017357023432850838, + 0.01273814681917429, + 0.017511285841464996, + 0.037918779999017715, + 0.01972317509353161, + 0.01608854904770851, + 0.011334802955389023, + 0.005061453208327293, + 0.005275986157357693, + 0.003665180178359151, + 0.00492629362270236, + 0.0041674100793898106, + 0.0052140080370008945, + 0.0034724532160907984, + 0.004317259415984154, + 0.0022869936656206846, + 0.003447331953793764, + 0.002837896114215255, + 0.00332553475163877, + 0.0033445251174271107, + 0.0033142429310828447, + 0.005162711255252361, + 0.00492103211581707, + 0.003926493227481842, + 0.00665821461006999, + 0.0038021733053028584, + 0.0036971629597246647, + 0.007109358906745911, + 0.0037061995826661587, + 0.004735586233437061, + 0.004986829590052366, + 0.00469115749001503, + 0.0065299710258841515, + 0.003635168308392167, + 0.004279396031051874, + 0.005423811264336109, + 0.00578092597424984, + 0.005443829111754894, + 0.006158860400319099, + 0.01243720855563879, + 0.01933225430548191, + 0.011776762083172798, + 0.009481089189648628, + 0.012476315721869469, + 0.007248084060847759, + 0.013315992429852486, + 0.008337052538990974, + 0.012451735325157642, + 0.012257111258804798, + 0.008478613570332527, + 0.007611114531755447, + 0.007442448288202286, + 0.01569220796227455, + 0.008401518687605858, + 0.00913241971284151, + 0.026900814846158028, + 0.021263787522912025, + 0.010106534697115421, + 0.01594608649611473, + 0.015815751627087593, + 0.013660833239555359, + 0.016363002359867096, + 0.009831514209508896, + 0.027237582951784134, + 0.02570470981299877, + 0.025730609893798828, + 0.03095281682908535, + 0.03202635049819946, + 0.017247308045625687, + 0.02995222806930542, + 0.04135347157716751, + 0.0495590902864933, + 0.061487168073654175, + 0.008472324348986149, + 0.009580543264746666, + 0.01643470861017704, + 0.007296612020581961, + 0.0064109014347195625, + 0.010642125271260738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.029811859130859375 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a vivid scene on a sun-drenched schoolyard. Four young boys, about eight to ten years old, huddled on cracked concrete, huddled joyfully in a tight, happy circle, engrossed in a game. A boy with messy blond hair and a bright red hoodie was laughing, his hands clenched into tight, rock-like fists. Across from him, a boy in a blue striped T-shirt countered with an open, paper-like hand. They were chattering when their playful bubble suddenly burst. Two much taller teenage boys walked directly through their circle, forcing the younger boys to spread out slightly. The first intruder, wearing an oversized black hoodie with the hood pulled down and baggy jeans, didn't break stride. His companion, with short hair, followed closely behind. They paused a few feet away from the group and, with practiced indifference, posed against a weathered brick wall in the background. The first man leaned back, shoving his thumbs in his pockets, while the second stood with his arms crossed, a stony expression on his face. They immediately launched into their own lively conversation, speaking and gesturing only to each other, their body language creating a clear boundary. The four young boys, forgetting their game, now sat down to watch the duo, their expressions a mix of silent annoyance and curiosity. The warm golden light of the late afternoon cast long shadows, creating an atmosphere of palpable social dynamics.", + "seed": 27107, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_15/index_3145_seed_27107_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0709996536374092, + "vlm_logprob": -0.0709996536374092, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.0709996536374092, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched schoolyard with cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four young boys (approx. 8-10 years old) initially huddled in a tight circle playing a game?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do two much taller teenage boys walk through the circle and pose against a weathered brick wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the teenage boys wearing an oversized black hoodie with the hood pulled down and baggy jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with warm golden late afternoon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14039362967014313 + ], + "opposite_logprob_tokens": [ + -2.1403937339782715 + ], + "logprob_reward": -0.14039362967014313, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.14039362967014313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04308672249317169, + 0.030027087777853012, + 0.030608471482992172, + 0.09994666278362274, + 0.02412685565650463, + 0.04560154676437378, + 0.058081645518541336, + 0.10491906106472015, + 0.19248856604099274, + 0.3711134195327759 + ], + "credit_map": [ + 0.004147889092564583, + 0.007878622971475124, + 0.005512563977390528, + 0.006190342362970114, + 0.008177581243216991, + 0.003899687435477972, + 0.0037932905834168196, + 0.003486743662506342, + 0.0034420201554894447, + 0.00441888440400362, + 0.0034449808299541473, + 0.0026896330527961254, + 0.005216656252741814, + 0.004135055001825094, + 0.0031852449756115675, + 0.003494614502415061, + 0.003838213859125972, + 0.004312998149544001, + 0.003030662192031741, + 0.0033229647669941187, + 0.005386343225836754, + 0.0038616093806922436, + 0.004035117104649544, + 0.0028205611743032932, + 0.006735486444085836, + 0.006772883236408234, + 0.005030609667301178, + 0.004600179847329855, + 0.006485824938863516, + 0.028847189620137215, + 0.03250864893198013, + 0.008965836837887764, + 0.0026693211402744055, + 0.002542866161093116, + 0.0037222702521830797, + 0.003269932931289077, + 0.002417578361928463, + 0.0030011849012225866, + 0.002262179274111986, + 0.0042415219359099865, + 0.0037001261953264475, + 0.005192229058593512, + 0.005180954001843929, + 0.006249968893826008, + 0.005428113508969545, + 0.006096603814512491, + 0.0067586288787424564, + 0.006994919385761023, + 0.008383153937757015, + 0.007256262470036745, + 0.006493768189102411, + 0.004390249028801918, + 0.008011194877326488, + 0.00758665194734931, + 0.007552022580057383, + 0.008408347144722939, + 0.008967670612037182, + 0.009432689286768436, + 0.006234769709408283, + 0.004838687367737293, + 0.007153613958507776, + 0.03727923333644867, + 0.011494495905935764, + 0.01951790042221546, + 0.011761403642594814, + 0.020044449716806412, + 0.009791483171284199, + 0.014490576460957527, + 0.01150421891361475, + 0.06625845283269882, + 0.012600364163517952, + 0.04603762552142143, + 0.059483036398887634, + 0.027081722393631935, + 0.004848385229706764, + 0.004149651620537043, + 0.005580850411206484, + 0.17061467468738556, + 0.008349200710654259, + 0.09100589156150818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.14039362967014313 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched schoolyard with cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four young boys (approx. 8-10 years old) initially huddled in a tight circle playing a game?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do two much taller teenage boys walk through the circle and pose against a weathered brick wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the teenage boys wearing an oversized black hoodie with the hood pulled down and baggy jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with warm golden late afternoon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038126103579998016 + ], + "opposite_logprob_tokens": [ + -3.288125991821289 + ], + "logprob_reward": -0.038126103579998016, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038126103579998016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028423495590686798, + 0.03940536081790924, + 0.0490306131541729, + 0.07317661494016647, + 0.015114832669496536, + 0.015507780015468597, + 0.020704030990600586, + 0.01859433762729168, + 0.009327271953225136, + 0.7307156324386597 + ], + "credit_map": [ + 0.0019603134132921696, + 0.0034492260310798883, + 0.0034854470286518335, + 0.002899301005527377, + 0.006118657998740673, + 0.0036137960851192474, + 0.0031248482409864664, + 0.0037719071842730045, + 0.0033851100597530603, + 0.00396821228787303, + 0.00490952841937542, + 0.0058741457760334015, + 0.006712877657264471, + 0.005912360735237598, + 0.0037410459481179714, + 0.0049020773731172085, + 0.006376961246132851, + 0.00924613606184721, + 0.004527864512056112, + 0.006123894825577736, + 0.007166780065745115, + 0.005889269057661295, + 0.0053866771049797535, + 0.004313031677156687, + 0.007276116404682398, + 0.0074226874858140945, + 0.01054577250033617, + 0.00510039646178484, + 0.008284728042781353, + 0.004891892895102501, + 0.022911373525857925, + 0.006743648089468479, + 0.0021388556342571974, + 0.0022226020228117704, + 0.002145756734535098, + 0.001961490837857127, + 0.0014694418059661984, + 0.0021014201920479536, + 0.0018509221263229847, + 0.001224344945512712, + 0.001739793922752142, + 0.0014504239661619067, + 0.0028477683663368225, + 0.0014690494863316417, + 0.002154499525204301, + 0.002019817940890789, + 0.002003296511247754, + 0.001823130063712597, + 0.002502446062862873, + 0.002476456807926297, + 0.0017551424680277705, + 0.0016347095370292664, + 0.0035352802369743586, + 0.0031158332712948322, + 0.003169095143675804, + 0.0025150675792247057, + 0.0017746197991073132, + 0.0015233345329761505, + 0.0013652037596330047, + 0.0012008886551484466, + 0.0019700501579791307, + 0.005653881933540106, + 0.0025017056614160538, + 0.0026046549901366234, + 0.0010351898381486535, + 0.0010875263251364231, + 0.0009384825825691223, + 0.0009173240978270769, + 0.0008438913500867784, + 0.00219821953214705, + 0.00100591906812042, + 0.0013007193338125944, + 0.11794103682041168, + 0.05262269079685211, + 0.008612516336143017, + 0.008136139251291752, + 0.010160685516893864, + 0.3375018835067749, + 0.01610764116048813, + 0.17963308095932007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 2 + ], + "selected_reward": -0.038126103579998016 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched schoolyard with cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four young boys (approx. 8-10 years old) initially huddled in a tight circle playing a game?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do two much taller teenage boys walk through the circle and pose against a weathered brick wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the teenage boys wearing an oversized black hoodie with the hood pulled down and baggy jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with warm golden late afternoon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029809430241584778 + ], + "opposite_logprob_tokens": [ + -3.5298094749450684 + ], + "logprob_reward": -0.029809430241584778, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029809430241584778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039226122200489044, + 0.014676054008305073, + 0.028770677745342255, + 0.030604299157857895, + 0.02516075223684311, + 0.025235317647457123, + 0.03330255299806595, + 0.08613485842943192, + 0.21350614726543427, + 0.503383219242096 + ], + "credit_map": [ + 0.0036343757528811693, + 0.006169385276734829, + 0.004588440526276827, + 0.0034234472550451756, + 0.00935456808656454, + 0.004243341740220785, + 0.004194133449345827, + 0.0036184305790811777, + 0.0015083093894645572, + 0.0021897037513554096, + 0.0018038940615952015, + 0.002010937314480543, + 0.0023308636154979467, + 0.0018169083632528782, + 0.0017733641434460878, + 0.0012420732527971268, + 0.005819100886583328, + 0.0035056567285209894, + 0.0033323357347398996, + 0.0030602954793721437, + 0.003922518342733383, + 0.002946160500869155, + 0.0037956899031996727, + 0.0023889211006462574, + 0.002524116775020957, + 0.003247953252866864, + 0.0023875704500824213, + 0.0020043719559907913, + 0.0029011976439505816, + 0.007364294026046991, + 0.007418245077133179, + 0.002756550209596753, + 0.0032398696057498455, + 0.002900291234254837, + 0.004277192987501621, + 0.0026315846480429173, + 0.002750718966126442, + 0.0033272330183535814, + 0.002879370702430606, + 0.003154490841552615, + 0.0027555031701922417, + 0.002834114944562316, + 0.003532809903845191, + 0.0029588297475129366, + 0.0038013593293726444, + 0.003183918073773384, + 0.0033439649268984795, + 0.002824817318469286, + 0.0035826198291033506, + 0.003761387662962079, + 0.00452034268528223, + 0.0031256363727152348, + 0.005045998841524124, + 0.004620418883860111, + 0.004410375840961933, + 0.0042357733473181725, + 0.006447953637689352, + 0.007920886389911175, + 0.005477562081068754, + 0.004822276532649994, + 0.005901300814002752, + 0.030773820355534554, + 0.00744132325053215, + 0.0173497311770916, + 0.012196473777294159, + 0.0176265649497509, + 0.008193914778530598, + 0.017048846930265427, + 0.011968616396188736, + 0.07548964768648148, + 0.012003262527287006, + 0.058978818356990814, + 0.08047184348106384, + 0.03646215796470642, + 0.006038866005837917, + 0.0056550647132098675, + 0.008276971988379955, + 0.230821430683136, + 0.011742755770683289, + 0.12391414493322372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.029809430241584778 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched schoolyard with cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four young boys (approx. 8-10 years old) initially huddled in a tight circle playing a game?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do two much taller teenage boys walk through the circle and pose against a weathered brick wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the teenage boys wearing an oversized black hoodie with the hood pulled down and baggy jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with warm golden late afternoon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371216356754303 + ], + "opposite_logprob_tokens": [ + -3.408712148666382 + ], + "logprob_reward": -0.03371216356754303, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371216356754303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04399482533335686, + 0.052463456988334656, + 0.03768613189458847, + 0.023719752207398415, + 0.03906095400452614, + 0.036405306309461594, + 0.2626023292541504, + 0.06504370272159576, + 0.06658586859703064, + 0.3724376857280731 + ], + "credit_map": [ + 0.004830195102840662, + 0.005630519241094589, + 0.0042068543843925, + 0.0045600952580571175, + 0.009013617411255836, + 0.005269134417176247, + 0.00393483554944396, + 0.00654957490041852, + 0.005764486733824015, + 0.006265239790081978, + 0.006299067288637161, + 0.006542150862514973, + 0.007234536111354828, + 0.008347784169018269, + 0.005749102216213942, + 0.006261088885366917, + 0.0052857534028589725, + 0.004116185940802097, + 0.003948034252971411, + 0.004147547297179699, + 0.004066944122314453, + 0.006251664832234383, + 0.005028571002185345, + 0.004841431509703398, + 0.0017727998783811927, + 0.001804988831281662, + 0.0024469082709401846, + 0.0018653224688023329, + 0.00325935916043818, + 0.007027600426226854, + 0.003765009343624115, + 0.0017777648754417896, + 0.003859872929751873, + 0.0045420327223837376, + 0.005495955236256123, + 0.0052827210165560246, + 0.005108397454023361, + 0.004927525296807289, + 0.004904116038233042, + 0.004940334241837263, + 0.003080779453739524, + 0.0043861232697963715, + 0.004125232808291912, + 0.005379124078899622, + 0.0038857164327055216, + 0.004366510547697544, + 0.00533796101808548, + 0.005843857303261757, + 0.027530353516340256, + 0.03396233916282654, + 0.025251977145671844, + 0.0195654034614563, + 0.048962146043777466, + 0.03975915536284447, + 0.04087692126631737, + 0.026694022119045258, + 0.006363567430526018, + 0.006301120854914188, + 0.005354031920433044, + 0.005048316903412342, + 0.006911797448992729, + 0.015153450891375542, + 0.010115345939993858, + 0.009796072728931904, + 0.004813017789274454, + 0.006403143517673016, + 0.003465251764282584, + 0.005456837825477123, + 0.004188974387943745, + 0.021699486300349236, + 0.0043516261503100395, + 0.01620752550661564, + 0.054932139813899994, + 0.028902646154165268, + 0.0147180100902915, + 0.011131398379802704, + 0.01070509385317564, + 0.1531612128019333, + 0.013305364176630974, + 0.08558183908462524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.03371216356754303 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched schoolyard with cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four young boys (approx. 8-10 years old) initially huddled in a tight circle playing a game?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do two much taller teenage boys walk through the circle and pose against a weathered brick wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the teenage boys wearing an oversized black hoodie with the hood pulled down and baggy jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with warm golden late afternoon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129569411277771 + ], + "opposite_logprob_tokens": [ + -2.237957000732422 + ], + "logprob_reward": -0.1129569411277771, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129569411277771, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054251108318567276, + 0.026152193546295166, + 0.03143506124615669, + 0.060735370963811874, + 0.04511447995901108, + 0.03258950635790825, + 0.027848493307828903, + 0.021286511793732643, + 0.026236223056912422, + 0.674351155757904 + ], + "credit_map": [ + 0.005804143846035004, + 0.00898681115359068, + 0.00607358617708087, + 0.005826284177601337, + 0.012162255123257637, + 0.005412112455815077, + 0.005334325134754181, + 0.004651585128158331, + 0.0017795783933252096, + 0.0037892507389187813, + 0.0028576739132404327, + 0.004519462585449219, + 0.005355944391340017, + 0.00346594350412488, + 0.0020164737943559885, + 0.0023678645957261324, + 0.006931744981557131, + 0.0030740012880414724, + 0.0030440064147114754, + 0.0027033791411668062, + 0.006969522684812546, + 0.003203156404197216, + 0.00336350011639297, + 0.0021457450930029154, + 0.005149818956851959, + 0.004342116415500641, + 0.013398568145930767, + 0.00432267040014267, + 0.005803721956908703, + 0.021663526073098183, + 0.002906671492382884, + 0.00314826937392354, + 0.0042948126792907715, + 0.006202553864568472, + 0.007989359088242054, + 0.005996438208967447, + 0.005647741723805666, + 0.005882184486836195, + 0.004704891238361597, + 0.004396494012326002, + 0.0025823505129665136, + 0.003425824223086238, + 0.002922795480117202, + 0.003522701794281602, + 0.0029401856008917093, + 0.00462180795148015, + 0.004973733797669411, + 0.007600104436278343, + 0.0041772183030843735, + 0.0020603190641850233, + 0.003998120315372944, + 0.0023580666165798903, + 0.002613233169540763, + 0.003979465924203396, + 0.003852937603369355, + 0.0048091295175254345, + 0.0019651902839541435, + 0.0015450367936864495, + 0.0017229736549779773, + 0.0018230038695037365, + 0.001909840852022171, + 0.0060613323003053665, + 0.0025664435233920813, + 0.0036926893517374992, + 0.002075426746159792, + 0.0019154775654897094, + 0.0012633617734536529, + 0.0023038883227854967, + 0.0022060335613787174, + 0.007952610030770302, + 0.0018822294659912586, + 0.006637194659560919, + 0.10885391384363174, + 0.048642370849847794, + 0.0079044783487916, + 0.007906582206487656, + 0.009545587934553623, + 0.31114551424980164, + 0.01506606861948967, + 0.165286585688591 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.1129569411277771 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_3/index_tensor([9236, 9236])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_3/index_tensor([9236, 9236])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..19ea40dd29247732832c2c586501ec7962c4e96c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_3/index_tensor([9236, 9236])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, overhead close-up shot focuses on the majestic Monkey King, hovering above a swirling sea of golden clouds. He wears an intricately carved phoenix-wing purple-gold crown, two slender pheasant feathers swaying gently with an invisible celestial aura. He clad in shimmering golden chainmail armor, its plates gleaming brilliantly in the brilliant sunlight. A crimson cape rustles behind him, adding to his heroic presence. He casually slung the heavy Ruyi Jingu Bang (Ruyi Jingu Cudgel) over his right shoulder, its golden hoops gleaming with brilliance. His posture is composed and commanding. A close-up captures his distinctive face, his lips curled slightly in confidence, his golden eyes gleaming with a hint of rebelliousness, a smile both powerful and slightly playful. Brilliant golden light shines down from above, outlining sharp golden edges on his armor and facial features. The background is the vaguely visible jade buildings in the distance. The whole picture is full of epic mythological atmosphere and unparalleled heroism.", + "seed": 1408, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_3/index_9236_seed_1408_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04856735179200768, + "vlm_logprob": -0.04856735179200768, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.04856735179200768, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static overhead close-up focusing on the Monkey King?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of golden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing an intricately carved phoenix-wing purple-gold crown with two slender pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King clad in shimmering golden chainmail armor with a crimson cape rustling behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the heavy Ruyi Jingu Bang slung over the character's right shoulder with golden hoops gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11603734642267227 + ], + "opposite_logprob_tokens": [ + -2.241037368774414 + ], + "logprob_reward": -0.11603734642267227, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11603734642267227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07102814316749573, + 0.08986604958772659, + 0.14285460114479065, + 0.05149053782224655, + 0.2689119577407837, + 0.05203457176685333, + 0.07028115540742874, + 0.09005574882030487, + 0.12034519761800766, + 0.04313202574849129 + ], + "credit_map": [ + 0.006977658253163099, + 0.006416350603103638, + 0.012714261189103127, + 0.009168111719191074, + 0.01239708624780178, + 0.006545816548168659, + 0.009005707688629627, + 0.007803153712302446, + 0.011074507609009743, + 0.010413436219096184, + 0.010795042850077152, + 0.009767678566277027, + 0.0189815741032362, + 0.0069829984568059444, + 0.011008658446371555, + 0.010842151008546352, + 0.012905348092317581, + 0.014667687937617302, + 0.021118806675076485, + 0.01595747284591198, + 0.028487667441368103, + 0.02053060568869114, + 0.015620146878063679, + 0.013566850684583187, + 0.006589606404304504, + 0.004664245061576366, + 0.006015625316649675, + 0.004957765806466341, + 0.007154695224016905, + 0.0076028211042284966, + 0.007513074204325676, + 0.006992703769356012, + 0.03865402936935425, + 0.017068976536393166, + 0.0236346498131752, + 0.010965685360133648, + 0.02320922166109085, + 0.1141347885131836, + 0.016587981954216957, + 0.02465660870075226, + 0.006671522743999958, + 0.0061530000530183315, + 0.009701285511255264, + 0.008872346952557564, + 0.006937575060874224, + 0.004781924653798342, + 0.0059263622388243675, + 0.002990555018186569, + 0.010359907522797585, + 0.006937236525118351, + 0.011493363417685032, + 0.010242835618555546, + 0.009581665508449078, + 0.00756424805149436, + 0.009359074756503105, + 0.0047428300604224205, + 0.012814508751034737, + 0.011763916350901127, + 0.017508843913674355, + 0.010746438056230545, + 0.009110895916819572, + 0.009745733812451363, + 0.012144177220761776, + 0.006221229676157236, + 0.01768367923796177, + 0.017320673912763596, + 0.01170169748365879, + 0.015373309142887592, + 0.014837910421192646, + 0.019448237493634224, + 0.01339513435959816, + 0.010584560222923756, + 0.005531769711524248, + 0.008917964063584805, + 0.003708910895511508, + 0.005691951606422663, + 0.006417274475097656, + 0.004908759146928787, + 0.0046792663633823395, + 0.0032761297188699245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 8 + ], + "selected_reward": -0.11603734642267227 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static overhead close-up focusing on the Monkey King?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of golden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing an intricately carved phoenix-wing purple-gold crown with two slender pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King clad in shimmering golden chainmail armor with a crimson cape rustling behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the heavy Ruyi Jingu Bang slung over the character's right shoulder with golden hoops gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982042171061039 + ], + "opposite_logprob_tokens": [ + -3.529820442199707 + ], + "logprob_reward": -0.02982042171061039, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982042171061039, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21648652851581573, + 0.12158362567424774, + 0.05761679261922836, + 0.12254168093204498, + 0.24062037467956543, + 0.05903767794370651, + 0.030641213059425354, + 0.07485301047563553, + 0.056967515498399734, + 0.01965155266225338 + ], + "credit_map": [ + 0.020915741100907326, + 0.018888220191001892, + 0.0521114245057106, + 0.031049972400069237, + 0.022787094116210938, + 0.018557103350758553, + 0.018905891105532646, + 0.03327108174562454, + 0.0186941996216774, + 0.014790588989853859, + 0.020160097628831863, + 0.012551626190543175, + 0.018148161470890045, + 0.009130685590207577, + 0.011804415844380856, + 0.01630384661257267, + 0.005234458949416876, + 0.006620173808187246, + 0.007289254572242498, + 0.005119025241583586, + 0.011241083964705467, + 0.010135835967957973, + 0.006386542227119207, + 0.005590420216321945, + 0.016499746590852737, + 0.011556696146726608, + 0.02837035432457924, + 0.01955324038863182, + 0.012313060462474823, + 0.015066183172166348, + 0.007917285896837711, + 0.011265106499195099, + 0.029504261910915375, + 0.012495184317231178, + 0.02257726714015007, + 0.01258451584726572, + 0.020709527656435966, + 0.10351542383432388, + 0.016143539920449257, + 0.02309066615998745, + 0.008290410973131657, + 0.005056014284491539, + 0.007453008089214563, + 0.00920031126588583, + 0.007492543198168278, + 0.005961718037724495, + 0.010794430039823055, + 0.00478924298658967, + 0.003551310393959284, + 0.004321683663874865, + 0.004720814060419798, + 0.004536239895969629, + 0.0036160345189273357, + 0.003876961302012205, + 0.003171046730130911, + 0.002847123658284545, + 0.011130739003419876, + 0.009527776390314102, + 0.012477573938667774, + 0.011668476276099682, + 0.007994220592081547, + 0.009168707765638828, + 0.007541273254901171, + 0.0053442432545125484, + 0.009487083181738853, + 0.010003851726651192, + 0.007435234729200602, + 0.005162851419299841, + 0.007518616504967213, + 0.007270687725394964, + 0.005939707159996033, + 0.004149481654167175, + 0.0028282832354307175, + 0.003916003275662661, + 0.0014551356434822083, + 0.002321780426427722, + 0.0022865491919219494, + 0.0020870212465524673, + 0.003014056710526347, + 0.0017427225830033422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.02982042171061039 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static overhead close-up focusing on the Monkey King?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of golden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing an intricately carved phoenix-wing purple-gold crown with two slender pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King clad in shimmering golden chainmail armor with a crimson cape rustling behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the heavy Ruyi Jingu Bang slung over the character's right shoulder with golden hoops gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607259176671505 + ], + "opposite_logprob_tokens": [ + -4.141072750091553 + ], + "logprob_reward": -0.01607259176671505, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607259176671505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1442044973373413, + 0.05493171140551567, + 0.09654262661933899, + 0.03952295333147049, + 0.2595307230949402, + 0.07587185502052307, + 0.11141983419656754, + 0.0829806998372078, + 0.0976082980632782, + 0.037386830896139145 + ], + "credit_map": [ + 0.014961675740778446, + 0.011570862494409084, + 0.028340881690382957, + 0.020353252068161964, + 0.016678176820278168, + 0.014920799992978573, + 0.015585791319608688, + 0.0217930618673563, + 0.006452139467000961, + 0.005465622991323471, + 0.008752766996622086, + 0.005315291229635477, + 0.008671733550727367, + 0.006219925358891487, + 0.005466839764267206, + 0.008587390184402466, + 0.0075362748466432095, + 0.008223406970500946, + 0.01824110746383667, + 0.01270698755979538, + 0.012893722392618656, + 0.01674390211701393, + 0.008848346769809723, + 0.01134887058287859, + 0.00365377776324749, + 0.003662894479930401, + 0.006364517379552126, + 0.004308365285396576, + 0.0050926548428833485, + 0.005728018470108509, + 0.004834286402910948, + 0.005878441035747528, + 0.03480800241231918, + 0.02431689016520977, + 0.02131723426282406, + 0.011733383871614933, + 0.022738546133041382, + 0.10406867414712906, + 0.017268730327486992, + 0.023279264569282532, + 0.009790762327611446, + 0.007675816770642996, + 0.00830432865768671, + 0.007661260198801756, + 0.011448198929429054, + 0.008249158039689064, + 0.017471542581915855, + 0.005270785186439753, + 0.013866325840353966, + 0.014768410474061966, + 0.01907994970679283, + 0.012881902046501637, + 0.01972970739006996, + 0.010920369066298008, + 0.011400004848837852, + 0.008773167617619038, + 0.012327026575803757, + 0.01154462993144989, + 0.012923338450491428, + 0.009314131923019886, + 0.01250082440674305, + 0.008258292451500893, + 0.008775313384830952, + 0.007337143644690514, + 0.015487127006053925, + 0.01958778128027916, + 0.007801900617778301, + 0.009885754436254501, + 0.014565183781087399, + 0.014352157711982727, + 0.009920509532094002, + 0.006007874384522438, + 0.005036911461502314, + 0.004943130537867546, + 0.004655088298022747, + 0.004167579580098391, + 0.005526818800717592, + 0.0032942232210189104, + 0.005450592376291752, + 0.0043124845251441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.01607259176671505 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static overhead close-up focusing on the Monkey King?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of golden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing an intricately carved phoenix-wing purple-gold crown with two slender pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King clad in shimmering golden chainmail armor with a crimson cape rustling behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the heavy Ruyi Jingu Bang slung over the character's right shoulder with golden hoops gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019407500512897968 + ], + "opposite_logprob_tokens": [ + -6.251940727233887 + ], + "logprob_reward": -0.0019407500512897968, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019407500512897968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1709531545639038, + 0.0989103764295578, + 0.18028037250041962, + 0.06652043014764786, + 0.2184249758720398, + 0.05497405678033829, + 0.048323728144168854, + 0.07043147832155228, + 0.061769045889377594, + 0.029412351548671722 + ], + "credit_map": [ + 0.016674986109137535, + 0.012910869903862476, + 0.03822233900427818, + 0.025625187903642654, + 0.020285267382860184, + 0.018520347774028778, + 0.01562528684735298, + 0.023088868707418442, + 0.01393794920295477, + 0.012128750793635845, + 0.013340518809854984, + 0.009739314205944538, + 0.014619129709899426, + 0.013426794670522213, + 0.009108424186706543, + 0.012609494850039482, + 0.01986571215093136, + 0.015994489192962646, + 0.03336518630385399, + 0.02552163042128086, + 0.021847035735845566, + 0.02647637203335762, + 0.016885625198483467, + 0.02032431960105896, + 0.006834710016846657, + 0.0062553780153393745, + 0.008617594838142395, + 0.008467708714306355, + 0.00956969615072012, + 0.00913857202976942, + 0.007560873869806528, + 0.010075900703668594, + 0.03299850970506668, + 0.013389811851084232, + 0.018106810748577118, + 0.008997155353426933, + 0.019667508080601692, + 0.09144728630781174, + 0.013942458666861057, + 0.019875438883900642, + 0.008238531649112701, + 0.005437434185296297, + 0.009056149050593376, + 0.006503861863166094, + 0.007773876190185547, + 0.0068776290863752365, + 0.007097674533724785, + 0.003988900221884251, + 0.007464727386832237, + 0.007354586385190487, + 0.006182102486491203, + 0.006437219679355621, + 0.007499141618609428, + 0.004988808650523424, + 0.004867339041084051, + 0.003529801731929183, + 0.009717519395053387, + 0.011833776719868183, + 0.008409283123910427, + 0.008447961881756783, + 0.00992082804441452, + 0.008329818956553936, + 0.00726056145504117, + 0.006511729676276445, + 0.010119627229869366, + 0.011297844350337982, + 0.009573016315698624, + 0.005864886566996574, + 0.008297336287796497, + 0.006510882638394833, + 0.006598108448088169, + 0.003507340094074607, + 0.003497591707855463, + 0.00389547529630363, + 0.0029687080532312393, + 0.0034138422925025225, + 0.0047487798146903515, + 0.00358049594797194, + 0.004120583180338144, + 0.00318687641993165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.0019407500512897968 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static overhead close-up focusing on the Monkey King?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of golden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing an intricately carved phoenix-wing purple-gold crown with two slender pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King clad in shimmering golden chainmail armor with a crimson cape rustling behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the heavy Ruyi Jingu Bang slung over the character's right shoulder with golden hoops gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896564900875092 + ], + "opposite_logprob_tokens": [ + -2.578965663909912 + ], + "logprob_reward": -0.07896564900875092, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896564900875092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1532137095928192, + 0.0379377081990242, + 0.07434713840484619, + 0.0704674944281578, + 0.42387455701828003, + 0.03377804160118103, + 0.0436864048242569, + 0.038734812289476395, + 0.09646418690681458, + 0.02749594859778881 + ], + "credit_map": [ + 0.011063625104725361, + 0.01189916767179966, + 0.026312261819839478, + 0.026396242901682854, + 0.01864764466881752, + 0.018286198377609253, + 0.019519690424203873, + 0.021088890731334686, + 0.00370030221529305, + 0.0026654882822185755, + 0.00718751922249794, + 0.0036011619959026575, + 0.006895648315548897, + 0.004336994141340256, + 0.003304321551695466, + 0.006246273405849934, + 0.007665825076401234, + 0.013315480202436447, + 0.00796562060713768, + 0.007635083980858326, + 0.01147266011685133, + 0.010031333193182945, + 0.007354310248047113, + 0.008906825445592403, + 0.008168592117726803, + 0.006885780021548271, + 0.008472559042274952, + 0.00761972414329648, + 0.009599512442946434, + 0.009622014127671719, + 0.009150715544819832, + 0.01094860304147005, + 0.05661601945757866, + 0.027978749945759773, + 0.035640887916088104, + 0.016817288473248482, + 0.03687097504734993, + 0.18264524638652802, + 0.027083659544587135, + 0.04022173956036568, + 0.004670954309403896, + 0.0029117579106241465, + 0.005345568526536226, + 0.004140010103583336, + 0.005479753483086824, + 0.004277247469872236, + 0.003470502095296979, + 0.0034822458401322365, + 0.005089478101581335, + 0.004994097165763378, + 0.005606852006167173, + 0.005199710372835398, + 0.005345613230019808, + 0.006039176136255264, + 0.006369500886648893, + 0.005041979253292084, + 0.0048477486707270145, + 0.0037488581147044897, + 0.008920514024794102, + 0.004614407662302256, + 0.005469331983476877, + 0.004853767342865467, + 0.0035285346675664186, + 0.002751653315499425, + 0.014820714481174946, + 0.017688695341348648, + 0.01060480996966362, + 0.009321784600615501, + 0.012788261286914349, + 0.015646904706954956, + 0.00840496551245451, + 0.007188050542026758, + 0.0035974455531686544, + 0.00399770075455308, + 0.0024255188181996346, + 0.002879870356991887, + 0.003908934071660042, + 0.003974789287894964, + 0.002786288969218731, + 0.003925402648746967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.07896564900875092 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, overhead close-up shot focuses on the majestic Monkey King, hovering above a swirling sea of golden clouds. He wears an intricately carved phoenix-wing purple-gold crown, two slender pheasant feathers swaying gently with an invisible celestial aura. He clad in shimmering golden chainmail armor, its plates gleaming brilliantly in the brilliant sunlight. A crimson cape rustles behind him, adding to his heroic presence. He casually slung the heavy Ruyi Jingu Bang (Ruyi Jingu Cudgel) over his right shoulder, its golden hoops gleaming with brilliance. His posture is composed and commanding. A close-up captures his distinctive face, his lips curled slightly in confidence, his golden eyes gleaming with a hint of rebelliousness, a smile both powerful and slightly playful. Brilliant golden light shines down from above, outlining sharp golden edges on his armor and facial features. The background is the vaguely visible jade buildings in the distance. The whole picture is full of epic mythological atmosphere and unparalleled heroism.", + "seed": 1409, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_3/index_9236_seed_1409_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07489723395556211, + "vlm_logprob": -0.07489723395556211, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.07489723395556211, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static overhead close-up focusing on the Monkey King?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of golden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing an intricately carved phoenix-wing purple-gold crown with two slender pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King clad in shimmering golden chainmail armor with a crimson cape rustling behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the heavy Ruyi Jingu Bang slung over the character's right shoulder with golden hoops gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1294412910938263 + ], + "opposite_logprob_tokens": [ + -2.129441261291504 + ], + "logprob_reward": -0.1294412910938263, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1294412910938263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07563500106334686, + 0.14385519921779633, + 0.06101134791970253, + 0.09432514756917953, + 0.13773374259471893, + 0.054509758949279785, + 0.21367639303207397, + 0.07190842181444168, + 0.06057039275765419, + 0.08677463233470917 + ], + "credit_map": [ + 0.008154506795108318, + 0.005780581384897232, + 0.011857261881232262, + 0.012305372394621372, + 0.004749450366944075, + 0.007946417666971684, + 0.009284269995987415, + 0.01555713452398777, + 0.019446153193712234, + 0.009191757999360561, + 0.055387917906045914, + 0.015774594619870186, + 0.008299319073557854, + 0.009289849549531937, + 0.00851829070597887, + 0.017947321757674217, + 0.00504078296944499, + 0.004678552038967609, + 0.010714592412114143, + 0.008942827582359314, + 0.005820123013108969, + 0.007902313023805618, + 0.006545145995914936, + 0.011367013677954674, + 0.007503871340304613, + 0.007205619942396879, + 0.01069070864468813, + 0.009342839941382408, + 0.018279166892170906, + 0.011209935881197453, + 0.015627915039658546, + 0.014465092681348324, + 0.01438991166651249, + 0.011203053407371044, + 0.024130480363965034, + 0.01953839510679245, + 0.016979344189167023, + 0.017459681257605553, + 0.01683218404650688, + 0.017200682312250137, + 0.006720529869198799, + 0.007128643337637186, + 0.010768961161375046, + 0.005940479226410389, + 0.006245897617191076, + 0.004624467343091965, + 0.0072653209790587425, + 0.005815456621348858, + 0.02890661172568798, + 0.0423007607460022, + 0.020822061225771904, + 0.02056657336652279, + 0.02882685698568821, + 0.02443789876997471, + 0.019627481698989868, + 0.028188154101371765, + 0.010215414687991142, + 0.006784314289689064, + 0.011646471917629242, + 0.0052969628013670444, + 0.009366892278194427, + 0.009460248984396458, + 0.008477816358208656, + 0.010660299099981785, + 0.008373447693884373, + 0.0052644796669483185, + 0.011709633283317089, + 0.005057999398559332, + 0.008533576503396034, + 0.005671111401170492, + 0.008031132631003857, + 0.007929009385406971, + 0.011305207386612892, + 0.017641540616750717, + 0.010718554258346558, + 0.00772083317860961, + 0.013590767048299313, + 0.00930143240839243, + 0.010019791312515736, + 0.006476510316133499 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 4 + ], + "selected_reward": -0.1294412910938263 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static overhead close-up focusing on the Monkey King?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of golden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing an intricately carved phoenix-wing purple-gold crown with two slender pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King clad in shimmering golden chainmail armor with a crimson cape rustling behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the heavy Ruyi Jingu Bang slung over the character's right shoulder with golden hoops gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06208929046988487 + ], + "opposite_logprob_tokens": [ + -2.812089204788208 + ], + "logprob_reward": -0.06208929046988487, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06208929046988487, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2647863030433655, + 0.03774722293019295, + 0.058815840631723404, + 0.059537842869758606, + 0.04846109077334404, + 0.28169628977775574, + 0.04050903022289276, + 0.04687323048710823, + 0.028699124231934547, + 0.1328740119934082 + ], + "credit_map": [ + 0.025990920141339302, + 0.013499044813215733, + 0.03772479668259621, + 0.0456729456782341, + 0.014645791612565517, + 0.017496848478913307, + 0.03222481906414032, + 0.07753115147352219, + 0.004731543827801943, + 0.004750213120132685, + 0.0050676544196903706, + 0.006367685738950968, + 0.004778595641255379, + 0.003380702342838049, + 0.0032484843395650387, + 0.005422342102974653, + 0.005041336640715599, + 0.003363741096109152, + 0.00965909380465746, + 0.009229174815118313, + 0.009952902793884277, + 0.005994325969368219, + 0.0076356977224349976, + 0.007939569652080536, + 0.005366163328289986, + 0.004293071571737528, + 0.01078660786151886, + 0.00604885583743453, + 0.007080739364027977, + 0.00599212059751153, + 0.006773505825549364, + 0.013196774758398533, + 0.006143758539110422, + 0.006870863493531942, + 0.008320633322000504, + 0.005495582241564989, + 0.007512266747653484, + 0.005078142508864403, + 0.004643378779292107, + 0.004396461416035891, + 0.033743757754564285, + 0.03978224843740463, + 0.05730057880282402, + 0.034313004463911057, + 0.03403455764055252, + 0.016775794327259064, + 0.025718839839100838, + 0.04002752900123596, + 0.005649757571518421, + 0.006949336268007755, + 0.004262067843228579, + 0.0033831954933702946, + 0.009003100916743279, + 0.002879748586565256, + 0.003966616466641426, + 0.004415207076817751, + 0.0057855648919939995, + 0.005892380606383085, + 0.007310429587960243, + 0.004421590361744165, + 0.00812117476016283, + 0.005216794554144144, + 0.0061934576369822025, + 0.003931835293769836, + 0.003868900239467621, + 0.0028492272831499577, + 0.004472078289836645, + 0.0028475874569267035, + 0.004166198894381523, + 0.0030300032813102007, + 0.0034873278345912695, + 0.003977801650762558, + 0.01757340505719185, + 0.013492265716195107, + 0.0246500913053751, + 0.009431910701096058, + 0.02319619245827198, + 0.015506645664572716, + 0.016291582956910133, + 0.012731927447021008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.06208929046988487 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static overhead close-up focusing on the Monkey King?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of golden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing an intricately carved phoenix-wing purple-gold crown with two slender pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King clad in shimmering golden chainmail armor with a crimson cape rustling behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the heavy Ruyi Jingu Bang slung over the character's right shoulder with golden hoops gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10035194456577301 + ], + "opposite_logprob_tokens": [ + -2.3503520488739014 + ], + "logprob_reward": -0.10035194456577301, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10035194456577301, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10162390023469925, + 0.08837351948022842, + 0.04271995648741722, + 0.0444149523973465, + 0.08574895560741425, + 0.0710119903087616, + 0.14451122283935547, + 0.13508068025112152, + 0.07162689417600632, + 0.21488794684410095 + ], + "credit_map": [ + 0.008481116034090519, + 0.00810773205012083, + 0.02714865654706955, + 0.015329853631556034, + 0.006661606952548027, + 0.007757900282740593, + 0.01235937885940075, + 0.01577765680849552, + 0.011001176200807095, + 0.007581138983368874, + 0.017452595755457878, + 0.009273885749280453, + 0.009448655880987644, + 0.01408053282648325, + 0.009146382100880146, + 0.010389144532382488, + 0.003936280030757189, + 0.00273121683858335, + 0.006758320145308971, + 0.005085282027721405, + 0.003952320199459791, + 0.005732123740017414, + 0.00464519951492548, + 0.009879212826490402, + 0.003934466280043125, + 0.004696578718721867, + 0.00437113456428051, + 0.0038411326240748167, + 0.008937430568039417, + 0.005914574954658747, + 0.006420711055397987, + 0.006298924330621958, + 0.009279624558985233, + 0.008065041154623032, + 0.012966543436050415, + 0.007875053212046623, + 0.01843399927020073, + 0.00943006295710802, + 0.010404043830931187, + 0.009294584393501282, + 0.008873834274709225, + 0.011236307211220264, + 0.011634593829512596, + 0.007066454738378525, + 0.008989968337118626, + 0.005607901141047478, + 0.007492824457585812, + 0.010110109113156796, + 0.019368577748537064, + 0.024134255945682526, + 0.014228712767362595, + 0.010379035025835037, + 0.03341478481888771, + 0.009701269678771496, + 0.02016511559486389, + 0.013119473122060299, + 0.018764754757285118, + 0.011279292404651642, + 0.019642015919089317, + 0.008127614855766296, + 0.02543216198682785, + 0.01693943701684475, + 0.018318116664886475, + 0.01657726988196373, + 0.009277483448386192, + 0.006914621219038963, + 0.013444201089441776, + 0.005851305555552244, + 0.01100967824459076, + 0.00904444232583046, + 0.008644165471196175, + 0.007440994028002024, + 0.028113417327404022, + 0.019034873694181442, + 0.03635565564036369, + 0.014264136552810669, + 0.042186345905065536, + 0.02660846710205078, + 0.027838708832859993, + 0.020486343652009964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.10035194456577301 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static overhead close-up focusing on the Monkey King?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of golden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing an intricately carved phoenix-wing purple-gold crown with two slender pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King clad in shimmering golden chainmail armor with a crimson cape rustling behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the heavy Ruyi Jingu Bang slung over the character's right shoulder with golden hoops gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036202622577548027 + ], + "opposite_logprob_tokens": [ + -5.628620147705078 + ], + "logprob_reward": -0.0036202622577548027, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036202622577548027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13105811178684235, + 0.06822509318590164, + 0.06720093637704849, + 0.17537462711334229, + 0.05806311219930649, + 0.1517200618982315, + 0.07773363590240479, + 0.06677921861410141, + 0.17087246477603912, + 0.03297281637787819 + ], + "credit_map": [ + 0.012960086576640606, + 0.014240428805351257, + 0.02853666990995407, + 0.018011322245001793, + 0.011045648716390133, + 0.010706707835197449, + 0.018533259630203247, + 0.01702398806810379, + 0.008128147572278976, + 0.006562648341059685, + 0.007812591269612312, + 0.007256499957293272, + 0.009393060579895973, + 0.01205196138471365, + 0.008480794727802277, + 0.008539380505681038, + 0.006342039909213781, + 0.004898316226899624, + 0.008981973864138126, + 0.007982785813510418, + 0.0070129758678376675, + 0.008110256865620613, + 0.010227429680526257, + 0.013645155355334282, + 0.013427386991679668, + 0.013537363149225712, + 0.019307328388094902, + 0.016003165394067764, + 0.04593386873602867, + 0.015652021393179893, + 0.02148330956697464, + 0.03003017045557499, + 0.0064193326979875565, + 0.007526557892560959, + 0.006452507339417934, + 0.006739193107932806, + 0.006448163650929928, + 0.005491076037287712, + 0.0075195711106061935, + 0.01146670337766409, + 0.018015917390584946, + 0.024913575500249863, + 0.023414693772792816, + 0.017739152535796165, + 0.020701171830296516, + 0.015619267709553242, + 0.014288735575973988, + 0.017027538269758224, + 0.011125645600259304, + 0.011437499895691872, + 0.00787391047924757, + 0.00863416027277708, + 0.014856263063848019, + 0.005828112829476595, + 0.007579809054732323, + 0.010398229584097862, + 0.009349475614726543, + 0.008536706678569317, + 0.010963859036564827, + 0.0064661456272006035, + 0.008186145685613155, + 0.008166984654963017, + 0.006061718799173832, + 0.00904818158596754, + 0.02340400591492653, + 0.01757785864174366, + 0.03112606517970562, + 0.013149084523320198, + 0.025209365412592888, + 0.020131388679146767, + 0.021158719435334206, + 0.019115949049592018, + 0.0042413268238306046, + 0.005230315495282412, + 0.004843628965318203, + 0.0027665579691529274, + 0.004742095246911049, + 0.003475094446912408, + 0.0036389546003192663, + 0.004034838173538446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 5 + ], + "selected_reward": -0.0036202622577548027 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static overhead close-up focusing on the Monkey King?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of golden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing an intricately carved phoenix-wing purple-gold crown with two slender pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King clad in shimmering golden chainmail armor with a crimson cape rustling behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the heavy Ruyi Jingu Bang slung over the character's right shoulder with golden hoops gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789833813905716 + ], + "opposite_logprob_tokens": [ + -2.5789833068847656 + ], + "logprob_reward": -0.0789833813905716, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789833813905716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10684715211391449, + 0.09605734050273895, + 0.2129218429327011, + 0.15289010107517242, + 0.14022196829319, + 0.09390144050121307, + 0.051559604704380035, + 0.03946736082434654, + 0.08816760033369064, + 0.0179655272513628 + ], + "credit_map": [ + 0.00973830372095108, + 0.01050140056759119, + 0.023870930075645447, + 0.015668900683522224, + 0.00841074250638485, + 0.01037719938904047, + 0.015701068565249443, + 0.012578600086271763, + 0.01595490425825119, + 0.0069551486521959305, + 0.01622961089015007, + 0.008593328297138214, + 0.007537519093602896, + 0.010084785521030426, + 0.006965542212128639, + 0.023736506700515747, + 0.018935177475214005, + 0.015918590128421783, + 0.04260379448533058, + 0.04003601521253586, + 0.017918629571795464, + 0.02020668052136898, + 0.024932043626904488, + 0.0323709174990654, + 0.011571009643375874, + 0.010315435007214546, + 0.020576197654008865, + 0.021037261933088303, + 0.03260635957121849, + 0.01253083162009716, + 0.019516441971063614, + 0.024736568331718445, + 0.014776557683944702, + 0.019976384937763214, + 0.01785389333963394, + 0.01971747726202011, + 0.02253032475709915, + 0.013174175284802914, + 0.015536283142864704, + 0.01665688119828701, + 0.010427400469779968, + 0.013828213326632977, + 0.015222110785543919, + 0.012931627221405506, + 0.010467383079230785, + 0.009480699896812439, + 0.009332551620900631, + 0.012211458757519722, + 0.006472911220043898, + 0.005263078957796097, + 0.008804171346127987, + 0.004818078596144915, + 0.007796036079525948, + 0.005659331567585468, + 0.005706931930035353, + 0.007039063144475222, + 0.005503186956048012, + 0.004150843247771263, + 0.006547441706061363, + 0.0031871867831796408, + 0.00509295891970396, + 0.004560789559036493, + 0.003978394903242588, + 0.006446557119488716, + 0.011938255280256271, + 0.010534902103245258, + 0.01586170122027397, + 0.007561668287962675, + 0.012533623725175858, + 0.010977737605571747, + 0.00996073242276907, + 0.00879898015409708, + 0.002404291182756424, + 0.001705914386548102, + 0.003030596300959587, + 0.0013608220033347607, + 0.0035397594328969717, + 0.0016111781587824225, + 0.0023690725211054087, + 0.0019438921008259058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 4 + ], + "selected_reward": -0.0789833813905716 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, overhead close-up shot focuses on the majestic Monkey King, hovering above a swirling sea of golden clouds. He wears an intricately carved phoenix-wing purple-gold crown, two slender pheasant feathers swaying gently with an invisible celestial aura. He clad in shimmering golden chainmail armor, its plates gleaming brilliantly in the brilliant sunlight. A crimson cape rustles behind him, adding to his heroic presence. He casually slung the heavy Ruyi Jingu Bang (Ruyi Jingu Cudgel) over his right shoulder, its golden hoops gleaming with brilliance. His posture is composed and commanding. A close-up captures his distinctive face, his lips curled slightly in confidence, his golden eyes gleaming with a hint of rebelliousness, a smile both powerful and slightly playful. Brilliant golden light shines down from above, outlining sharp golden edges on his armor and facial features. The background is the vaguely visible jade buildings in the distance. The whole picture is full of epic mythological atmosphere and unparalleled heroism.", + "seed": 1410, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_3/index_9236_seed_1410_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08357186615467072, + "vlm_logprob": -0.08357186615467072, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.08357186615467072, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static overhead close-up focusing on the Monkey King?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of golden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing an intricately carved phoenix-wing purple-gold crown with two slender pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King clad in shimmering golden chainmail armor with a crimson cape rustling behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the heavy Ruyi Jingu Bang slung over the character's right shoulder with golden hoops gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1621786504983902 + ], + "opposite_logprob_tokens": [ + -1.912178635597229 + ], + "logprob_reward": -0.1621786504983902, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1621786504983902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16549475491046906, + 0.08317027986049652, + 0.05038108304142952, + 0.05718384310603142, + 0.13959641754627228, + 0.04671258106827736, + 0.04542165622115135, + 0.11473195999860764, + 0.18729552626609802, + 0.11001183092594147 + ], + "credit_map": [ + 0.019967399537563324, + 0.013071524910628796, + 0.01985882595181465, + 0.016612466424703598, + 0.02368050254881382, + 0.017109835520386696, + 0.029821135103702545, + 0.025373084470629692, + 0.006823204457759857, + 0.006859125103801489, + 0.011916635558009148, + 0.009391846135258675, + 0.01614893600344658, + 0.008196279406547546, + 0.01217914093285799, + 0.011655105277895927, + 0.004874018952250481, + 0.006741659715771675, + 0.0059213414788246155, + 0.007325784303247929, + 0.010201907716691494, + 0.005142161622643471, + 0.004997530020773411, + 0.005176681559532881, + 0.008850716054439545, + 0.00590771809220314, + 0.006425437051802874, + 0.006646978203207254, + 0.008183885365724564, + 0.009270542301237583, + 0.00609766086563468, + 0.00580090144649148, + 0.012213752605021, + 0.017657959833741188, + 0.01925017312169075, + 0.017882149666547775, + 0.034844450652599335, + 0.00993097759783268, + 0.01580004394054413, + 0.01201691199094057, + 0.007324893958866596, + 0.004826442338526249, + 0.005823911167681217, + 0.005308563355356455, + 0.005346446298062801, + 0.004808112978935242, + 0.008219332434237003, + 0.005054880399256945, + 0.0046503543853759766, + 0.004537347238510847, + 0.006390833295881748, + 0.006904250476509333, + 0.006697159726172686, + 0.004202237352728844, + 0.004214688204228878, + 0.007824789732694626, + 0.01088064443320036, + 0.012463221326470375, + 0.010468787513673306, + 0.008295505307614803, + 0.020763756707310677, + 0.015689881518483162, + 0.015873296186327934, + 0.020296873524785042, + 0.01794726587831974, + 0.025306882336735725, + 0.022521456703543663, + 0.016162393614649773, + 0.05184587836265564, + 0.019217170774936676, + 0.01724168099462986, + 0.017052801325917244, + 0.012651994824409485, + 0.022327015176415443, + 0.014254322275519371, + 0.012554531916975975, + 0.01775374636054039, + 0.013668769970536232, + 0.008119319565594196, + 0.00868213176727295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.1621786504983902 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static overhead close-up focusing on the Monkey King?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of golden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing an intricately carved phoenix-wing purple-gold crown with two slender pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King clad in shimmering golden chainmail armor with a crimson cape rustling behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the heavy Ruyi Jingu Bang slung over the character's right shoulder with golden hoops gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270873248577118 + ], + "opposite_logprob_tokens": [ + -2.127087354660034 + ], + "logprob_reward": -0.1270873248577118, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270873248577118, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3988267481327057, + 0.06282099336385727, + 0.045520588755607605, + 0.12887902557849884, + 0.11477609723806381, + 0.01639077067375183, + 0.022867336869239807, + 0.0440535731613636, + 0.05340483784675598, + 0.11246000975370407 + ], + "credit_map": [ + 0.03321655094623566, + 0.03677972033619881, + 0.04729417338967323, + 0.04155782237648964, + 0.06626798957586288, + 0.0473196767270565, + 0.07461146265268326, + 0.0517793633043766, + 0.004478317219763994, + 0.005820420570671558, + 0.008824863471090794, + 0.007700228597968817, + 0.011647425591945648, + 0.006151816342025995, + 0.00941645447164774, + 0.008781472221016884, + 0.003961100708693266, + 0.005399332381784916, + 0.006229223683476448, + 0.008740217424929142, + 0.008638015016913414, + 0.003946735057979822, + 0.005296641029417515, + 0.003309322288259864, + 0.02133827470242977, + 0.017095543444156647, + 0.01580624096095562, + 0.017703739926218987, + 0.014301512390375137, + 0.01758044771850109, + 0.015161621384322643, + 0.009891648776829243, + 0.010750673711299896, + 0.013988149352371693, + 0.016828248277306557, + 0.014388932846486568, + 0.029537588357925415, + 0.0067955381236970425, + 0.012639856897294521, + 0.009847116656601429, + 0.0018759104423224926, + 0.002021047519519925, + 0.0019056496676057577, + 0.0019638489466160536, + 0.002062843879684806, + 0.0022854749113321304, + 0.0021868590265512466, + 0.00208913697861135, + 0.0032359594479203224, + 0.003242042614147067, + 0.0018485465552657843, + 0.003450923366472125, + 0.0037217196077108383, + 0.0019126132829114795, + 0.002784359734505415, + 0.0026711714453995228, + 0.00625977898016572, + 0.005682448856532574, + 0.004508313722908497, + 0.0029552639462053776, + 0.006558299995958805, + 0.005620636977255344, + 0.005077774170786142, + 0.007391061168164015, + 0.006831797771155834, + 0.006844428833574057, + 0.007815934717655182, + 0.005098268389701843, + 0.010236681438982487, + 0.006063634529709816, + 0.0037587357219308615, + 0.006755354348570108, + 0.010837739333510399, + 0.01985960453748703, + 0.016925420612096786, + 0.01250125840306282, + 0.01858438178896904, + 0.015342322178184986, + 0.008807751350104809, + 0.009601536206901073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.1270873248577118 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static overhead close-up focusing on the Monkey King?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of golden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing an intricately carved phoenix-wing purple-gold crown with two slender pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King clad in shimmering golden chainmail armor with a crimson cape rustling behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the heavy Ruyi Jingu Bang slung over the character's right shoulder with golden hoops gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495971441268921 + ], + "opposite_logprob_tokens": [ + -2.929959774017334 + ], + "logprob_reward": -0.05495971441268921, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495971441268921, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2953341603279114, + 0.12066356837749481, + 0.07368876785039902, + 0.05581942945718765, + 0.04515029489994049, + 0.037042081356048584, + 0.04345808923244476, + 0.07397053390741348, + 0.10447489470243454, + 0.15039828419685364 + ], + "credit_map": [ + 0.02803114242851734, + 0.02712266892194748, + 0.035578999668359756, + 0.02849130518734455, + 0.046947162598371506, + 0.03509830683469772, + 0.055315058678388596, + 0.03874948248267174, + 0.009169574826955795, + 0.012823250144720078, + 0.0169072262942791, + 0.014872051775455475, + 0.02188168466091156, + 0.010509196668863297, + 0.016601722687482834, + 0.01789885014295578, + 0.014571543782949448, + 0.00738608418032527, + 0.011784524656832218, + 0.009209146723151207, + 0.009546304121613503, + 0.006799457594752312, + 0.006885637994855642, + 0.00750606507062912, + 0.010410365648567677, + 0.007266692817211151, + 0.006932728923857212, + 0.006872090976685286, + 0.006345017813146114, + 0.007289922796189785, + 0.006497160065919161, + 0.004205445758998394, + 0.007739272899925709, + 0.004552809987217188, + 0.007795342709869146, + 0.003406872507184744, + 0.011828646063804626, + 0.0026999746914952993, + 0.004307611845433712, + 0.0028197607025504112, + 0.0030666470993310213, + 0.004352430813014507, + 0.003601474454626441, + 0.004138650838285685, + 0.005997639615088701, + 0.004172879736870527, + 0.00825035572052002, + 0.003462000982835889, + 0.003929349593818188, + 0.005577482748776674, + 0.0032351412810385227, + 0.006939122918993235, + 0.008925898000597954, + 0.003589379135519266, + 0.0065619442611932755, + 0.004699766170233488, + 0.007132077123969793, + 0.009131214581429958, + 0.005823106970638037, + 0.005509196314960718, + 0.020936734974384308, + 0.009313791990280151, + 0.008185415528714657, + 0.007938986644148827, + 0.009754125960171223, + 0.012324517592787743, + 0.012032730504870415, + 0.009048283100128174, + 0.030696507543325424, + 0.011269678361713886, + 0.010412294417619705, + 0.008936744183301926, + 0.017424745485186577, + 0.02361443266272545, + 0.01802102103829384, + 0.018996015191078186, + 0.020518941804766655, + 0.018335990607738495, + 0.01864468865096569, + 0.014842439442873001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.05495971441268921 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static overhead close-up focusing on the Monkey King?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of golden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing an intricately carved phoenix-wing purple-gold crown with two slender pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King clad in shimmering golden chainmail armor with a crimson cape rustling behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the heavy Ruyi Jingu Bang slung over the character's right shoulder with golden hoops gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003616936504840851 + ], + "opposite_logprob_tokens": [ + -5.628616809844971 + ], + "logprob_reward": -0.003616936504840851, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003616936504840851, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3511524498462677, + 0.031498413532972336, + 0.04561170935630798, + 0.07870300114154816, + 0.04370700940489769, + 0.022320520132780075, + 0.04342924430966377, + 0.04002208262681961, + 0.08353220671415329, + 0.2600233554840088 + ], + "credit_map": [ + 0.03071451373398304, + 0.03160044178366661, + 0.042063046246767044, + 0.03713235259056091, + 0.0574679933488369, + 0.040256209671497345, + 0.06543409824371338, + 0.04648377373814583, + 0.005181432701647282, + 0.0025434033013880253, + 0.004416546784341335, + 0.0037300169933587313, + 0.004154853988438845, + 0.0035183331929147243, + 0.0040155695751309395, + 0.003938254434615374, + 0.007384646683931351, + 0.005100727081298828, + 0.005274361465126276, + 0.005694363731890917, + 0.00815922673791647, + 0.004035612568259239, + 0.005715987179428339, + 0.004246785305440426, + 0.015500855632126331, + 0.009078974835574627, + 0.009684642776846886, + 0.008760068565607071, + 0.006993580609560013, + 0.014680691994726658, + 0.008055802434682846, + 0.005948374513536692, + 0.005489420145750046, + 0.0057121929712593555, + 0.005001432728022337, + 0.004781176336109638, + 0.009609959088265896, + 0.0036538969725370407, + 0.0054938592948019505, + 0.003965070471167564, + 0.0020032089669257402, + 0.0028403664473444223, + 0.002398047363385558, + 0.002493442501872778, + 0.00349297933280468, + 0.0021562206093221903, + 0.004510188940912485, + 0.0024260657373815775, + 0.004285082221031189, + 0.004155268892645836, + 0.005397493951022625, + 0.00563072832301259, + 0.007133340463042259, + 0.00466035446152091, + 0.005654471460729837, + 0.006512505933642387, + 0.004834463819861412, + 0.004301555920392275, + 0.003940758761018515, + 0.0031469278037548065, + 0.006270421203225851, + 0.006323980167508125, + 0.005872095003724098, + 0.005331879947334528, + 0.009997652843594551, + 0.008667646907269955, + 0.010722897946834564, + 0.009506922215223312, + 0.016762740910053253, + 0.009615263901650906, + 0.008922016248106956, + 0.009337065741419792, + 0.03305211663246155, + 0.05272809788584709, + 0.04193926975131035, + 0.02351798675954342, + 0.04125998914241791, + 0.031068969517946243, + 0.019653720781207085, + 0.01680322177708149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.003616936504840851 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static overhead close-up focusing on the Monkey King?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of golden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing an intricately carved phoenix-wing purple-gold crown with two slender pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King clad in shimmering golden chainmail armor with a crimson cape rustling behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the heavy Ruyi Jingu Bang slung over the character's right shoulder with golden hoops gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001670449972153 + ], + "opposite_logprob_tokens": [ + -2.695016622543335 + ], + "logprob_reward": -0.07001670449972153, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001670449972153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2403385043144226, + 0.1167742908000946, + 0.039964403957128525, + 0.13512898981571198, + 0.04737108200788498, + 0.11156713217496872, + 0.05308392643928528, + 0.026700180023908615, + 0.05554952472448349, + 0.1735219657421112 + ], + "credit_map": [ + 0.019913123920559883, + 0.021988507360219955, + 0.028466632589697838, + 0.025198400020599365, + 0.04088309407234192, + 0.02909926511347294, + 0.04403025284409523, + 0.030759209766983986, + 0.006799743510782719, + 0.012155475094914436, + 0.016019823029637337, + 0.014728358946740627, + 0.021914247423410416, + 0.010579811409115791, + 0.01738700084388256, + 0.017189830541610718, + 0.00980453472584486, + 0.004071489907801151, + 0.005673075094819069, + 0.002822891343384981, + 0.006906560622155666, + 0.004152549430727959, + 0.0040268003940582275, + 0.00250650173984468, + 0.023141631856560707, + 0.012963294051587582, + 0.015358746983110905, + 0.017320021986961365, + 0.0161125548183918, + 0.02339966595172882, + 0.012395760044455528, + 0.01443732064217329, + 0.008610771037638187, + 0.004156045615673065, + 0.004557199776172638, + 0.005144518334418535, + 0.005313211120665073, + 0.006311238277703524, + 0.00453179981559515, + 0.008746296167373657, + 0.009210777468979359, + 0.01415970828384161, + 0.01131469663232565, + 0.012129505164921284, + 0.014219320379197598, + 0.017108218744397163, + 0.019174771383404732, + 0.014250130392611027, + 0.006625337991863489, + 0.0044355629943311214, + 0.006114903371781111, + 0.006467646919190884, + 0.00803336314857006, + 0.005059259012341499, + 0.00537364836782217, + 0.010974203236401081, + 0.002490769373252988, + 0.00348662119358778, + 0.0022317387629300356, + 0.001885287812910974, + 0.006138037424534559, + 0.0038766900543123484, + 0.002420675940811634, + 0.0041703591123223305, + 0.007631752174347639, + 0.008260571397840977, + 0.007432785350829363, + 0.0052646612748503685, + 0.00941668450832367, + 0.007089786231517792, + 0.005961612798273563, + 0.004491669591516256, + 0.022338178008794785, + 0.021246634423732758, + 0.025824671611189842, + 0.017898892983794212, + 0.015645647421479225, + 0.02350570447742939, + 0.014461859129369259, + 0.03260039538145065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.07001670449972153 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, overhead close-up shot focuses on the majestic Monkey King, hovering above a swirling sea of golden clouds. He wears an intricately carved phoenix-wing purple-gold crown, two slender pheasant feathers swaying gently with an invisible celestial aura. He clad in shimmering golden chainmail armor, its plates gleaming brilliantly in the brilliant sunlight. A crimson cape rustles behind him, adding to his heroic presence. He casually slung the heavy Ruyi Jingu Bang (Ruyi Jingu Cudgel) over his right shoulder, its golden hoops gleaming with brilliance. His posture is composed and commanding. A close-up captures his distinctive face, his lips curled slightly in confidence, his golden eyes gleaming with a hint of rebelliousness, a smile both powerful and slightly playful. Brilliant golden light shines down from above, outlining sharp golden edges on his armor and facial features. The background is the vaguely visible jade buildings in the distance. The whole picture is full of epic mythological atmosphere and unparalleled heroism.", + "seed": 1411, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_3/index_9236_seed_1411_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04546421067789197, + "vlm_logprob": -0.04546421067789197, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.04546421067789197, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static overhead close-up focusing on the Monkey King?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of golden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing an intricately carved phoenix-wing purple-gold crown with two slender pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King clad in shimmering golden chainmail armor with a crimson cape rustling behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the heavy Ruyi Jingu Bang slung over the character's right shoulder with golden hoops gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0717197135090828 + ], + "opposite_logprob_tokens": [ + -2.6967196464538574 + ], + "logprob_reward": -0.0717197135090828, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0717197135090828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11924990266561508, + 0.10194011777639389, + 0.09933631867170334, + 0.13986174762248993, + 0.08749492466449738, + 0.09629841893911362, + 0.10323361307382584, + 0.10264726728200912, + 0.05255984142422676, + 0.09737785905599594 + ], + "credit_map": [ + 0.013774813152849674, + 0.011989844031631947, + 0.021672304719686508, + 0.014085705392062664, + 0.014227555133402348, + 0.010683268308639526, + 0.012877974659204483, + 0.01993843913078308, + 0.006536803673952818, + 0.008913529105484486, + 0.009183607064187527, + 0.020962072536349297, + 0.008612260222434998, + 0.0071360869333148, + 0.01142801158130169, + 0.029167750850319862, + 0.008194849826395512, + 0.009625496342778206, + 0.018608389422297478, + 0.013491077348589897, + 0.009470384567975998, + 0.01014486514031887, + 0.012770886532962322, + 0.017030369490385056, + 0.008985897526144981, + 0.014492318034172058, + 0.013539876788854599, + 0.029074721038341522, + 0.014043192379176617, + 0.030769789591431618, + 0.01132137794047594, + 0.017634578049182892, + 0.008976873941719532, + 0.006486673839390278, + 0.013974138535559177, + 0.010160420089960098, + 0.009302971884608269, + 0.008642022497951984, + 0.012771973386406898, + 0.01717984862625599, + 0.010768886655569077, + 0.008928713388741016, + 0.02667473815381527, + 0.00777136767283082, + 0.011178125627338886, + 0.011548501439392567, + 0.010415354743599892, + 0.009012737311422825, + 0.010539050213992596, + 0.014955422841012478, + 0.019098931923508644, + 0.011641927063465118, + 0.015991460531949997, + 0.008644475601613522, + 0.012774959206581116, + 0.009587393142282963, + 0.01044903788715601, + 0.014245809987187386, + 0.016272341832518578, + 0.00918828696012497, + 0.01449307519942522, + 0.00908644124865532, + 0.014680773951113224, + 0.014231501147150993, + 0.00525704538449645, + 0.006520495284348726, + 0.007329617161303759, + 0.0037869811058044434, + 0.008404574356973171, + 0.005112056154757738, + 0.005927267950028181, + 0.01022180262953043, + 0.02730400115251541, + 0.006856522057205439, + 0.016283081844449043, + 0.0042006466537714005, + 0.004962367005646229, + 0.006121669430285692, + 0.007058946881443262, + 0.024590617045760155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.0717197135090828 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static overhead close-up focusing on the Monkey King?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of golden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing an intricately carved phoenix-wing purple-gold crown with two slender pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King clad in shimmering golden chainmail armor with a crimson cape rustling behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the heavy Ruyi Jingu Bang slung over the character's right shoulder with golden hoops gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868520796298981 + ], + "opposite_logprob_tokens": [ + -3.048685312271118 + ], + "logprob_reward": -0.04868520796298981, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868520796298981, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17170462012290955, + 0.11379168927669525, + 0.2585553526878357, + 0.03076641634106636, + 0.04753786325454712, + 0.07299283891916275, + 0.027114376425743103, + 0.06754209101200104, + 0.1045435220003128, + 0.10545122623443604 + ], + "credit_map": [ + 0.020400622859597206, + 0.021100964397192, + 0.0208563432097435, + 0.024918988347053528, + 0.020817190408706665, + 0.018457060679793358, + 0.0256209634244442, + 0.019532492384314537, + 0.006842664908617735, + 0.008559301495552063, + 0.009542291052639484, + 0.03636893630027771, + 0.00638606259599328, + 0.006954811979085207, + 0.007660491392016411, + 0.03147711977362633, + 0.018344203010201454, + 0.023405324667692184, + 0.03863905370235443, + 0.04082447290420532, + 0.031827811151742935, + 0.03394726291298866, + 0.035549622029066086, + 0.036017581820487976, + 0.0025225766003131866, + 0.0029818210750818253, + 0.00420404365286231, + 0.004693714901804924, + 0.005912583786994219, + 0.0031926108058542013, + 0.0036509667988866568, + 0.003608100116252899, + 0.005576273892074823, + 0.0049858056008815765, + 0.014630966819822788, + 0.004265144467353821, + 0.004554347135126591, + 0.0030798998195677996, + 0.004640768747776747, + 0.0058046551421284676, + 0.0068573616445064545, + 0.009720991365611553, + 0.01723487861454487, + 0.006221186835318804, + 0.009495184756815434, + 0.008395402692258358, + 0.009129829704761505, + 0.005938004236668348, + 0.0021480503492057323, + 0.0029869189020246267, + 0.004833288490772247, + 0.0029602632857859135, + 0.005110672675073147, + 0.002986383391544223, + 0.0031732639763504267, + 0.0029155355878174305, + 0.005784742534160614, + 0.008549872785806656, + 0.01198974996805191, + 0.0061439587734639645, + 0.00785126630216837, + 0.006005376111716032, + 0.007563351187855005, + 0.0136537691578269, + 0.009303364902734756, + 0.008010447956621647, + 0.023397445678710938, + 0.006191920023411512, + 0.011146115139126778, + 0.009184593334794044, + 0.009697100147604942, + 0.027612539008259773, + 0.005668343510478735, + 0.005359261762350798, + 0.020145054906606674, + 0.0047880844213068485, + 0.004442273173481226, + 0.005051269195973873, + 0.01549599226564169, + 0.04450094699859619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.04868520796298981 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static overhead close-up focusing on the Monkey King?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of golden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing an intricately carved phoenix-wing purple-gold crown with two slender pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King clad in shimmering golden chainmail armor with a crimson cape rustling behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the heavy Ruyi Jingu Bang slung over the character's right shoulder with golden hoops gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033707089722156525 + ], + "opposite_logprob_tokens": [ + -3.4087071418762207 + ], + "logprob_reward": -0.033707089722156525, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033707089722156525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09983489662408829, + 0.12182606756687164, + 0.11626211553812027, + 0.2529469430446625, + 0.08846455812454224, + 0.11885230988264084, + 0.03468247875571251, + 0.06073128432035446, + 0.07944075763225555, + 0.026958607137203217 + ], + "credit_map": [ + 0.010944712907075882, + 0.012576380744576454, + 0.013210287317633629, + 0.01943311281502247, + 0.010702504776418209, + 0.0076883286237716675, + 0.010902921669185162, + 0.014376644045114517, + 0.006760569289326668, + 0.008638039231300354, + 0.010874206200242043, + 0.027974722906947136, + 0.01053637731820345, + 0.008840476162731647, + 0.010512121953070164, + 0.037689562886953354, + 0.010062778368592262, + 0.014505757950246334, + 0.015223423019051552, + 0.01401639636605978, + 0.014066504314541817, + 0.016013745218515396, + 0.013767319731414318, + 0.01860620453953743, + 0.02487841434776783, + 0.029856234788894653, + 0.029690925031900406, + 0.030637364834547043, + 0.03540872782468796, + 0.03729073330760002, + 0.023309700191020966, + 0.04187484458088875, + 0.007179487496614456, + 0.009780457243323326, + 0.016579873859882355, + 0.009461048990488052, + 0.015185016207396984, + 0.011685341596603394, + 0.008481010794639587, + 0.010112319141626358, + 0.00975954718887806, + 0.013445965014398098, + 0.02092757821083069, + 0.012801690958440304, + 0.01567729189991951, + 0.016256412491202354, + 0.016624797135591507, + 0.01335903163999319, + 0.003637955291196704, + 0.006708985660225153, + 0.005401985254138708, + 0.002873288467526436, + 0.005180587060749531, + 0.003410486038774252, + 0.0033666063100099564, + 0.004102584905922413, + 0.004592894576489925, + 0.0077548460103571415, + 0.015704285353422165, + 0.004623345099389553, + 0.005941558163613081, + 0.0035051850136369467, + 0.009484142996370792, + 0.009125027805566788, + 0.007411367725580931, + 0.008751101791858673, + 0.014559073373675346, + 0.005413897335529327, + 0.008965580724179745, + 0.010862925089895725, + 0.008523689582943916, + 0.014953119680285454, + 0.0037072342820465565, + 0.003237633965909481, + 0.006062062922865152, + 0.0016269944608211517, + 0.0028378956485539675, + 0.0028329023625701666, + 0.002169431187212467, + 0.00448444951325655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 5 + ], + "selected_reward": -0.033707089722156525 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static overhead close-up focusing on the Monkey King?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of golden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing an intricately carved phoenix-wing purple-gold crown with two slender pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King clad in shimmering golden chainmail armor with a crimson cape rustling behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the heavy Ruyi Jingu Bang slung over the character's right shoulder with golden hoops gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003192449454218149 + ], + "opposite_logprob_tokens": [ + -5.75319242477417 + ], + "logprob_reward": -0.003192449454218149, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003192449454218149, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10689320415258408, + 0.1229078620672226, + 0.11062923818826675, + 0.2382390797138214, + 0.06168094277381897, + 0.04733587056398392, + 0.04887717589735985, + 0.04748469591140747, + 0.03545171394944191, + 0.18050022423267365 + ], + "credit_map": [ + 0.009041361510753632, + 0.010812053456902504, + 0.015786880627274513, + 0.015542714856564999, + 0.016176333650946617, + 0.011468796990811825, + 0.015034139156341553, + 0.013030917383730412, + 0.006927772890776396, + 0.012028159573674202, + 0.010980254970490932, + 0.03130851686000824, + 0.007192825898528099, + 0.008976036682724953, + 0.008591946214437485, + 0.036902349442243576, + 0.00921572931110859, + 0.010567511431872845, + 0.01892649009823799, + 0.01547669991850853, + 0.009728781878948212, + 0.011826138012111187, + 0.01484293956309557, + 0.020044943317770958, + 0.020924415439367294, + 0.02986382693052292, + 0.029583998024463654, + 0.022428231313824654, + 0.037420257925987244, + 0.04284150153398514, + 0.023198436945676804, + 0.03197840228676796, + 0.005438060965389013, + 0.005232722498476505, + 0.011371735483407974, + 0.008559124544262886, + 0.009513243101537228, + 0.0068504177033901215, + 0.007000614423304796, + 0.007715030573308468, + 0.004894276615232229, + 0.004654231481254101, + 0.013349371962249279, + 0.004294050391763449, + 0.004006688483059406, + 0.00460879597812891, + 0.005929069593548775, + 0.005599386990070343, + 0.004562146496027708, + 0.005922687239944935, + 0.010315018706023693, + 0.0050561572425067425, + 0.008100028149783611, + 0.005223357118666172, + 0.0054317801259458065, + 0.004265999887138605, + 0.003037102287635207, + 0.0072935218922793865, + 0.006329311057925224, + 0.0036626048386096954, + 0.006534204352647066, + 0.00395334605127573, + 0.006333870347589254, + 0.010340734384953976, + 0.002445328515022993, + 0.003027749713510275, + 0.009625500999391079, + 0.0021611631382256746, + 0.0030544279143214226, + 0.0022166557610034943, + 0.0030175955034792423, + 0.00990329124033451, + 0.038905415683984756, + 0.008268495090305805, + 0.016200393438339233, + 0.005581784062087536, + 0.0070911371149122715, + 0.011325313709676266, + 0.018969140946865082, + 0.07415854930877686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 1 + ], + "selected_reward": -0.003192449454218149 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static overhead close-up focusing on the Monkey King?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of golden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing an intricately carved phoenix-wing purple-gold crown with two slender pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King clad in shimmering golden chainmail armor with a crimson cape rustling behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the heavy Ruyi Jingu Bang slung over the character's right shoulder with golden hoops gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001659274101257 + ], + "opposite_logprob_tokens": [ + -2.695016622543335 + ], + "logprob_reward": -0.07001659274101257, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001659274101257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16615961492061615, + 0.09720531851053238, + 0.04242360219359398, + 0.12133084237575531, + 0.0985669493675232, + 0.04318762570619583, + 0.019411107525229454, + 0.06525042653083801, + 0.031437087804079056, + 0.31502750515937805 + ], + "credit_map": [ + 0.020413950085639954, + 0.020616352558135986, + 0.019858064129948616, + 0.023947227746248245, + 0.020904524251818657, + 0.016106978058815002, + 0.022213879972696304, + 0.022098615765571594, + 0.0054609449580311775, + 0.006109524052590132, + 0.008306130766868591, + 0.022838182747364044, + 0.007791995536535978, + 0.0071669225580990314, + 0.005824246443808079, + 0.033707354217767715, + 0.003484641434624791, + 0.0031263770069926977, + 0.007630292791873217, + 0.006089480128139257, + 0.004765686113387346, + 0.00467170774936676, + 0.005449801683425903, + 0.007205612026154995, + 0.012473098933696747, + 0.014161641709506512, + 0.01613970287144184, + 0.014971013180911541, + 0.016212651506066322, + 0.01653205417096615, + 0.011741835623979568, + 0.019098831340670586, + 0.010876921005547047, + 0.008446726948022842, + 0.02847147546708584, + 0.008051875047385693, + 0.010373697616159916, + 0.01038255263119936, + 0.0098368925973773, + 0.012126795947551727, + 0.004182836972177029, + 0.005292160902172327, + 0.01060954574495554, + 0.004663331899791956, + 0.005092464853078127, + 0.0035440896172076464, + 0.005894474219530821, + 0.003908717073500156, + 0.0015727992868050933, + 0.0029642311856150627, + 0.0032971547916531563, + 0.0015574608696624637, + 0.0034699002280831337, + 0.0019592344760894775, + 0.002074947115033865, + 0.002515376778319478, + 0.00475316634401679, + 0.006805251352488995, + 0.013685638085007668, + 0.005432254169136286, + 0.007134255487471819, + 0.004746390972286463, + 0.009998917579650879, + 0.012694540433585644, + 0.003466256195679307, + 0.002974965376779437, + 0.006601319182664156, + 0.002240232890471816, + 0.0031119929626584053, + 0.004553241655230522, + 0.002860499080270529, + 0.005628576502203941, + 0.053823038935661316, + 0.011342339217662811, + 0.03348936140537262, + 0.008450144901871681, + 0.009355524554848671, + 0.016040438786149025, + 0.040152452886104584, + 0.1423741728067398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.07001659274101257 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_7/index_tensor([8117, 8117])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_7/index_tensor([8117, 8117])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cb12e4bcf71b8ae8bd8b2368a6509408fbc572e4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_7/index_tensor([8117, 8117])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from a level angle, the camera moving incredibly slowly forward, focuses on a young hero seated cross-legged at the center of a massive lotus throne. His long, lustrous black hair is bound into two delicate buns on either side of his head by red ribbons. His face is as handsome as jade, his expression as serene as a sleeping sculpture. The lotus throne beneath him is carved from flawless white jade, its layers of petals emanating a soft, holy light, enveloping him in a hazy, ethereal aura. Suddenly, his long, thick eyelashes begin to tremble gently and rhythmically, the first sign of awakening after millennia of silence. Then, his eyelids slowly lift upward, the movement steady and deliberate, as if emerging from a deep, timeless dream. The moment his eyes fully opened, a brilliant golden light suddenly shone from the once-empty, deep pupils. This radiance was sharp and powerful, as if imbued with the wisdom and majesty of the dawning universe, heralding the true return of a legendary hero. The entire scene, under a soft and holy light, with a base of pure white and ethereal blue, was illuminated by the bright red of the youth's robe and the golden light in his eyes, creating a solemn, sacred, and powerful atmosphere of rebirth.", + "seed": 12404, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_7/index_8117_seed_12404_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.09131453931331635, + "vlm_logprob": -0.09131453931331635, + "vlm_yes_no_margin": 2.5, + "vlm_token_credit": -0.09131453931331635, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a level angle with the camera moving slowly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young hero seated cross-legged at the center of a massive lotus throne?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero have long black hair bound into two buns by red ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hero's eyelids slowly lift to reveal eyes emitting a brilliant golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lotus throne carved from white jade and emitting a soft holy light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06420256942510605 + ], + "opposite_logprob_tokens": [ + -3.3142025470733643 + ], + "logprob_reward": -0.06420256942510605, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.06420256942510605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21372485160827637, + 0.0748802199959755, + 0.04616736248135567, + 0.03722744062542915, + 0.07949462532997131, + 0.05065961554646492, + 0.1235160306096077, + 0.07934712618589401, + 0.07166916131973267, + 0.22331355512142181 + ], + "credit_map": [ + 0.043264057487249374, + 0.02706207148730755, + 0.04052705690264702, + 0.018806440755724907, + 0.01497857365757227, + 0.02689647488296032, + 0.02000625617802143, + 0.02218392863869667, + 0.012448050081729889, + 0.006803776603192091, + 0.012799030169844627, + 0.0058173201978206635, + 0.0076548936776816845, + 0.008804494515061378, + 0.00861055962741375, + 0.011942108161747456, + 0.007201990112662315, + 0.004710646346211433, + 0.008839367888867855, + 0.003610751125961542, + 0.005509176757186651, + 0.005366264842450619, + 0.005003107711672783, + 0.005926062352955341, + 0.00619804160669446, + 0.003408366348594427, + 0.004925219342112541, + 0.002882411004975438, + 0.005547155160456896, + 0.004774671513587236, + 0.004953613970428705, + 0.004537961911410093, + 0.012296011671423912, + 0.007109904661774635, + 0.015241694636642933, + 0.004702927079051733, + 0.008154802024364471, + 0.008596143685281277, + 0.009070794098079205, + 0.014322355389595032, + 0.006226872093975544, + 0.0060910554602742195, + 0.007053818553686142, + 0.004380485042929649, + 0.007974496111273766, + 0.006040630862116814, + 0.005903946235775948, + 0.006988315377384424, + 0.022473441436886787, + 0.009036271832883358, + 0.02522537112236023, + 0.00755838630720973, + 0.01235742587596178, + 0.019901171326637268, + 0.011521757580339909, + 0.01544220931828022, + 0.01118095126003027, + 0.0060081384144723415, + 0.013874681666493416, + 0.005796306300908327, + 0.01183401420712471, + 0.011868707835674286, + 0.010082642547786236, + 0.0087016886100173, + 0.010508586652576923, + 0.006790708750486374, + 0.008092760108411312, + 0.004184255376458168, + 0.009390369057655334, + 0.010469830594956875, + 0.008226249366998672, + 0.014006408862769604, + 0.03296418488025665, + 0.01787073165178299, + 0.04220566898584366, + 0.013371013104915619, + 0.03272072598338127, + 0.03260834515094757, + 0.02677040733397007, + 0.024802470579743385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.06420256942510605 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a level angle with the camera moving slowly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young hero seated cross-legged at the center of a massive lotus throne?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero have long black hair bound into two buns by red ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hero's eyelids slowly lift to reveal eyes emitting a brilliant golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lotus throne carved from white jade and emitting a soft holy light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07906996458768845 + ], + "opposite_logprob_tokens": [ + -2.5790698528289795 + ], + "logprob_reward": -0.07906996458768845, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07906996458768845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2592568099498749, + 0.12959934771060944, + 0.05827109515666962, + 0.032317571341991425, + 0.056108035147190094, + 0.0844951868057251, + 0.04414815455675125, + 0.07504668086767197, + 0.0682043507695198, + 0.1925528198480606 + ], + "credit_map": [ + 0.04821870103478432, + 0.03262410685420036, + 0.049574144184589386, + 0.01837281510233879, + 0.031122634187340736, + 0.031281087547540665, + 0.023316312581300735, + 0.02474701590836048, + 0.0265139602124691, + 0.013514283113181591, + 0.020948132500052452, + 0.00841127336025238, + 0.013992397114634514, + 0.016369659453630447, + 0.012608825229108334, + 0.01724081113934517, + 0.011774624697864056, + 0.0043198601342737675, + 0.009277894161641598, + 0.004017214756458998, + 0.006161631550639868, + 0.007358943577855825, + 0.005213776603341103, + 0.010147152468562126, + 0.006928144488483667, + 0.0031602107919752598, + 0.006365609355270863, + 0.0019758958369493484, + 0.0027825275901705027, + 0.00363632058724761, + 0.0029608532786369324, + 0.0045080105774104595, + 0.007910572923719883, + 0.0056401449255645275, + 0.007842736318707466, + 0.0044458708725869656, + 0.007724699564278126, + 0.00560256140306592, + 0.007553600240498781, + 0.009387853555381298, + 0.0125595573335886, + 0.004367248620837927, + 0.013784284703433514, + 0.005628024693578482, + 0.00913394708186388, + 0.010298230685293674, + 0.008434662595391273, + 0.020289231091737747, + 0.006674807518720627, + 0.0037579804193228483, + 0.006859662476927042, + 0.004381626844406128, + 0.007008843589574099, + 0.0052605317905545235, + 0.005025419406592846, + 0.005179284140467644, + 0.01475563831627369, + 0.005887344945222139, + 0.015338781289756298, + 0.006888191681355238, + 0.009363530203700066, + 0.008212560787796974, + 0.006743572652339935, + 0.007857056334614754, + 0.008712503127753735, + 0.005602741148322821, + 0.006599084939807653, + 0.005890014581382275, + 0.013382373377680779, + 0.00765953166410327, + 0.00991386640816927, + 0.010444235056638718, + 0.038278549909591675, + 0.011110021732747555, + 0.049785319715738297, + 0.011107751168310642, + 0.011294867843389511, + 0.03631265461444855, + 0.01389631349593401, + 0.020767346024513245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.07906996458768845 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a level angle with the camera moving slowly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young hero seated cross-legged at the center of a massive lotus throne?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero have long black hair bound into two buns by red ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hero's eyelids slowly lift to reveal eyes emitting a brilliant golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lotus throne carved from white jade and emitting a soft holy light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14284437894821167 + ], + "opposite_logprob_tokens": [ + -2.0178444385528564 + ], + "logprob_reward": -0.14284437894821167, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14284437894821167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22090986371040344, + 0.06947899609804153, + 0.08124690502882004, + 0.030509384348988533, + 0.0605640672147274, + 0.22321370244026184, + 0.03934295102953911, + 0.08167856931686401, + 0.0800539180636406, + 0.11300164461135864 + ], + "credit_map": [ + 0.03270501643419266, + 0.027908898890018463, + 0.03323311358690262, + 0.022325312718749046, + 0.03153559938073158, + 0.022316405549645424, + 0.032759446650743484, + 0.018126077950000763, + 0.007646827958524227, + 0.004899461288005114, + 0.008965061977505684, + 0.008381323888897896, + 0.010799609124660492, + 0.007985549978911877, + 0.010390368290245533, + 0.010410789400339127, + 0.015176781453192234, + 0.005986799485981464, + 0.007800985127687454, + 0.005723379552364349, + 0.009508739225566387, + 0.010177732445299625, + 0.00848692748695612, + 0.01838556118309498, + 0.006429647095501423, + 0.0030126748606562614, + 0.005525033455342054, + 0.0017912747571244836, + 0.003465330693870783, + 0.003596286289393902, + 0.0034706627484411, + 0.003218473168089986, + 0.008373924531042576, + 0.005186993163079023, + 0.008618623949587345, + 0.005031075328588486, + 0.007578413467854261, + 0.006290734745562077, + 0.007641758304089308, + 0.011842545121908188, + 0.026417888700962067, + 0.028181785717606544, + 0.02277507446706295, + 0.01918342337012291, + 0.04032883793115616, + 0.031632620841264725, + 0.03847464546561241, + 0.01621943525969982, + 0.00613190233707428, + 0.0039734188467264175, + 0.0076661608181893826, + 0.0027824679855257273, + 0.005850676912814379, + 0.00406570453196764, + 0.005217121914029121, + 0.003655496519058943, + 0.014061074703931808, + 0.009916246868669987, + 0.012075994163751602, + 0.005864083766937256, + 0.01187824085354805, + 0.0106446398422122, + 0.009152763523161411, + 0.008085524663329124, + 0.011378521099686623, + 0.006761525757610798, + 0.008336635306477547, + 0.006008530035614967, + 0.008940807543694973, + 0.00965025369077921, + 0.009165510535240173, + 0.01981213130056858, + 0.01707339473068714, + 0.008423089981079102, + 0.018272072076797485, + 0.009222500957548618, + 0.01794653758406639, + 0.012166379019618034, + 0.014246692880988121, + 0.015650974586606026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.14284437894821167 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a level angle with the camera moving slowly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young hero seated cross-legged at the center of a massive lotus throne?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero have long black hair bound into two buns by red ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hero's eyelids slowly lift to reveal eyes emitting a brilliant golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lotus throne carved from white jade and emitting a soft holy light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10036305338144302 + ], + "opposite_logprob_tokens": [ + -2.35036301612854 + ], + "logprob_reward": -0.10036305338144302, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10036305338144302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11802715808153152, + 0.05910104140639305, + 0.14869989454746246, + 0.07979719340801239, + 0.11557193100452423, + 0.12710252404212952, + 0.0358508937060833, + 0.03948502987623215, + 0.0711536705493927, + 0.2052106261253357 + ], + "credit_map": [ + 0.015434229746460915, + 0.01789667271077633, + 0.013627531006932259, + 0.011867105029523373, + 0.014575198292732239, + 0.011468138545751572, + 0.014722554944455624, + 0.018435735255479813, + 0.0073868101462721825, + 0.0051287016831338406, + 0.005963634699583054, + 0.005080133210867643, + 0.009217397309839725, + 0.006232913583517075, + 0.01002398319542408, + 0.010067470371723175, + 0.025764379650354385, + 0.011378665454685688, + 0.029180828481912613, + 0.009731266647577286, + 0.011294541880488396, + 0.019417479634284973, + 0.017313562333583832, + 0.024619165807962418, + 0.014851086772978306, + 0.006486767437309027, + 0.01623832993209362, + 0.0052198562771081924, + 0.007193679455667734, + 0.009739578701555729, + 0.010784141719341278, + 0.009283752180635929, + 0.018613547086715698, + 0.00858976785093546, + 0.027813805267214775, + 0.005827966146171093, + 0.012684859335422516, + 0.012190782465040684, + 0.01113242655992508, + 0.018718775361776352, + 0.02156151458621025, + 0.010226204991340637, + 0.026387156918644905, + 0.009096016176044941, + 0.012806430459022522, + 0.014526116661727428, + 0.013805573806166649, + 0.018693508580327034, + 0.006203970871865749, + 0.0028073247522115707, + 0.008982235565781593, + 0.002577116945758462, + 0.0031793485395610332, + 0.004896273370832205, + 0.004155050031840801, + 0.0030495752580463886, + 0.004878016654402018, + 0.0033505549654364586, + 0.0057119345292449, + 0.002875095931813121, + 0.007539904676377773, + 0.00604760879650712, + 0.005324991419911385, + 0.0037569256965070963, + 0.01055421307682991, + 0.0060281227342784405, + 0.007804322987794876, + 0.007311034481972456, + 0.008984845131635666, + 0.007767374627292156, + 0.009789631702005863, + 0.01291411928832531, + 0.0419502817094326, + 0.013440818525850773, + 0.03750753030180931, + 0.015106944367289543, + 0.021645259112119675, + 0.020415497943758965, + 0.02190067619085312, + 0.03324362263083458 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 5 + ], + "selected_reward": -0.10036305338144302 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a level angle with the camera moving slowly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young hero seated cross-legged at the center of a massive lotus throne?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero have long black hair bound into two buns by red ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hero's eyelids slowly lift to reveal eyes emitting a brilliant golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lotus throne carved from white jade and emitting a soft holy light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07009273022413254 + ], + "opposite_logprob_tokens": [ + -2.6950926780700684 + ], + "logprob_reward": -0.07009273022413254, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07009273022413254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3516053557395935, + 0.12059354037046432, + 0.053309377282857895, + 0.022764327004551888, + 0.09204862266778946, + 0.030082672834396362, + 0.03281105309724808, + 0.05264539271593094, + 0.06176060810685158, + 0.1823790967464447 + ], + "credit_map": [ + 0.06925931572914124, + 0.04524454474449158, + 0.06738536804914474, + 0.025262907147407532, + 0.03941626474261284, + 0.043262239545583725, + 0.02711011841893196, + 0.034664593636989594, + 0.02509419620037079, + 0.011925920844078064, + 0.021096041426062584, + 0.00639576418325305, + 0.012486420571804047, + 0.01581481657922268, + 0.010020097717642784, + 0.01776028797030449, + 0.010927745141088963, + 0.003928920719772577, + 0.009117214009165764, + 0.0033736468758434057, + 0.006826904136687517, + 0.006626575719565153, + 0.0050704400055110455, + 0.007437930442392826, + 0.0036676651798188686, + 0.0019571317825466394, + 0.003998097497969866, + 0.0016547930426895618, + 0.0020639433059841394, + 0.0023936289362609386, + 0.0026824953965842724, + 0.004346572328358889, + 0.012407987378537655, + 0.005225697997957468, + 0.018260149285197258, + 0.007329954765737057, + 0.012972453609108925, + 0.009019956924021244, + 0.013312176801264286, + 0.01352024544030428, + 0.004379360470920801, + 0.002152298577129841, + 0.004167984239757061, + 0.00216102646663785, + 0.003361649112775922, + 0.0036925640888512135, + 0.003352053463459015, + 0.006815734785050154, + 0.006467344705015421, + 0.002700638957321644, + 0.005611177533864975, + 0.0025918048340827227, + 0.003770992159843445, + 0.005298013333231211, + 0.0025867929216474295, + 0.003784287953749299, + 0.011449912562966347, + 0.0049498118460178375, + 0.009173616766929626, + 0.004299762658774853, + 0.005835164804011583, + 0.007942866533994675, + 0.00485215475782752, + 0.004142101388424635, + 0.008202850818634033, + 0.00480116019025445, + 0.007772891782224178, + 0.0043998840264976025, + 0.008933165110647678, + 0.006520974449813366, + 0.008384701795876026, + 0.01274497713893652, + 0.03730965405702591, + 0.010797165334224701, + 0.04275600612163544, + 0.011276531964540482, + 0.010828554630279541, + 0.036132000386714935, + 0.012724000960588455, + 0.02055518329143524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.07009273022413254 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a level angle, the camera moving incredibly slowly forward, focuses on a young hero seated cross-legged at the center of a massive lotus throne. His long, lustrous black hair is bound into two delicate buns on either side of his head by red ribbons. His face is as handsome as jade, his expression as serene as a sleeping sculpture. The lotus throne beneath him is carved from flawless white jade, its layers of petals emanating a soft, holy light, enveloping him in a hazy, ethereal aura. Suddenly, his long, thick eyelashes begin to tremble gently and rhythmically, the first sign of awakening after millennia of silence. Then, his eyelids slowly lift upward, the movement steady and deliberate, as if emerging from a deep, timeless dream. The moment his eyes fully opened, a brilliant golden light suddenly shone from the once-empty, deep pupils. This radiance was sharp and powerful, as if imbued with the wisdom and majesty of the dawning universe, heralding the true return of a legendary hero. The entire scene, under a soft and holy light, with a base of pure white and ethereal blue, was illuminated by the bright red of the youth's robe and the golden light in his eyes, creating a solemn, sacred, and powerful atmosphere of rebirth.", + "seed": 12405, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_7/index_8117_seed_12405_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09723029583692551, + "vlm_logprob": -0.09723029583692551, + "vlm_yes_no_margin": 2.4, + "vlm_token_credit": -0.09723029583692551, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a level angle with the camera moving slowly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young hero seated cross-legged at the center of a massive lotus throne?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero have long black hair bound into two buns by red ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hero's eyelids slowly lift to reveal eyes emitting a brilliant golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lotus throne carved from white jade and emitting a soft holy light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.045801907777786255 + ], + "opposite_logprob_tokens": [ + -3.295801877975464 + ], + "logprob_reward": -0.045801907777786255, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.045801907777786255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27150219678878784, + 0.1101895272731781, + 0.1065078005194664, + 0.0666598454117775, + 0.04627135023474693, + 0.07591301202774048, + 0.057436782866716385, + 0.12078852951526642, + 0.06403003633022308, + 0.08070090413093567 + ], + "credit_map": [ + 0.04686889797449112, + 0.02386895753443241, + 0.024133605882525444, + 0.02047012746334076, + 0.013510785065591335, + 0.0706048235297203, + 0.015782572329044342, + 0.05626242980360985, + 0.016905970871448517, + 0.017268341034650803, + 0.018990011885762215, + 0.011363793164491653, + 0.008814390748739243, + 0.014379763044416904, + 0.00710076279938221, + 0.015366503037512302, + 0.013971005566418171, + 0.012268096208572388, + 0.01868393085896969, + 0.011135932058095932, + 0.008565569296479225, + 0.01570947840809822, + 0.01033098716288805, + 0.015842806547880173, + 0.008418234996497631, + 0.010936581529676914, + 0.0073745399713516235, + 0.009652802720665932, + 0.008534453809261322, + 0.00778538640588522, + 0.006265380885452032, + 0.007692471146583557, + 0.005968505050987005, + 0.010122464969754219, + 0.00759601965546608, + 0.005209048744291067, + 0.0035364734940230846, + 0.006302915047854185, + 0.002350177615880966, + 0.005185750313103199, + 0.018835438415408134, + 0.0092457951977849, + 0.005627056583762169, + 0.00681519927456975, + 0.005371944047510624, + 0.01919746771454811, + 0.0036794396582990885, + 0.007140681613236666, + 0.0126803545281291, + 0.00641564279794693, + 0.004517990164458752, + 0.0066146282479166985, + 0.004173353314399719, + 0.013575594872236252, + 0.004338822327554226, + 0.005120403133332729, + 0.01839093491435051, + 0.023674417287111282, + 0.015089609660208225, + 0.012915857136249542, + 0.014788719825446606, + 0.011904856190085411, + 0.010309492237865925, + 0.013714650645852089, + 0.007812992669641972, + 0.015438306145370007, + 0.005944341421127319, + 0.0072648474015295506, + 0.00856083631515503, + 0.006803380325436592, + 0.005859644617885351, + 0.006345687434077263, + 0.01945396699011326, + 0.008947718888521194, + 0.007275854237377644, + 0.0092984763905406, + 0.008405646309256554, + 0.012150020338594913, + 0.005582521203905344, + 0.009586702100932598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.045801907777786255 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a level angle with the camera moving slowly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young hero seated cross-legged at the center of a massive lotus throne?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero have long black hair bound into two buns by red ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hero's eyelids slowly lift to reveal eyes emitting a brilliant golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lotus throne carved from white jade and emitting a soft holy light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07914254814386368 + ], + "opposite_logprob_tokens": [ + -2.5791425704956055 + ], + "logprob_reward": -0.07914254814386368, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07914254814386368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2732110619544983, + 0.0830993726849556, + 0.16497765481472015, + 0.19580283761024475, + 0.04360691457986832, + 0.12694697082042694, + 0.04035204276442528, + 0.02396499365568161, + 0.01654648594558239, + 0.031491734087467194 + ], + "credit_map": [ + 0.0492192842066288, + 0.01854645274579525, + 0.027569852769374847, + 0.012130804359912872, + 0.013803641311824322, + 0.07823774963617325, + 0.01125505194067955, + 0.0624481737613678, + 0.011767457239329815, + 0.01485845260322094, + 0.012389693409204483, + 0.009163438342511654, + 0.006923801265656948, + 0.011660076677799225, + 0.0067646256648004055, + 0.009571816772222519, + 0.022074254229664803, + 0.015626104548573494, + 0.029632894322276115, + 0.019643183797597885, + 0.010478939861059189, + 0.023084409534931183, + 0.018169419839978218, + 0.026268426328897476, + 0.024780740961432457, + 0.01752125658094883, + 0.03490457311272621, + 0.021493345499038696, + 0.019124800339341164, + 0.02628333307802677, + 0.01601923629641533, + 0.035675548017024994, + 0.0055978368036448956, + 0.0041041686199605465, + 0.0068145557306706905, + 0.005715812556445599, + 0.0027531932573765516, + 0.0065324436873197556, + 0.0036475539673119783, + 0.008441345766186714, + 0.05058140680193901, + 0.004270687233656645, + 0.003669528290629387, + 0.0026123549323529005, + 0.003937050700187683, + 0.05517913028597832, + 0.0024177508894354105, + 0.00427905423566699, + 0.006291770841926336, + 0.006624241825193167, + 0.0036128670908510685, + 0.005527776665985584, + 0.004161265213042498, + 0.005459723528474569, + 0.0039685191586613655, + 0.004705874715000391, + 0.003744169371202588, + 0.003789498470723629, + 0.003159925574436784, + 0.0027701871003955603, + 0.002799210138618946, + 0.0024807967711240053, + 0.0017154124798253179, + 0.0035057906061410904, + 0.0019939737394452095, + 0.0026626877952367067, + 0.002195435343310237, + 0.0017647786298766732, + 0.0022725157905369997, + 0.0015998154412955046, + 0.0014439502265304327, + 0.0026133281644433737, + 0.004286088049411774, + 0.0039050967898219824, + 0.004609737079590559, + 0.0041196877136826515, + 0.003983761183917522, + 0.0035274443216621876, + 0.0032883293461054564, + 0.003771586809307337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.07914254814386368 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a level angle with the camera moving slowly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young hero seated cross-legged at the center of a massive lotus throne?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero have long black hair bound into two buns by red ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hero's eyelids slowly lift to reveal eyes emitting a brilliant golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lotus throne carved from white jade and emitting a soft holy light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10034126788377762 + ], + "opposite_logprob_tokens": [ + -2.350341320037842 + ], + "logprob_reward": -0.10034126788377762, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10034126788377762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37348100543022156, + 0.03345510736107826, + 0.08924977481365204, + 0.06857043504714966, + 0.08723465353250504, + 0.16732168197631836, + 0.03028588928282261, + 0.059614282101392746, + 0.05178492143750191, + 0.03900224342942238 + ], + "credit_map": [ + 0.04525464028120041, + 0.0500628687441349, + 0.034867510199546814, + 0.0343901552259922, + 0.025661105290055275, + 0.07472435384988785, + 0.03052304871380329, + 0.07799732685089111, + 0.004458404146134853, + 0.0075749619863927364, + 0.003452987875789404, + 0.004228486213833094, + 0.0036322269588708878, + 0.003353914711624384, + 0.0037764739245176315, + 0.0029776531737297773, + 0.009120117872953415, + 0.021866362541913986, + 0.012121841311454773, + 0.009002363309264183, + 0.00895850732922554, + 0.009446771815419197, + 0.00865289755165577, + 0.0100809121504426, + 0.00796128436923027, + 0.010564214549958706, + 0.009783172979950905, + 0.008603770285844803, + 0.0065740663558244705, + 0.006954859476536512, + 0.007681981194764376, + 0.010447083972394466, + 0.011052029207348824, + 0.017546214163303375, + 0.007788175251334906, + 0.012118696235120296, + 0.010296031832695007, + 0.011179468594491482, + 0.008942665532231331, + 0.008311375044286251, + 0.01994253695011139, + 0.024117879569530487, + 0.019626248627901077, + 0.023712096735835075, + 0.016482500359416008, + 0.02070949599146843, + 0.023855246603488922, + 0.01887567900121212, + 0.004421311430633068, + 0.006024346686899662, + 0.0035350441467016935, + 0.002706501167267561, + 0.002667222637683153, + 0.00481451628729701, + 0.00248122401535511, + 0.0036357229109853506, + 0.008891289122402668, + 0.010224915109574795, + 0.007744618225842714, + 0.00658246548846364, + 0.006567710544914007, + 0.006449729669839144, + 0.0052034310065209866, + 0.007950123399496078, + 0.007081529125571251, + 0.007718642707914114, + 0.006729481741786003, + 0.007145310752093792, + 0.00673076743260026, + 0.004581141751259565, + 0.00544627895578742, + 0.006351771764457226, + 0.005359210539609194, + 0.006947372108697891, + 0.004820706322789192, + 0.004532619845122099, + 0.004372462630271912, + 0.00398947112262249, + 0.004044105298817158, + 0.004936294164508581 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.10034126788377762 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a level angle with the camera moving slowly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young hero seated cross-legged at the center of a massive lotus throne?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero have long black hair bound into two buns by red ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hero's eyelids slowly lift to reveal eyes emitting a brilliant golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lotus throne carved from white jade and emitting a soft holy light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10036585479974747 + ], + "opposite_logprob_tokens": [ + -2.3503658771514893 + ], + "logprob_reward": -0.10036585479974747, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10036585479974747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15367482602596283, + 0.0844540074467659, + 0.06271401792764664, + 0.06285009533166885, + 0.039180342108011246, + 0.1223548948764801, + 0.13056036829948425, + 0.1265806406736374, + 0.16156712174415588, + 0.056063685566186905 + ], + "credit_map": [ + 0.030358867719769478, + 0.02018682286143303, + 0.013685768470168114, + 0.014264761470258236, + 0.011090240441262722, + 0.028045905753970146, + 0.009693716652691364, + 0.026348737999796867, + 0.014081720262765884, + 0.01592007838189602, + 0.012384416535496712, + 0.008170144632458687, + 0.0071369316428899765, + 0.00992273073643446, + 0.006332067307084799, + 0.010505921207368374, + 0.0056237345561385155, + 0.011491462588310242, + 0.009118281304836273, + 0.00916345790028572, + 0.006213274784386158, + 0.006414972711354494, + 0.00690986355766654, + 0.007778972387313843, + 0.007355456706136465, + 0.009698422625660896, + 0.007914725691080093, + 0.010230222716927528, + 0.008138857781887054, + 0.0067369588650763035, + 0.005761015694588423, + 0.0070144375786185265, + 0.008010211400687695, + 0.004894899670034647, + 0.0032907130662351847, + 0.003391076112166047, + 0.003320361254736781, + 0.009295391850173473, + 0.0030958845745772123, + 0.003881806507706642, + 0.03869492560625076, + 0.011538728140294552, + 0.005583113990724087, + 0.006526226177811623, + 0.007014057599008083, + 0.04111310467123985, + 0.0051832059398293495, + 0.006701522506773472, + 0.021646598353981972, + 0.021172143518924713, + 0.01374170370399952, + 0.017392583191394806, + 0.011112901382148266, + 0.02408083714544773, + 0.010334957391023636, + 0.011078646406531334, + 0.0189824178814888, + 0.02861098386347294, + 0.012616069987416267, + 0.013488340191543102, + 0.019155224785208702, + 0.012926295399665833, + 0.009410222992300987, + 0.011391093023121357, + 0.018258124589920044, + 0.04128265380859375, + 0.01743592508137226, + 0.016529161483049393, + 0.023892439901828766, + 0.016389962285757065, + 0.012788836844265461, + 0.014990029856562614, + 0.009116214700043201, + 0.01023623626679182, + 0.0058593954890966415, + 0.006107129622250795, + 0.006715669296681881, + 0.006124392617493868, + 0.004518883768469095, + 0.0073857661336660385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.10036585479974747 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a level angle with the camera moving slowly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young hero seated cross-legged at the center of a massive lotus throne?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero have long black hair bound into two buns by red ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hero's eyelids slowly lift to reveal eyes emitting a brilliant golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lotus throne carved from white jade and emitting a soft holy light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16049990057945251 + ], + "opposite_logprob_tokens": [ + -1.910499930381775 + ], + "logprob_reward": -0.16049990057945251, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16049990057945251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22518467903137207, + 0.06658005714416504, + 0.1551280915737152, + 0.12788379192352295, + 0.04536733031272888, + 0.1619480550289154, + 0.046567514538764954, + 0.04258080571889877, + 0.07122429460287094, + 0.05753539875149727 + ], + "credit_map": [ + 0.0495896078646183, + 0.014165296219289303, + 0.023421023041009903, + 0.01141746249049902, + 0.010371999815106392, + 0.06810031831264496, + 0.009113996289670467, + 0.03900497034192085, + 0.01310209184885025, + 0.00906084943562746, + 0.009452753700315952, + 0.008444828912615776, + 0.0057200780138373375, + 0.009355800226330757, + 0.004375874530524015, + 0.007067783735692501, + 0.019206590950489044, + 0.01476428471505642, + 0.0301708672195673, + 0.016767051070928574, + 0.01037830300629139, + 0.02123754657804966, + 0.01628146320581436, + 0.02632199227809906, + 0.015224157832562923, + 0.011538918130099773, + 0.023431049659848213, + 0.015306376852095127, + 0.01233027782291174, + 0.01563194766640663, + 0.010624325834214687, + 0.023796748369932175, + 0.0068173836916685104, + 0.005813521333038807, + 0.004615702200680971, + 0.005046546924859285, + 0.0034015963319689035, + 0.007695212960243225, + 0.005250351969152689, + 0.006727016530930996, + 0.05132550373673439, + 0.010814465582370758, + 0.009732085280120373, + 0.008893287740647793, + 0.007297866512089968, + 0.05767616629600525, + 0.007186227478086948, + 0.009022463113069534, + 0.006170750129967928, + 0.008301311172544956, + 0.005265567917376757, + 0.005861968733370304, + 0.004261606372892857, + 0.006818708963692188, + 0.004722516983747482, + 0.005165083799511194, + 0.0062620858661830425, + 0.010432777926325798, + 0.00365630560554564, + 0.004723197314888239, + 0.0054919240064918995, + 0.004302571061998606, + 0.002982641337439418, + 0.004729302600026131, + 0.008845883421599865, + 0.013877786695957184, + 0.009811562485992908, + 0.00754897203296423, + 0.008270042948424816, + 0.0065038204193115234, + 0.007115081883966923, + 0.009251147508621216, + 0.013353046961128712, + 0.00863537285476923, + 0.004518670029938221, + 0.005582655314356089, + 0.006772964261472225, + 0.009309940971434116, + 0.0036287831608206034, + 0.005733966361731291 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.16049990057945251 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a level angle, the camera moving incredibly slowly forward, focuses on a young hero seated cross-legged at the center of a massive lotus throne. His long, lustrous black hair is bound into two delicate buns on either side of his head by red ribbons. His face is as handsome as jade, his expression as serene as a sleeping sculpture. The lotus throne beneath him is carved from flawless white jade, its layers of petals emanating a soft, holy light, enveloping him in a hazy, ethereal aura. Suddenly, his long, thick eyelashes begin to tremble gently and rhythmically, the first sign of awakening after millennia of silence. Then, his eyelids slowly lift upward, the movement steady and deliberate, as if emerging from a deep, timeless dream. The moment his eyes fully opened, a brilliant golden light suddenly shone from the once-empty, deep pupils. This radiance was sharp and powerful, as if imbued with the wisdom and majesty of the dawning universe, heralding the true return of a legendary hero. The entire scene, under a soft and holy light, with a base of pure white and ethereal blue, was illuminated by the bright red of the youth's robe and the golden light in his eyes, creating a solemn, sacred, and powerful atmosphere of rebirth.", + "seed": 12406, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_7/index_8117_seed_12406_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04874028004705906, + "vlm_logprob": -0.04874028004705906, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.04874028004705906, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a level angle with the camera moving slowly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young hero seated cross-legged at the center of a massive lotus throne?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero have long black hair bound into two buns by red ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hero's eyelids slowly lift to reveal eyes emitting a brilliant golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lotus throne carved from white jade and emitting a soft holy light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027194133028388023 + ], + "opposite_logprob_tokens": [ + -4.527194023132324 + ], + "logprob_reward": -0.027194133028388023, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.027194133028388023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1574818193912506, + 0.010064727626740932, + 0.715057373046875, + 0.011154860258102417, + 0.009443165734410286, + 0.014684714376926422, + 0.01531420461833477, + 0.024916475638747215, + 0.02040315791964531, + 0.021479502320289612 + ], + "credit_map": [ + 0.016814153641462326, + 0.0028017812874168158, + 0.014519044198095798, + 0.0028586399275809526, + 0.005232661031186581, + 0.061809465289115906, + 0.004291598219424486, + 0.049154482781887054, + 0.0016666430747136474, + 0.000758361245971173, + 0.0019614300690591335, + 0.0009466623887419701, + 0.0007726785261183977, + 0.001953771570697427, + 0.000580682186409831, + 0.001424498506821692, + 0.08445283025503159, + 0.010293174535036087, + 0.06369417905807495, + 0.010282615199685097, + 0.015807118266820908, + 0.2678258717060089, + 0.013657480478286743, + 0.24904409050941467, + 0.0017785308882594109, + 0.001034085056744516, + 0.0015453383093699813, + 0.0007075915345922112, + 0.0011310548288747668, + 0.0022077744361013174, + 0.0011305399239063263, + 0.0016199455130845308, + 0.0013460839400067925, + 0.0012012096121907234, + 0.0017802382353693247, + 0.0007737788837403059, + 0.001097384374588728, + 0.001247693900950253, + 0.0010137585923075676, + 0.0009830183116719127, + 0.002864261157810688, + 0.0012378249084576964, + 0.0014539025723934174, + 0.0009381364216096699, + 0.0023387677501887083, + 0.002514590509235859, + 0.0012312744511291385, + 0.0021059575956314802, + 0.0016169368755072355, + 0.0017798561602830887, + 0.001948092714883387, + 0.0018917513079941273, + 0.0023193219676613808, + 0.0018511901143938303, + 0.0018361323745921254, + 0.0020709228701889515, + 0.0033442957792431116, + 0.004026380833238363, + 0.0022509689442813396, + 0.003257060656324029, + 0.003617699723690748, + 0.002810136415064335, + 0.0024816440418362617, + 0.003128289943560958, + 0.002197618130594492, + 0.003529404057189822, + 0.0016612104373052716, + 0.002563444199040532, + 0.0026541128754615784, + 0.002245255745947361, + 0.002797825261950493, + 0.002754287561401725, + 0.002305469708517194, + 0.0032688057981431484, + 0.001172153395600617, + 0.0025769288185983896, + 0.0041158320382237434, + 0.003244984894990921, + 0.00248980475589633, + 0.002305523958057165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.027194133028388023 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a level angle with the camera moving slowly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young hero seated cross-legged at the center of a massive lotus throne?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero have long black hair bound into two buns by red ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hero's eyelids slowly lift to reveal eyes emitting a brilliant golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lotus throne carved from white jade and emitting a soft holy light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07906743139028549 + ], + "opposite_logprob_tokens": [ + -2.5790674686431885 + ], + "logprob_reward": -0.07906743139028549, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07906743139028549, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13085275888442993, + 0.14928989112377167, + 0.12763693928718567, + 0.12063425034284592, + 0.06523396819829941, + 0.062064796686172485, + 0.0671955794095993, + 0.10297543555498123, + 0.09023091942071915, + 0.08388540148735046 + ], + "credit_map": [ + 0.016084501519799232, + 0.011961699463427067, + 0.014685980044305325, + 0.015678483992815018, + 0.012732469476759434, + 0.028920689597725868, + 0.012196529656648636, + 0.018592409789562225, + 0.033754847943782806, + 0.010726263746619225, + 0.019115490838885307, + 0.011535827070474625, + 0.012842145748436451, + 0.03218312934041023, + 0.008799458853900433, + 0.020332738757133484, + 0.011337446980178356, + 0.003773143980652094, + 0.01540518831461668, + 0.003883661702275276, + 0.004325623158365488, + 0.038429491221904755, + 0.005288063548505306, + 0.04519433155655861, + 0.018359646201133728, + 0.008322780951857567, + 0.01904148980975151, + 0.013864283449947834, + 0.009528378024697304, + 0.024490641430020332, + 0.014876789413392544, + 0.012150254100561142, + 0.007471943739801645, + 0.009902714751660824, + 0.00819824542850256, + 0.007228090427815914, + 0.010070056654512882, + 0.008497428148984909, + 0.0077803065069019794, + 0.006085183005779982, + 0.01059664972126484, + 0.008551094681024551, + 0.00781568605452776, + 0.005450816825032234, + 0.009139270521700382, + 0.00913848727941513, + 0.0056853704154491425, + 0.00568742910400033, + 0.008029884658753872, + 0.00993152242153883, + 0.008518817834556103, + 0.009036636911332607, + 0.01016115304082632, + 0.0073531679809093475, + 0.008263379335403442, + 0.00590102095156908, + 0.012800803408026695, + 0.017665958032011986, + 0.007362435571849346, + 0.01710071973502636, + 0.01988113671541214, + 0.010257994756102562, + 0.010839021764695644, + 0.007067369297146797, + 0.008879289031028748, + 0.017273342236876488, + 0.008477219380438328, + 0.012241304852068424, + 0.015180852264165878, + 0.008783474564552307, + 0.007844231091439724, + 0.011551212519407272, + 0.006878838874399662, + 0.01104819867759943, + 0.005640188232064247, + 0.015318114310503006, + 0.01203871425241232, + 0.0101614473387599, + 0.011808246374130249, + 0.010991650633513927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.07906743139028549 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a level angle with the camera moving slowly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young hero seated cross-legged at the center of a massive lotus throne?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero have long black hair bound into two buns by red ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hero's eyelids slowly lift to reveal eyes emitting a brilliant golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lotus throne carved from white jade and emitting a soft holy light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033724263310432434 + ], + "opposite_logprob_tokens": [ + -3.408724308013916 + ], + "logprob_reward": -0.033724263310432434, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033724263310432434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08081497251987457, + 0.04330788552761078, + 0.6410971879959106, + 0.012647794559597969, + 0.011557166464626789, + 0.023594168946146965, + 0.02186877653002739, + 0.033542510122060776, + 0.05853424593806267, + 0.07303529232740402 + ], + "credit_map": [ + 0.008330589160323143, + 0.0019981602672487497, + 0.007750058081001043, + 0.003018039045855403, + 0.0035240950528532267, + 0.030235929414629936, + 0.0034199764486402273, + 0.022538116201758385, + 0.00943054724484682, + 0.0027753585018217564, + 0.008434290997684002, + 0.0037569941487163305, + 0.002177409129217267, + 0.008580639027059078, + 0.003114261431619525, + 0.005038384813815355, + 0.07625405490398407, + 0.009774887003004551, + 0.05683588236570358, + 0.009608225896954536, + 0.01405077613890171, + 0.24044184386730194, + 0.011890984140336514, + 0.22224053740501404, + 0.0011268483940511942, + 0.0013320565922185779, + 0.0019557708874344826, + 0.0019772520754486322, + 0.001209869864396751, + 0.0016877508023753762, + 0.0020940881222486496, + 0.0012641578214243054, + 0.0012857080437242985, + 0.0013980860821902752, + 0.0015739903319627047, + 0.0010041712084785104, + 0.0015795836225152016, + 0.001803145743906498, + 0.0010426279623061419, + 0.00186985416803509, + 0.002871001372113824, + 0.002619880484417081, + 0.0029371175915002823, + 0.0014736095909029245, + 0.002207490149885416, + 0.004736330825835466, + 0.00209629163146019, + 0.004652447532862425, + 0.0030602223705500364, + 0.003044068580493331, + 0.0025065287481993437, + 0.0028962886426597834, + 0.002401870908215642, + 0.002483245451003313, + 0.0026430319994688034, + 0.0028335200622677803, + 0.004087148234248161, + 0.006065786816179752, + 0.0027362010441720486, + 0.004941587336361408, + 0.006661052815616131, + 0.003457128768786788, + 0.0031046816147863865, + 0.0024889239575713873, + 0.005849106702953577, + 0.009572772309184074, + 0.006160346791148186, + 0.009834875352680683, + 0.008948066271841526, + 0.0057820770889520645, + 0.006906711496412754, + 0.005480289459228516, + 0.006320129614323378, + 0.011619661934673786, + 0.005947576370090246, + 0.01408083364367485, + 0.01050878968089819, + 0.007602243684232235, + 0.010225014761090279, + 0.006731044966727495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.033724263310432434 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a level angle with the camera moving slowly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young hero seated cross-legged at the center of a massive lotus throne?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero have long black hair bound into two buns by red ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hero's eyelids slowly lift to reveal eyes emitting a brilliant golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lotus throne carved from white jade and emitting a soft holy light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05499142035841942 + ], + "opposite_logprob_tokens": [ + -2.9299914836883545 + ], + "logprob_reward": -0.05499142035841942, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05499142035841942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08032440394163132, + 0.016394391655921936, + 0.6558957099914551, + 0.019151583313941956, + 0.010759934782981873, + 0.01430464442819357, + 0.03941134735941887, + 0.026435598731040955, + 0.03573185205459595, + 0.10159049928188324 + ], + "credit_map": [ + 0.008802637457847595, + 0.002088838256895542, + 0.009235176257789135, + 0.0038028056733310223, + 0.0034181701485067606, + 0.030213652178645134, + 0.003150154137983918, + 0.019612964242696762, + 0.0022771870717406273, + 0.0012775870272889733, + 0.0020963309798389673, + 0.0026074242778122425, + 0.0015094515401870012, + 0.002632542047649622, + 0.001926587545312941, + 0.0020672809332609177, + 0.07779497653245926, + 0.009743401780724525, + 0.05759720131754875, + 0.009364744648337364, + 0.014927838928997517, + 0.24654832482337952, + 0.012641648761928082, + 0.22727756202220917, + 0.0018155776197090745, + 0.0016718258848413825, + 0.003757994621992111, + 0.0025365669280290604, + 0.0018017619149759412, + 0.002734518377110362, + 0.0028027098160237074, + 0.0020306273363530636, + 0.001124454545788467, + 0.001073284656740725, + 0.0015800485853105783, + 0.000571868964470923, + 0.0009728078148327768, + 0.0024738688953220844, + 0.0008504789439029992, + 0.002113122260197997, + 0.0017751321429386735, + 0.002775211352854967, + 0.0019644340500235558, + 0.001331051578745246, + 0.00175057805608958, + 0.001856345683336258, + 0.0013001625193282962, + 0.0015517300926148891, + 0.005830144509673119, + 0.005618660245090723, + 0.004936566110700369, + 0.0047412957064807415, + 0.0051645380444824696, + 0.004567625932395458, + 0.0038985766004770994, + 0.004653941374272108, + 0.0033283971715718508, + 0.0038994753267616034, + 0.002332617063075304, + 0.00362632074393332, + 0.004752239212393761, + 0.0028245856519788504, + 0.0028701990377157927, + 0.002801764290779829, + 0.00386143010109663, + 0.007272847928106785, + 0.002294096164405346, + 0.0054961140267550945, + 0.005212741903960705, + 0.0031510714907199144, + 0.004167288541793823, + 0.004276262130588293, + 0.010340721346437931, + 0.009781597182154655, + 0.008485778234899044, + 0.011881662532687187, + 0.012716027908027172, + 0.013773743063211441, + 0.012539582327008247, + 0.022071389481425285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.05499142035841942 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a level angle with the camera moving slowly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young hero seated cross-legged at the center of a massive lotus throne?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero have long black hair bound into two buns by red ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hero's eyelids slowly lift to reveal eyes emitting a brilliant golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lotus throne carved from white jade and emitting a soft holy light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04872415214776993 + ], + "opposite_logprob_tokens": [ + -3.0487241744995117 + ], + "logprob_reward": -0.04872415214776993, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04872415214776993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03231095150113106, + 0.0374743714928627, + 0.6522042155265808, + 0.047938235104084015, + 0.04814697057008743, + 0.02139190398156643, + 0.020002789795398712, + 0.01918404921889305, + 0.021248094737529755, + 0.10009843856096268 + ], + "credit_map": [ + 0.004066083580255508, + 0.002074797870591283, + 0.006599774584174156, + 0.0032325058709830046, + 0.0024942271411418915, + 0.0033546951599419117, + 0.0024243867956101894, + 0.008064483292400837, + 0.007582811173051596, + 0.00288886995986104, + 0.006159541197121143, + 0.0028874787967652082, + 0.003436553990468383, + 0.007298795506358147, + 0.0024602417834103107, + 0.0047600786201655865, + 0.07581396400928497, + 0.009081977419555187, + 0.05958310142159462, + 0.009423605166375637, + 0.014078725129365921, + 0.24207782745361328, + 0.012413861230015755, + 0.22973115742206573, + 0.007526462897658348, + 0.003557470627129078, + 0.007773969322443008, + 0.004147938918322325, + 0.0046189920976758, + 0.009725123643875122, + 0.0054282741621136665, + 0.005160005297511816, + 0.006370830349624157, + 0.006653177551925182, + 0.007343284320086241, + 0.004498243797570467, + 0.00665697455406189, + 0.007302726153284311, + 0.004572692792862654, + 0.004749041981995106, + 0.0029426405671983957, + 0.002038822043687105, + 0.0024904971942305565, + 0.0016632376937195659, + 0.0033858660608530045, + 0.0031386222690343857, + 0.0017027963185682893, + 0.004029421601444483, + 0.002561277011409402, + 0.003104923525825143, + 0.0022894099820405245, + 0.0025789544451981783, + 0.002434130758047104, + 0.0023284342605620623, + 0.0022249084431678057, + 0.0024807520676404238, + 0.003172816475853324, + 0.0027942194137722254, + 0.0013024170184507966, + 0.002763004507869482, + 0.0023583669681102037, + 0.002289076568558812, + 0.0015741443494334817, + 0.0029300053138285875, + 0.002353607676923275, + 0.0029484094120562077, + 0.002025231486186385, + 0.0037553771398961544, + 0.0027688490226864815, + 0.0021521898452192545, + 0.0026493151672184467, + 0.0025951149873435497, + 0.00834246352314949, + 0.01083011832088232, + 0.006127067841589451, + 0.017229851335287094, + 0.012436789460480213, + 0.011632991954684258, + 0.01691652648150921, + 0.016582628712058067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 4 + ], + "selected_reward": -0.04872415214776993 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a level angle, the camera moving incredibly slowly forward, focuses on a young hero seated cross-legged at the center of a massive lotus throne. His long, lustrous black hair is bound into two delicate buns on either side of his head by red ribbons. His face is as handsome as jade, his expression as serene as a sleeping sculpture. The lotus throne beneath him is carved from flawless white jade, its layers of petals emanating a soft, holy light, enveloping him in a hazy, ethereal aura. Suddenly, his long, thick eyelashes begin to tremble gently and rhythmically, the first sign of awakening after millennia of silence. Then, his eyelids slowly lift upward, the movement steady and deliberate, as if emerging from a deep, timeless dream. The moment his eyes fully opened, a brilliant golden light suddenly shone from the once-empty, deep pupils. This radiance was sharp and powerful, as if imbued with the wisdom and majesty of the dawning universe, heralding the true return of a legendary hero. The entire scene, under a soft and holy light, with a base of pure white and ethereal blue, was illuminated by the bright red of the youth's robe and the golden light in his eyes, creating a solemn, sacred, and powerful atmosphere of rebirth.", + "seed": 12407, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_7/index_8117_seed_12407_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09977450370788574, + "vlm_logprob": -0.09977450370788574, + "vlm_yes_no_margin": 2.525, + "vlm_token_credit": -0.09977450370788574, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a level angle with the camera moving slowly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young hero seated cross-legged at the center of a massive lotus throne?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero have long black hair bound into two buns by red ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hero's eyelids slowly lift to reveal eyes emitting a brilliant golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lotus throne carved from white jade and emitting a soft holy light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04389386251568794 + ], + "opposite_logprob_tokens": [ + -3.793893814086914 + ], + "logprob_reward": -0.04389386251568794, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.04389386251568794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22805345058441162, + 0.04801298305392265, + 0.030364494770765305, + 0.04426715523004532, + 0.05003787949681282, + 0.06387989223003387, + 0.10230463743209839, + 0.08003946393728256, + 0.21265099942684174, + 0.14038898050785065 + ], + "credit_map": [ + 0.044394493103027344, + 0.01777925156056881, + 0.030529070645570755, + 0.029165910556912422, + 0.017300603911280632, + 0.03746792674064636, + 0.030633242800831795, + 0.020782964304089546, + 0.007077840622514486, + 0.00462097954005003, + 0.0039002641569823027, + 0.005892485845834017, + 0.00401272950693965, + 0.010486369952559471, + 0.006018104963004589, + 0.0060042100958526134, + 0.0053830379620194435, + 0.003990751691162586, + 0.0028459273744374514, + 0.002836527768522501, + 0.00269104796461761, + 0.006652989890426397, + 0.002726020058616996, + 0.0032381920609623194, + 0.006536278408020735, + 0.005861571524292231, + 0.0044954316690564156, + 0.004037196282297373, + 0.0027654829900711775, + 0.009267248213291168, + 0.0029133842326700687, + 0.008390558883547783, + 0.007314412854611874, + 0.010606320574879646, + 0.004891746677458286, + 0.005717902444303036, + 0.005866431165486574, + 0.005025968421250582, + 0.0049345772713422775, + 0.005680519621819258, + 0.0066149309277534485, + 0.01523012574762106, + 0.007770228199660778, + 0.009465234354138374, + 0.005130154080688953, + 0.007039014715701342, + 0.007291562855243683, + 0.005338642280548811, + 0.010144829750061035, + 0.020445799455046654, + 0.009860253892838955, + 0.014823139645159245, + 0.009947012178599834, + 0.010977661237120628, + 0.009419134818017483, + 0.016686812043190002, + 0.0064932736568152905, + 0.0161312036216259, + 0.005512670613825321, + 0.010684221051633358, + 0.006649937015026808, + 0.011785599403083324, + 0.005083281546831131, + 0.01769927330315113, + 0.023675894364714622, + 0.03170422837138176, + 0.014782258309423923, + 0.042756348848342896, + 0.016370229423046112, + 0.026350868865847588, + 0.017765646800398827, + 0.039245523512363434, + 0.01574430987238884, + 0.04194514453411102, + 0.009944702498614788, + 0.020231783390045166, + 0.011901202611625195, + 0.01280299760401249, + 0.009641299024224281, + 0.018177540972828865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.04389386251568794 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a level angle with the camera moving slowly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young hero seated cross-legged at the center of a massive lotus throne?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero have long black hair bound into two buns by red ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hero's eyelids slowly lift to reveal eyes emitting a brilliant golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lotus throne carved from white jade and emitting a soft holy light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18017053604125977 + ], + "opposite_logprob_tokens": [ + -1.8051705360412598 + ], + "logprob_reward": -0.18017053604125977, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18017053604125977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2804427742958069, + 0.09963060170412064, + 0.06463441997766495, + 0.05997176840901375, + 0.0843016505241394, + 0.0633246973156929, + 0.09383288025856018, + 0.10030370205640793, + 0.06439274549484253, + 0.08916476368904114 + ], + "credit_map": [ + 0.04962960258126259, + 0.024672089144587517, + 0.056205928325653076, + 0.025287894532084465, + 0.022585351020097733, + 0.04421013593673706, + 0.027548033744096756, + 0.030303725972771645, + 0.01655832678079605, + 0.005926282145082951, + 0.008669782429933548, + 0.011439817026257515, + 0.0064765820279717445, + 0.02517654001712799, + 0.011111574247479439, + 0.014271698892116547, + 0.010438577271997929, + 0.008759455755352974, + 0.0048772552981972694, + 0.008555687963962555, + 0.005686866585165262, + 0.014004007913172245, + 0.006873168516904116, + 0.005439401604235172, + 0.010103865526616573, + 0.008352695032954216, + 0.0049290782772004604, + 0.005951220169663429, + 0.003895314410328865, + 0.015163763426244259, + 0.005152686033397913, + 0.006423145066946745, + 0.01078470703214407, + 0.01702306978404522, + 0.01089872233569622, + 0.008070573210716248, + 0.00678869616240263, + 0.013426882214844227, + 0.00890297256410122, + 0.008406023494899273, + 0.008057833649218082, + 0.012447183020412922, + 0.005642309784889221, + 0.0067314752377569675, + 0.006132983602583408, + 0.012052708305418491, + 0.005992926191538572, + 0.006267277989536524, + 0.013034934177994728, + 0.015377392992377281, + 0.008369327522814274, + 0.009411121718585491, + 0.006553910207003355, + 0.01510712318122387, + 0.0075943926349282265, + 0.018384680151939392, + 0.011726685799658298, + 0.013179169967770576, + 0.011088534258306026, + 0.009242620319128036, + 0.006993796210736036, + 0.013818769715726376, + 0.0086815869435668, + 0.025572532787919044, + 0.005190502852201462, + 0.012221427634358406, + 0.004815523512661457, + 0.01492364052683115, + 0.004332595504820347, + 0.006721331272274256, + 0.006314264144748449, + 0.009873460978269577, + 0.014302892610430717, + 0.021213512867689133, + 0.0060328892432153225, + 0.012782266363501549, + 0.008533304557204247, + 0.01145393680781126, + 0.005642413627356291, + 0.00920354574918747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.18017053604125977 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a level angle with the camera moving slowly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young hero seated cross-legged at the center of a massive lotus throne?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero have long black hair bound into two buns by red ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hero's eyelids slowly lift to reveal eyes emitting a brilliant golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lotus throne carved from white jade and emitting a soft holy light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04869576916098595 + ], + "opposite_logprob_tokens": [ + -3.0486958026885986 + ], + "logprob_reward": -0.04869576916098595, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04869576916098595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2104661762714386, + 0.11695799976587296, + 0.22902069985866547, + 0.06957492977380753, + 0.03623216226696968, + 0.028344901278614998, + 0.10158047825098038, + 0.052451714873313904, + 0.05807154253125191, + 0.09729941189289093 + ], + "credit_map": [ + 0.028464658185839653, + 0.02109033800661564, + 0.02760140597820282, + 0.025026781484484673, + 0.018013659864664078, + 0.031112561002373695, + 0.027847658842802048, + 0.0313091017305851, + 0.012725891545414925, + 0.023529808968305588, + 0.009324993938207626, + 0.019817117601633072, + 0.012577512301504612, + 0.010346722789108753, + 0.018464582040905952, + 0.010171366855502129, + 0.032086972147226334, + 0.03254246711730957, + 0.02735903114080429, + 0.024060431867837906, + 0.01722787693142891, + 0.0424356684088707, + 0.02348942682147026, + 0.029818832874298096, + 0.009061610326170921, + 0.012717547826468945, + 0.007265299558639526, + 0.007774995639920235, + 0.006607584189623594, + 0.010226985439658165, + 0.005469930358231068, + 0.01045097317546606, + 0.004876494407653809, + 0.0075899092480540276, + 0.0025884087663143873, + 0.006502396892756224, + 0.0034087130334228277, + 0.005485884845256805, + 0.003238287288695574, + 0.0025420666206628084, + 0.003604233032092452, + 0.005552792921662331, + 0.0024590720422565937, + 0.0035682087764143944, + 0.0026974270585924387, + 0.004170681349933147, + 0.0027562640607357025, + 0.0035362225025892258, + 0.014110657386481762, + 0.014166910201311111, + 0.010676116682589054, + 0.012297404929995537, + 0.011106101796030998, + 0.017176320776343346, + 0.007370582781732082, + 0.014676389284431934, + 0.004903330467641354, + 0.007100225891917944, + 0.00477308314293623, + 0.005669838283210993, + 0.003883222583681345, + 0.007675868459045887, + 0.0045702834613621235, + 0.013875864446163177, + 0.007796406280249357, + 0.007252020761370659, + 0.0039652492851018906, + 0.008198395371437073, + 0.003123840084299445, + 0.00806298479437828, + 0.003736604005098343, + 0.015936043113470078, + 0.015977680683135986, + 0.02066294103860855, + 0.007358535658568144, + 0.01790604554116726, + 0.0066123055294156075, + 0.010472031310200691, + 0.008163824677467346, + 0.010146046057343483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.04869576916098595 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a level angle with the camera moving slowly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young hero seated cross-legged at the center of a massive lotus throne?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero have long black hair bound into two buns by red ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hero's eyelids slowly lift to reveal eyes emitting a brilliant golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lotus throne carved from white jade and emitting a soft holy light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11300762742757797 + ], + "opposite_logprob_tokens": [ + -2.2380075454711914 + ], + "logprob_reward": -0.11300762742757797, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11300762742757797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10915809124708176, + 0.04183577373623848, + 0.048951368778944016, + 0.027409708127379417, + 0.026803577318787575, + 0.03067834861576557, + 0.27504387497901917, + 0.11011579632759094, + 0.08929384499788284, + 0.24070961773395538 + ], + "credit_map": [ + 0.01620347797870636, + 0.008767431601881981, + 0.016725871711969376, + 0.016773296520113945, + 0.008760549128055573, + 0.016127068549394608, + 0.01727639138698578, + 0.008524000644683838, + 0.007433885708451271, + 0.004912474192678928, + 0.003155339742079377, + 0.0035109384916722775, + 0.0033371334429830313, + 0.010226571932435036, + 0.0035805809311568737, + 0.005678848829120398, + 0.0072336052544415, + 0.009700076654553413, + 0.0038142018020153046, + 0.004348888527601957, + 0.005898145958781242, + 0.007056177593767643, + 0.0042764958925545216, + 0.006623777095228434, + 0.004034448880702257, + 0.004802996758371592, + 0.0032311517279595137, + 0.0032211728394031525, + 0.0021446559112519026, + 0.004655419383198023, + 0.002245260402560234, + 0.00307460012845695, + 0.003978738095611334, + 0.00509030744433403, + 0.0021839234977960587, + 0.0029552604537457228, + 0.003268021857365966, + 0.004096355754882097, + 0.002423310885205865, + 0.0028076583985239267, + 0.004264382179826498, + 0.006372231990098953, + 0.0027109866496175528, + 0.004108042921870947, + 0.003973481245338917, + 0.00418449379503727, + 0.002420315984636545, + 0.0026444129180163145, + 0.05642162635922432, + 0.03174804523587227, + 0.021889271214604378, + 0.02823561243712902, + 0.016902649775147438, + 0.05327265337109566, + 0.016660865396261215, + 0.04991312697529793, + 0.009708043187856674, + 0.018296627327799797, + 0.011144584976136684, + 0.011995230801403522, + 0.008273530751466751, + 0.015644224360585213, + 0.009552937000989914, + 0.025500614196062088, + 0.00942323263734579, + 0.013439980335533619, + 0.009600440040230751, + 0.022566024214029312, + 0.006186457816511393, + 0.009591291658580303, + 0.007917235605418682, + 0.010569177567958832, + 0.036620695143938065, + 0.06318756192922592, + 0.01727352850139141, + 0.035121649503707886, + 0.01713218353688717, + 0.025139711797237396, + 0.019507892429828644, + 0.026726389303803444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.11300762742757797 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a level angle with the camera moving slowly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young hero seated cross-legged at the center of a massive lotus throne?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero have long black hair bound into two buns by red ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hero's eyelids slowly lift to reveal eyes emitting a brilliant golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lotus throne carved from white jade and emitting a soft holy light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11310472339391708 + ], + "opposite_logprob_tokens": [ + -2.238104820251465 + ], + "logprob_reward": -0.11310472339391708, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11310472339391708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36776360869407654, + 0.1616324484348297, + 0.04768214374780655, + 0.04763089865446091, + 0.04887973889708519, + 0.04074571281671524, + 0.09021025896072388, + 0.06056967377662659, + 0.04954463616013527, + 0.08534091711044312 + ], + "credit_map": [ + 0.06908325105905533, + 0.03036942519247532, + 0.05626463145017624, + 0.027525296434760094, + 0.029423516243696213, + 0.06591466069221497, + 0.033052630722522736, + 0.05613017454743385, + 0.027352139353752136, + 0.008709605783224106, + 0.014415762387216091, + 0.013395451940596104, + 0.010917335748672485, + 0.041579652577638626, + 0.015377268195152283, + 0.02988523617386818, + 0.008048106916248798, + 0.006645960733294487, + 0.004392750561237335, + 0.0054556275717914104, + 0.0035697882995009422, + 0.010371821001172066, + 0.004361799918115139, + 0.00483628548681736, + 0.007777522783726454, + 0.006682002916932106, + 0.003913856111466885, + 0.004958013072609901, + 0.0033586667850613594, + 0.011929446831345558, + 0.00436668936163187, + 0.004644699860364199, + 0.006770751904696226, + 0.008532867766916752, + 0.005673909094184637, + 0.004955270793288946, + 0.003310712520033121, + 0.009301972575485706, + 0.004633849021047354, + 0.005700405687093735, + 0.005230283830314875, + 0.007024421822279692, + 0.0034822882153093815, + 0.003925353288650513, + 0.0043449304066598415, + 0.008707386441528797, + 0.0035215970128774643, + 0.004509450867772102, + 0.01594058983027935, + 0.011559986509382725, + 0.008862222544848919, + 0.008898785337805748, + 0.005430866032838821, + 0.014957535080611706, + 0.007457273546606302, + 0.017103001475334167, + 0.006684361957013607, + 0.010625229217112064, + 0.006280203815549612, + 0.006528046913444996, + 0.004636273253709078, + 0.007686604745686054, + 0.005009844899177551, + 0.013119108974933624, + 0.004045057110488415, + 0.008703421801328659, + 0.003280172124505043, + 0.01176576130092144, + 0.003438808722421527, + 0.0060433452017605305, + 0.0040472112596035, + 0.008220859803259373, + 0.008526264689862728, + 0.02583203837275505, + 0.00544540723785758, + 0.013511612080037594, + 0.007152384612709284, + 0.008538768626749516, + 0.005931852851063013, + 0.010402587242424488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.11310472339391708 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_9/index_tensor([4910, 4910])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_9/index_tensor([4910, 4910])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..634a18fbf1417cc4b75a063a2e109892acb5d70d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_9/index_tensor([4910, 4910])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping, high-angle drone shot smoothly tracks a vintage, faded cherry-red muscle car as it drives steadily forward down the center of a long, straight dirt road. The car's metal body is coated in a thin layer of light-colored dust, dulling its sheen. As it drives relentlessly, its tires kick up thick, billowing clouds of beige dust that snake along its wake and hang in the still air. The road itself carves a perfect, unwavering course across a vast, arid landscape, a flat land dotted with sparse, hardy desert shrubs and tufts of dry, yellowing grass. In the distance, casting sharp, dark shadows in the glare of the midday sun, a majestic mountain range looms on the horizon, its jagged peaks hued in purple against a pale, cloudless sky. The warm, earthy tones and the car's solitary, unwavering progress create a powerful cinematic atmosphere of isolation, freedom, and an endless, unwavering journey.", + "seed": 13897, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_9/index_4910_seed_13897_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.030415270617231725, + "vlm_logprob": -0.030415270617231725, + "vlm_yes_no_margin": 4.3249999523162845, + "vlm_token_credit": -0.030415270617231725, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vintage, faded cherry-red muscle car driving steadily forward down the center of the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's metal body coated in a thin layer of light-colored dust that dulls its sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's tires kick up thick, billowing clouds of beige dust that snake along its wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, arid landscape with a long, straight dirt road and a distant mountain range with purple peaks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11141536384820938 + ], + "opposite_logprob_tokens": [ + -2.486415386199951 + ], + "logprob_reward": -0.11141536384820938, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.11141536384820938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13934247195720673, + 0.04939570650458336, + 0.04784497246146202, + 0.21965423226356506, + 0.13027043640613556, + 0.04130256921052933, + 0.04625045508146286, + 0.18934151530265808, + 0.027782585471868515, + 0.1088150292634964 + ], + "credit_map": [ + 0.018341250717639923, + 0.01636628247797489, + 0.013203699141740799, + 0.017627088353037834, + 0.017352886497974396, + 0.027710676193237305, + 0.011427830904722214, + 0.01731276325881481, + 0.00633305823430419, + 0.0074235363863408566, + 0.004397382959723473, + 0.006941070314496756, + 0.006568775977939367, + 0.0074418107978999615, + 0.003877844661474228, + 0.006412225775420666, + 0.003921884577721357, + 0.004232986830174923, + 0.005482501350343227, + 0.009774021804332733, + 0.00536688044667244, + 0.004082539118826389, + 0.004010756500065327, + 0.010973402298986912, + 0.03172033280134201, + 0.015846820548176765, + 0.020745664834976196, + 0.03958059102296829, + 0.016133828088641167, + 0.03117992728948593, + 0.012239320203661919, + 0.05220774933695793, + 0.022041259333491325, + 0.012440959922969341, + 0.017445525154471397, + 0.024775013327598572, + 0.011460704728960991, + 0.017955627292394638, + 0.0061545176431536674, + 0.017996827140450478, + 0.004637848120182753, + 0.005034907720983028, + 0.004405945539474487, + 0.006115076132118702, + 0.004538168199360371, + 0.006567865144461393, + 0.0038279995787888765, + 0.006174759939312935, + 0.0033454163931310177, + 0.004648159258067608, + 0.006250106263905764, + 0.006091711111366749, + 0.005179950036108494, + 0.006204910576343536, + 0.008198430761694908, + 0.006331768352538347, + 0.012124603614211082, + 0.02443281002342701, + 0.04785536974668503, + 0.0188638623803854, + 0.008034806698560715, + 0.05671286582946777, + 0.010382942855358124, + 0.0109342560172081, + 0.004139973782002926, + 0.003973296377807856, + 0.0052481601014733315, + 0.00318344309926033, + 0.0022536516189575195, + 0.0035691827069967985, + 0.0030112965032458305, + 0.0024035831447690725, + 0.010924505069851875, + 0.026919908821582794, + 0.011796281673014164, + 0.01833282597362995, + 0.008081203326582909, + 0.012369146570563316, + 0.009863936342298985, + 0.010527226142585278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.11141536384820938 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vintage, faded cherry-red muscle car driving steadily forward down the center of the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's metal body coated in a thin layer of light-colored dust that dulls its sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's tires kick up thick, billowing clouds of beige dust that snake along its wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, arid landscape with a long, straight dirt road and a distant mountain range with purple peaks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014221273362636566 + ], + "opposite_logprob_tokens": [ + -4.26422119140625 + ], + "logprob_reward": -0.014221273362636566, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014221273362636566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19252870976924896, + 0.09172675758600235, + 0.049204446375370026, + 0.07805436104536057, + 0.12886373698711395, + 0.055489376187324524, + 0.04070736840367317, + 0.14935831725597382, + 0.09903102368116379, + 0.11503584682941437 + ], + "credit_map": [ + 0.019665637984871864, + 0.030285967513918877, + 0.017047781497240067, + 0.02793698199093342, + 0.021662982180714607, + 0.038456812500953674, + 0.01133818831294775, + 0.026134371757507324, + 0.0065275016240775585, + 0.010644062422215939, + 0.010165859013795853, + 0.013480077497661114, + 0.01605825312435627, + 0.00838771928101778, + 0.013181628659367561, + 0.013281663879752159, + 0.005548559594899416, + 0.005590751767158508, + 0.004426938481628895, + 0.011441067792475224, + 0.00420875521376729, + 0.005895834416151047, + 0.0036929266061633825, + 0.00839961413294077, + 0.008645478636026382, + 0.011259187944233418, + 0.009291405789554119, + 0.014858686365187168, + 0.007251191418617964, + 0.01102270558476448, + 0.0055172923021018505, + 0.010208416730165482, + 0.025104612112045288, + 0.0160741675645113, + 0.018061887472867966, + 0.01795404963195324, + 0.010477598756551743, + 0.017967894673347473, + 0.006145491264760494, + 0.01707804761826992, + 0.004910267423838377, + 0.00436263857409358, + 0.006225217133760452, + 0.008305183611810207, + 0.007123344112187624, + 0.006317093502730131, + 0.011235433630645275, + 0.007010206114500761, + 0.003033204236999154, + 0.003345661098137498, + 0.0054014138877391815, + 0.0052655357867479324, + 0.004285236354917288, + 0.006034504622220993, + 0.006841832306236029, + 0.006499982438981533, + 0.012329846620559692, + 0.02548583410680294, + 0.01410418190062046, + 0.019288774579763412, + 0.029660316184163094, + 0.0171204786747694, + 0.01816426031291485, + 0.01320464164018631, + 0.009341852739453316, + 0.022269798442721367, + 0.01169334165751934, + 0.011701320298016071, + 0.009882754646241665, + 0.011810160242021084, + 0.01489013060927391, + 0.007441665977239609, + 0.008469754830002785, + 0.019357437267899513, + 0.013513026759028435, + 0.02024674043059349, + 0.011398672126233578, + 0.022678565233945847, + 0.010838557966053486, + 0.008533104322850704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.014221273362636566 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vintage, faded cherry-red muscle car driving steadily forward down the center of the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's metal body coated in a thin layer of light-colored dust that dulls its sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's tires kick up thick, billowing clouds of beige dust that snake along its wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, arid landscape with a long, straight dirt road and a distant mountain range with purple peaks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0181919876486063 + ], + "opposite_logprob_tokens": [ + -4.018191814422607 + ], + "logprob_reward": -0.0181919876486063, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.0181919876486063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15293468534946442, + 0.029215525835752487, + 0.03708385303616524, + 0.06348076462745667, + 0.04095432162284851, + 0.029241694137454033, + 0.034914158284664154, + 0.5071113109588623, + 0.04191168397665024, + 0.0631520003080368 + ], + "credit_map": [ + 0.010072428733110428, + 0.015214084647595882, + 0.014699441380798817, + 0.027685826644301414, + 0.02154257893562317, + 0.02180900238454342, + 0.02106085978448391, + 0.020850470289587975, + 0.003165076021105051, + 0.003600532654672861, + 0.003347340039908886, + 0.005161521025002003, + 0.0033545184414833784, + 0.003847510553896427, + 0.001910670194774866, + 0.004828356672078371, + 0.003480254439637065, + 0.006614919286221266, + 0.0037745751906186342, + 0.005868738982826471, + 0.0035970620810985565, + 0.005657517816871405, + 0.002211323007941246, + 0.005879461299628019, + 0.008291730657219887, + 0.004276175517588854, + 0.003501362632960081, + 0.013275468721985817, + 0.003841891884803772, + 0.009313623420894146, + 0.0025730321649461985, + 0.01840747520327568, + 0.004368571564555168, + 0.008053758181631565, + 0.00403604144230485, + 0.007114197127521038, + 0.00475224619731307, + 0.0049001798033714294, + 0.002912160474807024, + 0.004817167762666941, + 0.0029869303107261658, + 0.004468484781682491, + 0.0035145983565598726, + 0.004403605591505766, + 0.0035484537947922945, + 0.004205022938549519, + 0.002857816405594349, + 0.0032567824237048626, + 0.0027010333724319935, + 0.0037765989545732737, + 0.004064999055117369, + 0.004984816536307335, + 0.0043164752423763275, + 0.004030295647680759, + 0.006241230759769678, + 0.0047987098805606365, + 0.03269875422120094, + 0.0706111416220665, + 0.12174899131059647, + 0.05674026161432266, + 0.021050922572612762, + 0.1497134268283844, + 0.025223178789019585, + 0.029324667528271675, + 0.0044044614769518375, + 0.010679468512535095, + 0.004229486919939518, + 0.006469654385000467, + 0.0035741874016821384, + 0.006081725470721722, + 0.0035666010808199644, + 0.002906096400693059, + 0.005195043049752712, + 0.010693979449570179, + 0.006883898749947548, + 0.009238305501639843, + 0.006215088535100222, + 0.006497509777545929, + 0.010708257555961609, + 0.00771991815418005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.0181919876486063 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vintage, faded cherry-red muscle car driving steadily forward down the center of the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's metal body coated in a thin layer of light-colored dust that dulls its sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's tires kick up thick, billowing clouds of beige dust that snake along its wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, arid landscape with a long, straight dirt road and a distant mountain range with purple peaks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003614679677411914 + ], + "opposite_logprob_tokens": [ + -5.628614902496338 + ], + "logprob_reward": -0.003614679677411914, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003614679677411914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09936482459306717, + 0.03516490384936333, + 0.034896742552518845, + 0.04359622299671173, + 0.08319756388664246, + 0.032555390149354935, + 0.058660659939050674, + 0.2443641871213913, + 0.34448695182800293, + 0.02371242083609104 + ], + "credit_map": [ + 0.008698997087776661, + 0.007288356311619282, + 0.010824337601661682, + 0.021531296893954277, + 0.015760695561766624, + 0.008893050253391266, + 0.01202135905623436, + 0.014346739277243614, + 0.004776299465447664, + 0.006033850833773613, + 0.0028017826844006777, + 0.004305226728320122, + 0.0034907106310129166, + 0.00532894441857934, + 0.002695530652999878, + 0.005732562392950058, + 0.00299362326040864, + 0.006802958436310291, + 0.0030016833916306496, + 0.005906766280531883, + 0.0033386163413524628, + 0.005229802802205086, + 0.0014861247036606073, + 0.006137167569249868, + 0.0044992840848863125, + 0.004139536060392857, + 0.005042879842221737, + 0.009773931466042995, + 0.0032984931021928787, + 0.005613327492028475, + 0.002394672017544508, + 0.008834104984998703, + 0.00830736942589283, + 0.01623198762536049, + 0.007332399487495422, + 0.019197596237063408, + 0.006958870682865381, + 0.008745372295379639, + 0.007097192108631134, + 0.009326787665486336, + 0.002466610400006175, + 0.0026562316343188286, + 0.0034553699661046267, + 0.00477331830188632, + 0.0031606557313352823, + 0.005728320684283972, + 0.006322483066469431, + 0.003992405720055103, + 0.0048456680960953236, + 0.005618357565253973, + 0.0061053023673594, + 0.007471256423741579, + 0.00707546342164278, + 0.00820147804915905, + 0.010966796427965164, + 0.008376344107091427, + 0.015107514336705208, + 0.02617564983665943, + 0.059377867728471756, + 0.028378019109368324, + 0.011231244541704655, + 0.07610297948122025, + 0.012396539561450481, + 0.015594422817230225, + 0.03465172275900841, + 0.09769372642040253, + 0.03934900835156441, + 0.034216850996017456, + 0.027745187282562256, + 0.05902232602238655, + 0.026509132236242294, + 0.025299029424786568, + 0.0022488178219646215, + 0.005017939954996109, + 0.0027404094580560923, + 0.0033634479623287916, + 0.002583224792033434, + 0.0034870151430368423, + 0.0025044765789061785, + 0.0017670922679826617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.003614679677411914 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vintage, faded cherry-red muscle car driving steadily forward down the center of the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's metal body coated in a thin layer of light-colored dust that dulls its sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's tires kick up thick, billowing clouds of beige dust that snake along its wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, arid landscape with a long, straight dirt road and a distant mountain range with purple peaks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633048549294472 + ], + "opposite_logprob_tokens": [ + -5.379632949829102 + ], + "logprob_reward": -0.004633048549294472, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633048549294472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13280946016311646, + 0.13156788051128387, + 0.061082471162080765, + 0.4025648534297943, + 0.07545172423124313, + 0.023893846198916435, + 0.021175650879740715, + 0.030387073755264282, + 0.06226314976811409, + 0.05880393460392952 + ], + "credit_map": [ + 0.0143485302105546, + 0.016902349889278412, + 0.012743692845106125, + 0.017745491117239, + 0.01593746244907379, + 0.022901538759469986, + 0.014235923998057842, + 0.0179944708943367, + 0.012247218750417233, + 0.017049668356776237, + 0.010082103312015533, + 0.019498739391565323, + 0.023077717050909996, + 0.015236476436257362, + 0.016777243465185165, + 0.017598707228899002, + 0.004098803270608187, + 0.006087012123316526, + 0.0040925913490355015, + 0.010709370486438274, + 0.008832679130136967, + 0.008001197129487991, + 0.009772024117410183, + 0.009488791227340698, + 0.047204356640577316, + 0.035507794469594955, + 0.04049300774931908, + 0.052445847541093826, + 0.03384449705481529, + 0.10690778493881226, + 0.031073514372110367, + 0.055088017135858536, + 0.013604553416371346, + 0.0064754909835755825, + 0.00897145364433527, + 0.01315776165574789, + 0.00608963193371892, + 0.010497432202100754, + 0.006569425575435162, + 0.010085967369377613, + 0.0025554178282618523, + 0.0021828352473676205, + 0.0021690381690859795, + 0.003435012185946107, + 0.0026591939385980368, + 0.003938690759241581, + 0.003168887924402952, + 0.0037847671192139387, + 0.0018561818869784474, + 0.0026550174225121737, + 0.0032391997519880533, + 0.0038898810744285583, + 0.0017497573280707002, + 0.002548009855672717, + 0.0016174991615116596, + 0.003620101837441325, + 0.0021667215041816235, + 0.0037651124875992537, + 0.006887211464345455, + 0.0034048310481011868, + 0.0027523464523255825, + 0.004956066142767668, + 0.003679074114188552, + 0.0027757107745856047, + 0.006282520946115255, + 0.013257911428809166, + 0.006774746347218752, + 0.0069637419655919075, + 0.005407856311649084, + 0.00947970524430275, + 0.008110501803457737, + 0.005986163858324289, + 0.005482405889779329, + 0.01124167162925005, + 0.008583702147006989, + 0.008425756357610226, + 0.00626359274610877, + 0.007851086556911469, + 0.005635467823594809, + 0.005320245865732431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.004633048549294472 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, high-angle drone shot smoothly tracks a vintage, faded cherry-red muscle car as it drives steadily forward down the center of a long, straight dirt road. The car's metal body is coated in a thin layer of light-colored dust, dulling its sheen. As it drives relentlessly, its tires kick up thick, billowing clouds of beige dust that snake along its wake and hang in the still air. The road itself carves a perfect, unwavering course across a vast, arid landscape, a flat land dotted with sparse, hardy desert shrubs and tufts of dry, yellowing grass. In the distance, casting sharp, dark shadows in the glare of the midday sun, a majestic mountain range looms on the horizon, its jagged peaks hued in purple against a pale, cloudless sky. The warm, earthy tones and the car's solitary, unwavering progress create a powerful cinematic atmosphere of isolation, freedom, and an endless, unwavering journey.", + "seed": 13898, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_9/index_4910_seed_13898_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11269332319498063, + "vlm_logprob": -0.11269332319498063, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.11269332319498063, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vintage, faded cherry-red muscle car driving steadily forward down the center of the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's metal body coated in a thin layer of light-colored dust that dulls its sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's tires kick up thick, billowing clouds of beige dust that snake along its wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, arid landscape with a long, straight dirt road and a distant mountain range with purple peaks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4907378554344177 + ], + "opposite_logprob_tokens": [ + -0.9907378554344177 + ], + "logprob_reward": -0.4907378554344177, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4907378554344177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14596568048000336, + 0.043056245893239975, + 0.028661387041211128, + 0.08411896228790283, + 0.07173827290534973, + 0.2559453248977661, + 0.0455021932721138, + 0.09548928588628769, + 0.21837756037712097, + 0.011145097203552723 + ], + "credit_map": [ + 0.009707449935376644, + 0.012557167559862137, + 0.015626003965735435, + 0.019371727481484413, + 0.023473020642995834, + 0.019646309316158295, + 0.021478833630681038, + 0.024105172604322433, + 0.005087060853838921, + 0.004041265230625868, + 0.004179417155683041, + 0.005151623394340277, + 0.005854348186403513, + 0.007485919166356325, + 0.005807424895465374, + 0.005449185613542795, + 0.002747137565165758, + 0.0020103571005165577, + 0.0029825102537870407, + 0.005140835884958506, + 0.003705535316839814, + 0.0028743441216647625, + 0.004200425930321217, + 0.005000241566449404, + 0.009627561084926128, + 0.006594722159206867, + 0.012885341420769691, + 0.013185257092118263, + 0.006035792175680399, + 0.009393676184117794, + 0.01242101937532425, + 0.013975591398775578, + 0.009727787226438522, + 0.004925365559756756, + 0.009067146107554436, + 0.010634681209921837, + 0.006368023343384266, + 0.008468776009976864, + 0.010352611541748047, + 0.012193875387310982, + 0.035661205649375916, + 0.021353397518396378, + 0.03005925938487053, + 0.03847334533929825, + 0.016850637272000313, + 0.025042489171028137, + 0.041968487203121185, + 0.046536486595869064, + 0.006048145703971386, + 0.0031663975678384304, + 0.007533862255513668, + 0.005140288732945919, + 0.005690186284482479, + 0.005365540739148855, + 0.006387195084244013, + 0.006170576438307762, + 0.01026786770671606, + 0.005525769200176001, + 0.013872426934540272, + 0.016947925090789795, + 0.008944367058575153, + 0.009854137897491455, + 0.007515451870858669, + 0.022561341524124146, + 0.02999066561460495, + 0.014834510162472725, + 0.01942450739443302, + 0.05084719881415367, + 0.014273041859269142, + 0.023462923243641853, + 0.010759864002466202, + 0.054784856736660004, + 0.0011974632507190108, + 0.0012463985476642847, + 0.0015366279985755682, + 0.0012377117527648807, + 0.0013413520064204931, + 0.0016704937443137169, + 0.0013954418245702982, + 0.0015196086606010795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.4907378554344177 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vintage, faded cherry-red muscle car driving steadily forward down the center of the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's metal body coated in a thin layer of light-colored dust that dulls its sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's tires kick up thick, billowing clouds of beige dust that snake along its wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, arid landscape with a long, straight dirt road and a distant mountain range with purple peaks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009814579039812088 + ], + "opposite_logprob_tokens": [ + -4.634814739227295 + ], + "logprob_reward": -0.009814579039812088, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009814579039812088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24521391093730927, + 0.10506483167409897, + 0.0962764173746109, + 0.10405369102954865, + 0.044148873537778854, + 0.06088586151599884, + 0.09942988306283951, + 0.08195669949054718, + 0.0931430533528328, + 0.06982681155204773 + ], + "credit_map": [ + 0.021793825551867485, + 0.01829739287495613, + 0.03141377121210098, + 0.032064031809568405, + 0.04739834740757942, + 0.03869134187698364, + 0.029238836839795113, + 0.026316368952393532, + 0.008840925060212612, + 0.008409541100263596, + 0.006528316996991634, + 0.01530818734318018, + 0.028848635032773018, + 0.011379108764231205, + 0.013579459860920906, + 0.012170654721558094, + 0.007513047661632299, + 0.007201524917036295, + 0.01716604083776474, + 0.015008237212896347, + 0.014657915569841862, + 0.008470674976706505, + 0.013667840510606766, + 0.012591136619448662, + 0.007322895806282759, + 0.009140919893980026, + 0.012165149673819542, + 0.011811656877398491, + 0.019278334453701973, + 0.010595899075269699, + 0.014433574862778187, + 0.019305255264043808, + 0.00334557075984776, + 0.004165390972048044, + 0.0044118245132267475, + 0.006694222334772348, + 0.006946630775928497, + 0.0055128056555986404, + 0.006370965391397476, + 0.006701461970806122, + 0.0071878680028021336, + 0.004019164014607668, + 0.006990063935518265, + 0.009696287102997303, + 0.007466803304851055, + 0.005953922867774963, + 0.008817658759653568, + 0.010754094459116459, + 0.006762024015188217, + 0.007653102744370699, + 0.009551111608743668, + 0.015042521059513092, + 0.022606629878282547, + 0.008648471906781197, + 0.015491319820284843, + 0.013674695044755936, + 0.008458857424557209, + 0.007083017844706774, + 0.008741079829633236, + 0.011934282258152962, + 0.016299396753311157, + 0.008815382607281208, + 0.010501287877559662, + 0.010123396292328835, + 0.009471054188907146, + 0.005958646535873413, + 0.006868292111903429, + 0.0221369918435812, + 0.009269182570278645, + 0.009073496796190739, + 0.00669513875618577, + 0.023670243099331856, + 0.008810953237116337, + 0.008171542547643185, + 0.005485834553837776, + 0.010453145951032639, + 0.010193582624197006, + 0.010991676710546017, + 0.006720251403748989, + 0.00899982638657093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.009814579039812088 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vintage, faded cherry-red muscle car driving steadily forward down the center of the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's metal body coated in a thin layer of light-colored dust that dulls its sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's tires kick up thick, billowing clouds of beige dust that snake along its wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, arid landscape with a long, straight dirt road and a distant mountain range with purple peaks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020585693418979645 + ], + "opposite_logprob_tokens": [ + -3.895585775375366 + ], + "logprob_reward": -0.020585693418979645, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020585693418979645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12832191586494446, + 0.07470275461673737, + 0.08865771442651749, + 0.09270838648080826, + 0.24745681881904602, + 0.03589444234967232, + 0.03904363885521889, + 0.09699482470750809, + 0.07265029847621918, + 0.12356922030448914 + ], + "credit_map": [ + 0.01015995629131794, + 0.011263932101428509, + 0.017688211053609848, + 0.020842351019382477, + 0.01867261342704296, + 0.009787285700440407, + 0.01909005269408226, + 0.020817505195736885, + 0.005672065075486898, + 0.005822896491736174, + 0.012327130883932114, + 0.010188693180680275, + 0.010781762190163136, + 0.009197001345455647, + 0.010165522806346416, + 0.010547671467065811, + 0.009068889543414116, + 0.00519254244863987, + 0.015950217843055725, + 0.011624655686318874, + 0.014840349555015564, + 0.009461445733904839, + 0.01264229416847229, + 0.009877311997115612, + 0.009050254710018635, + 0.006354683544486761, + 0.01021900400519371, + 0.01185378897935152, + 0.015033194795250893, + 0.012676711194217205, + 0.013708213344216347, + 0.013812534511089325, + 0.02218407206237316, + 0.021506600081920624, + 0.0326552651822567, + 0.045029036700725555, + 0.02384028211236, + 0.020647505298256874, + 0.039592523127794266, + 0.042001523077487946, + 0.0039451513439416885, + 0.0031578128691762686, + 0.005473120138049126, + 0.005169657524675131, + 0.004544282332062721, + 0.0035130379255861044, + 0.006262913346290588, + 0.0038284671027213335, + 0.0035920313093811274, + 0.00308176688849926, + 0.0063897669315338135, + 0.004898188170045614, + 0.0065141040831804276, + 0.004527530167251825, + 0.005682780873030424, + 0.004357465542852879, + 0.015145271085202694, + 0.00779846403747797, + 0.012309560552239418, + 0.013241533190011978, + 0.009748830460011959, + 0.015721816569566727, + 0.009365689009428024, + 0.013663653284311295, + 0.008853617124259472, + 0.005741179920732975, + 0.008228128775954247, + 0.010320298373699188, + 0.01289814431220293, + 0.007389395497739315, + 0.00844250712543726, + 0.010777025483548641, + 0.015873730182647705, + 0.01644197106361389, + 0.013368815183639526, + 0.017495615407824516, + 0.013369986787438393, + 0.016454461961984634, + 0.008510309271514416, + 0.02205432765185833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.020585693418979645 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vintage, faded cherry-red muscle car driving steadily forward down the center of the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's metal body coated in a thin layer of light-colored dust that dulls its sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's tires kick up thick, billowing clouds of beige dust that snake along its wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, arid landscape with a long, straight dirt road and a distant mountain range with purple peaks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029798785224556923 + ], + "opposite_logprob_tokens": [ + -3.529798746109009 + ], + "logprob_reward": -0.029798785224556923, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029798785224556923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25245264172554016, + 0.07787207514047623, + 0.0983559638261795, + 0.09195485711097717, + 0.20367445051670074, + 0.03475864231586456, + 0.05685918405652046, + 0.07538560032844543, + 0.02244235947728157, + 0.08624422550201416 + ], + "credit_map": [ + 0.01731409691274166, + 0.02505458891391754, + 0.037022389471530914, + 0.03300710767507553, + 0.04168994352221489, + 0.024338586255908012, + 0.03593391552567482, + 0.03809201717376709, + 0.007570600602775812, + 0.005810980219393969, + 0.008619305677711964, + 0.013812750577926636, + 0.007052864879369736, + 0.008539567701518536, + 0.012632627040147781, + 0.013833379372954369, + 0.010375403799116611, + 0.006695230957120657, + 0.01602022536098957, + 0.013985791243612766, + 0.013720909133553505, + 0.012893818318843842, + 0.012417098507285118, + 0.012247486971318722, + 0.009843776933848858, + 0.0066064829006791115, + 0.01067548617720604, + 0.013803990557789803, + 0.012890071608126163, + 0.012562833726406097, + 0.01416188944131136, + 0.011410328559577465, + 0.01986762136220932, + 0.01506143156439066, + 0.02087644301354885, + 0.03648588806390762, + 0.016535665839910507, + 0.018517209216952324, + 0.03492158651351929, + 0.04140860214829445, + 0.004114426206797361, + 0.0024782870896160603, + 0.004746646620333195, + 0.006074848119169474, + 0.0025298509281128645, + 0.004453849047422409, + 0.004784425720572472, + 0.005576306954026222, + 0.005028849933296442, + 0.004164937883615494, + 0.006983178202062845, + 0.008031340315937996, + 0.0066596744582057, + 0.006399952340871096, + 0.008623228408396244, + 0.01096801832318306, + 0.010729466564953327, + 0.006719004828482866, + 0.008572570979595184, + 0.01236679870635271, + 0.008222800679504871, + 0.010011743754148483, + 0.008638494648039341, + 0.010124722495675087, + 0.002105376683175564, + 0.001552939647808671, + 0.0017386479303240776, + 0.004373334348201752, + 0.0027029013726860285, + 0.0022574758622795343, + 0.0017907212022691965, + 0.005920963827520609, + 0.01138217095285654, + 0.010788961313664913, + 0.008573981001973152, + 0.011660831980407238, + 0.011209269054234028, + 0.014133631251752377, + 0.005503718741238117, + 0.01299165841192007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.029798785224556923 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vintage, faded cherry-red muscle car driving steadily forward down the center of the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's metal body coated in a thin layer of light-colored dust that dulls its sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's tires kick up thick, billowing clouds of beige dust that snake along its wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, arid landscape with a long, straight dirt road and a distant mountain range with purple peaks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012529702857136726 + ], + "opposite_logprob_tokens": [ + -4.3875298500061035 + ], + "logprob_reward": -0.012529702857136726, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012529702857136726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4215121269226074, + 0.14969377219676971, + 0.04882403090596199, + 0.02048663981258869, + 0.06466784328222275, + 0.09713369607925415, + 0.04167452082037926, + 0.029975324869155884, + 0.061542075127363205, + 0.06448998302221298 + ], + "credit_map": [ + 0.042182840406894684, + 0.04340899735689163, + 0.02596311829984188, + 0.042237136512994766, + 0.08598949760198593, + 0.0843130573630333, + 0.05834248661994934, + 0.039074983447790146, + 0.014748713001608849, + 0.013239314779639244, + 0.013444905169308186, + 0.016654767096042633, + 0.029357876628637314, + 0.02364453114569187, + 0.021006617695093155, + 0.017597053200006485, + 0.00421247398480773, + 0.0031844149343669415, + 0.008642269298434258, + 0.006838303525000811, + 0.006721756886690855, + 0.00522970873862505, + 0.006371129769831896, + 0.007623974699527025, + 0.002349566901102662, + 0.001379435881972313, + 0.002184306038543582, + 0.002727716928347945, + 0.0035898564383387566, + 0.0019697367679327726, + 0.002598338993266225, + 0.0036876823287457228, + 0.007965793833136559, + 0.004539396148175001, + 0.007915349677205086, + 0.010176981799304485, + 0.005994372069835663, + 0.006836067885160446, + 0.009995128959417343, + 0.011244756169617176, + 0.012598087079823017, + 0.007966887205839157, + 0.00929444469511509, + 0.015193558298051357, + 0.006972112227231264, + 0.008756808936595917, + 0.016020983457565308, + 0.020330820232629776, + 0.0058805025182664394, + 0.0034689714666455984, + 0.007219761610031128, + 0.005020007491111755, + 0.0037369350902736187, + 0.005677028559148312, + 0.0046758512035012245, + 0.005995461251586676, + 0.0027152260299772024, + 0.0025748175103217363, + 0.0032407075632363558, + 0.00395598728209734, + 0.005415167659521103, + 0.0030435819644480944, + 0.00384184205904603, + 0.0051879966631531715, + 0.0074117304757237434, + 0.004868381191045046, + 0.0052686287090182304, + 0.012605372816324234, + 0.007575869094580412, + 0.005847002845257521, + 0.0047409855760633945, + 0.013224107213318348, + 0.008756877854466438, + 0.007916857488453388, + 0.005652091931551695, + 0.010407155379652977, + 0.0058651212602853775, + 0.012171536684036255, + 0.005775498691946268, + 0.007944847457110882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.012529702857136726 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, high-angle drone shot smoothly tracks a vintage, faded cherry-red muscle car as it drives steadily forward down the center of a long, straight dirt road. The car's metal body is coated in a thin layer of light-colored dust, dulling its sheen. As it drives relentlessly, its tires kick up thick, billowing clouds of beige dust that snake along its wake and hang in the still air. The road itself carves a perfect, unwavering course across a vast, arid landscape, a flat land dotted with sparse, hardy desert shrubs and tufts of dry, yellowing grass. In the distance, casting sharp, dark shadows in the glare of the midday sun, a majestic mountain range looms on the horizon, its jagged peaks hued in purple against a pale, cloudless sky. The warm, earthy tones and the car's solitary, unwavering progress create a powerful cinematic atmosphere of isolation, freedom, and an endless, unwavering journey.", + "seed": 13899, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_9/index_4910_seed_13899_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06497551668435335, + "vlm_logprob": -0.06497551668435335, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.06497551668435335, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vintage, faded cherry-red muscle car driving steadily forward down the center of the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's metal body coated in a thin layer of light-colored dust that dulls its sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's tires kick up thick, billowing clouds of beige dust that snake along its wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, arid landscape with a long, straight dirt road and a distant mountain range with purple peaks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.24668271839618683 + ], + "opposite_logprob_tokens": [ + -1.6216827630996704 + ], + "logprob_reward": -0.24668271839618683, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.24668271839618683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06962836533784866, + 0.20592302083969116, + 0.019429517909884453, + 0.016087017953395844, + 0.018244992941617966, + 0.031849347054958344, + 0.34978166222572327, + 0.0372270792722702, + 0.23620882630348206, + 0.015620185062289238 + ], + "credit_map": [ + 0.003995527513325214, + 0.005572554189711809, + 0.005252559669315815, + 0.014408942312002182, + 0.007877835072577, + 0.010408574715256691, + 0.01334286853671074, + 0.008769506588578224, + 0.0047273761592805386, + 0.005234298296272755, + 0.004835932515561581, + 0.08360783755779266, + 0.006679556332528591, + 0.004679831210523844, + 0.08569014072418213, + 0.010468046180903912, + 0.001670516561716795, + 0.0025356325786560774, + 0.0014142004074528813, + 0.0024265116080641747, + 0.0031566584948450327, + 0.003258545882999897, + 0.0025082596112042665, + 0.002459191018715501, + 0.0012751229805871844, + 0.001960898283869028, + 0.00152115139644593, + 0.0021666015964001417, + 0.0020206242334097624, + 0.0027080827858299017, + 0.0022680836264044046, + 0.002166452817618847, + 0.0015131763648241758, + 0.0020728129893541336, + 0.001851031556725502, + 0.0029048575088381767, + 0.00233283219859004, + 0.0025260134134441614, + 0.00220899167470634, + 0.002835277933627367, + 0.00300555769354105, + 0.002919477876275778, + 0.002825327217578888, + 0.00621512345969677, + 0.0041625308804214, + 0.003560817800462246, + 0.005478214472532272, + 0.0036823011469095945, + 0.009714021347463131, + 0.011258469894528389, + 0.010280138812959194, + 0.13437655568122864, + 0.017191603779792786, + 0.010881427675485611, + 0.13525508344173431, + 0.0208243690431118, + 0.003128662472590804, + 0.0038901502266526222, + 0.0034646152053028345, + 0.00457179918885231, + 0.005953955464065075, + 0.004760080017149448, + 0.005872717592865229, + 0.00558509910479188, + 0.027283640578389168, + 0.021605899557471275, + 0.03047306276857853, + 0.046056248247623444, + 0.019205505028367043, + 0.02875627763569355, + 0.038553960621356964, + 0.02427424117922783, + 0.0024636033922433853, + 0.0017058203229680657, + 0.001329999533481896, + 0.0021470265928655863, + 0.001670182216912508, + 0.002205393509939313, + 0.0016895618755370378, + 0.0024085973855108023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 1 + ], + "selected_reward": -0.24668271839618683 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vintage, faded cherry-red muscle car driving steadily forward down the center of the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's metal body coated in a thin layer of light-colored dust that dulls its sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's tires kick up thick, billowing clouds of beige dust that snake along its wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, arid landscape with a long, straight dirt road and a distant mountain range with purple peaks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01256749127060175 + ], + "opposite_logprob_tokens": [ + -4.387567520141602 + ], + "logprob_reward": -0.01256749127060175, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01256749127060175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0912831723690033, + 0.15295031666755676, + 0.0559670552611351, + 0.035936422646045685, + 0.04516688361763954, + 0.042944058775901794, + 0.39286404848098755, + 0.07856054604053497, + 0.0634050965309143, + 0.04092241823673248 + ], + "credit_map": [ + 0.005967630539089441, + 0.008772051893174648, + 0.013203070499002934, + 0.01537795178592205, + 0.009217353537678719, + 0.014368677511811256, + 0.0137984249740839, + 0.010578007437288761, + 0.007711509708315134, + 0.012728952802717686, + 0.007452911697328091, + 0.03465710207819939, + 0.014224249869585037, + 0.014587637037038803, + 0.04294585436582565, + 0.018642080947756767, + 0.005019154399633408, + 0.006359864957630634, + 0.004144884180277586, + 0.008176811039447784, + 0.009941399097442627, + 0.0065246340818703175, + 0.009966853074729443, + 0.005833445582538843, + 0.002922971732914448, + 0.004010882694274187, + 0.003624228062108159, + 0.005939811002463102, + 0.005462752189487219, + 0.004404167179018259, + 0.004430894739925861, + 0.005140712019056082, + 0.00299476133659482, + 0.0036717138718813658, + 0.003930344711989164, + 0.009748347103595734, + 0.00588785856962204, + 0.005506690125912428, + 0.008152886293828487, + 0.005274274852126837, + 0.0037169698625802994, + 0.002947122324258089, + 0.001895778812468052, + 0.010220879688858986, + 0.005013625603169203, + 0.005413800943642855, + 0.01013545785099268, + 0.003600422292947769, + 0.010525929741561413, + 0.013067005202174187, + 0.013664893805980682, + 0.14624881744384766, + 0.027174003422260284, + 0.013149762526154518, + 0.14678411185741425, + 0.02224947139620781, + 0.0063557615503668785, + 0.008226349949836731, + 0.00794064812362194, + 0.007929928600788116, + 0.016179166734218597, + 0.009513051249086857, + 0.012317925691604614, + 0.010097706690430641, + 0.004538274835795164, + 0.00474824383854866, + 0.005582520272582769, + 0.008000980131328106, + 0.008959559723734856, + 0.00839115958660841, + 0.01352779008448124, + 0.00965656153857708, + 0.003178409533575177, + 0.0024854252114892006, + 0.0027709435671567917, + 0.008903414011001587, + 0.004642354790121317, + 0.005531314294785261, + 0.00510952016338706, + 0.008301034569740295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.01256749127060175 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vintage, faded cherry-red muscle car driving steadily forward down the center of the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's metal body coated in a thin layer of light-colored dust that dulls its sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's tires kick up thick, billowing clouds of beige dust that snake along its wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, arid landscape with a long, straight dirt road and a distant mountain range with purple peaks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329809032380581 + ], + "opposite_logprob_tokens": [ + -3.7732980251312256 + ], + "logprob_reward": -0.02329809032380581, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329809032380581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1683635413646698, + 0.435891717672348, + 0.03691117465496063, + 0.03738310560584068, + 0.05357380956411362, + 0.03431450203061104, + 0.12720005214214325, + 0.04350276291370392, + 0.020554840564727783, + 0.04230448231101036 + ], + "credit_map": [ + 0.008829302154481411, + 0.01681181974709034, + 0.015380318276584148, + 0.031136002391576767, + 0.025913257151842117, + 0.01881580986082554, + 0.037349507212638855, + 0.014127511531114578, + 0.011194886639714241, + 0.017851103097200394, + 0.016215452924370766, + 0.16128014028072357, + 0.02361985296010971, + 0.012833555229008198, + 0.15894049406051636, + 0.03395622596144676, + 0.0022526083048433065, + 0.003708341158926487, + 0.003970674239099026, + 0.006645645014941692, + 0.00448719784617424, + 0.004554198123514652, + 0.0055925422348082066, + 0.005699969362467527, + 0.002957726828753948, + 0.0030725905671715736, + 0.0034129885025322437, + 0.0076086996123194695, + 0.0041127391159534454, + 0.004395696334540844, + 0.006182733923196793, + 0.005639929790049791, + 0.003999279346317053, + 0.005268989596515894, + 0.0048592365346848965, + 0.009897880256175995, + 0.008255621418356895, + 0.005569410044699907, + 0.007869728840887547, + 0.00785366166383028, + 0.0014573147054761648, + 0.0016332188388332725, + 0.0016517022158950567, + 0.01137138344347477, + 0.0027176232542842627, + 0.0017932982882484794, + 0.01101062074303627, + 0.0026793412398546934, + 0.0027605807408690453, + 0.0035702588502317667, + 0.002655512886121869, + 0.0520329624414444, + 0.005436096340417862, + 0.002094613155350089, + 0.05158497020602226, + 0.007065064273774624, + 0.005536022130399942, + 0.004411558620631695, + 0.003787178546190262, + 0.006353749893605709, + 0.006079509388655424, + 0.005699859466403723, + 0.006662821397185326, + 0.0049720644019544125, + 0.001451278803870082, + 0.0017285890644416213, + 0.002546048955991864, + 0.003419300075620413, + 0.002904052147641778, + 0.0021639184560626745, + 0.003331598360091448, + 0.003010054351761937, + 0.003023884491994977, + 0.0016370847588405013, + 0.0024870336055755615, + 0.015119488351047039, + 0.0018770707538351417, + 0.0028749718330800533, + 0.003070057136937976, + 0.012214891612529755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.02329809032380581 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vintage, faded cherry-red muscle car driving steadily forward down the center of the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's metal body coated in a thin layer of light-colored dust that dulls its sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's tires kick up thick, billowing clouds of beige dust that snake along its wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, arid landscape with a long, straight dirt road and a distant mountain range with purple peaks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369821608066559 + ], + "opposite_logprob_tokens": [ + -3.408698320388794 + ], + "logprob_reward": -0.03369821608066559, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369821608066559, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16384297609329224, + 0.5814867615699768, + 0.05483599752187729, + 0.01642611064016819, + 0.04990500956773758, + 0.030669696629047394, + 0.04201854020357132, + 0.03438892588019371, + 0.008889508433640003, + 0.01753644272685051 + ], + "credit_map": [ + 0.007214120123535395, + 0.01659361459314823, + 0.01690221205353737, + 0.02914217673242092, + 0.024053074419498444, + 0.01988901011645794, + 0.03485045209527016, + 0.01519833318889141, + 0.00999997928738594, + 0.019720355048775673, + 0.01584281586110592, + 0.23546090722084045, + 0.022448547184467316, + 0.013449428603053093, + 0.2323450744152069, + 0.03221967816352844, + 0.0034397488925606012, + 0.006345278583467007, + 0.0051956805400550365, + 0.008533358573913574, + 0.008331216871738434, + 0.0069589014165103436, + 0.008198440074920654, + 0.007833367213606834, + 0.0014311198610812426, + 0.001352578168734908, + 0.0016442147316411138, + 0.0032019754871726036, + 0.0017859172075986862, + 0.001918133464641869, + 0.0023608130868524313, + 0.0027313600294291973, + 0.0032458738423883915, + 0.004786285106092691, + 0.004813128616660833, + 0.012296469882130623, + 0.005194024182856083, + 0.0042498912662267685, + 0.007745375856757164, + 0.007573961280286312, + 0.0016529022250324488, + 0.0016628149896860123, + 0.0023713973350822926, + 0.008483245968818665, + 0.002778991824015975, + 0.0020935693755745888, + 0.008256136439740658, + 0.0033706407994031906, + 0.0018143646884709597, + 0.002068630652502179, + 0.002601941814646125, + 0.01387077383697033, + 0.0028775411192327738, + 0.002723921090364456, + 0.013419335708022118, + 0.002642030129209161, + 0.0027992569375783205, + 0.0034465682692825794, + 0.0028864319901913404, + 0.005396833643317223, + 0.0040468256920576096, + 0.00485148373991251, + 0.006066794041544199, + 0.004894730169326067, + 0.0006119511090219021, + 0.0008232314139604568, + 0.0007813042029738426, + 0.002185257151722908, + 0.0009574173600412905, + 0.0007798626320436597, + 0.0012435228563845158, + 0.0015069622313603759, + 0.0019239383982494473, + 0.0015646327519789338, + 0.002193323103711009, + 0.0020597141701728106, + 0.002450976986438036, + 0.0029598120599985123, + 0.0019207869190722704, + 0.0024632588028907776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.03369821608066559 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vintage, faded cherry-red muscle car driving steadily forward down the center of the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's metal body coated in a thin layer of light-colored dust that dulls its sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's tires kick up thick, billowing clouds of beige dust that snake along its wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, arid landscape with a long, straight dirt road and a distant mountain range with purple peaks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008631067350506783 + ], + "opposite_logprob_tokens": [ + -4.758631229400635 + ], + "logprob_reward": -0.008631067350506783, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008631067350506783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18011444807052612, + 0.4675385653972626, + 0.0154220936819911, + 0.029128817841410637, + 0.019042914733290672, + 0.016323566436767578, + 0.1697329729795456, + 0.03504442423582077, + 0.028688590973615646, + 0.03896358981728554 + ], + "credit_map": [ + 0.01383923925459385, + 0.01957046613097191, + 0.008677342906594276, + 0.025621678680181503, + 0.03638272359967232, + 0.02781428024172783, + 0.02899935469031334, + 0.019209347665309906, + 0.0095474598929286, + 0.014047968201339245, + 0.01020941324532032, + 0.19422189891338348, + 0.014288943260908127, + 0.01028850581496954, + 0.19487766921520233, + 0.02005673572421074, + 0.0011111160274595022, + 0.001847527688369155, + 0.001322218682616949, + 0.0025705473963171244, + 0.0024895425885915756, + 0.0026535852812230587, + 0.0020373594015836716, + 0.0013901975471526384, + 0.002164503326639533, + 0.003411554265767336, + 0.0018259662901982665, + 0.004096782300621271, + 0.00511751975864172, + 0.005407152231782675, + 0.0035250112414360046, + 0.003580328542739153, + 0.0013606559950858355, + 0.001962351845577359, + 0.0013474441366270185, + 0.003225891850888729, + 0.004296030383557081, + 0.0020311831030994654, + 0.0032000229693949223, + 0.0016193340998142958, + 0.0014350076671689749, + 0.0012942133471369743, + 0.0013107635313645005, + 0.0031722048297524452, + 0.0027586095966398716, + 0.0013263191794976592, + 0.0024318331852555275, + 0.0025946146342903376, + 0.0033878961112350225, + 0.005056038033217192, + 0.004060766659677029, + 0.06734249740839005, + 0.00867713987827301, + 0.00376001070253551, + 0.06842950731515884, + 0.00901912059634924, + 0.004333991091698408, + 0.0042382399551570415, + 0.002205323427915573, + 0.0047127725556492805, + 0.004731958266347647, + 0.005367126315832138, + 0.004786291625350714, + 0.0046687196008861065, + 0.0021434600930660963, + 0.002824854338541627, + 0.0028593242168426514, + 0.0046693431213498116, + 0.003466743277385831, + 0.002971225418150425, + 0.005647069308906794, + 0.004106571897864342, + 0.00516708567738533, + 0.0029225251637399197, + 0.002950879978016019, + 0.009785301052033901, + 0.0036999438889324665, + 0.004430670291185379, + 0.0042367978021502495, + 0.005770385265350342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.008631067350506783 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, high-angle drone shot smoothly tracks a vintage, faded cherry-red muscle car as it drives steadily forward down the center of a long, straight dirt road. The car's metal body is coated in a thin layer of light-colored dust, dulling its sheen. As it drives relentlessly, its tires kick up thick, billowing clouds of beige dust that snake along its wake and hang in the still air. The road itself carves a perfect, unwavering course across a vast, arid landscape, a flat land dotted with sparse, hardy desert shrubs and tufts of dry, yellowing grass. In the distance, casting sharp, dark shadows in the glare of the midday sun, a majestic mountain range looms on the horizon, its jagged peaks hued in purple against a pale, cloudless sky. The warm, earthy tones and the car's solitary, unwavering progress create a powerful cinematic atmosphere of isolation, freedom, and an endless, unwavering journey.", + "seed": 13900, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_9/index_4910_seed_13900_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03632725076749921, + "vlm_logprob": -0.03632725076749921, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.03632725076749921, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vintage, faded cherry-red muscle car driving steadily forward down the center of the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's metal body coated in a thin layer of light-colored dust that dulls its sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's tires kick up thick, billowing clouds of beige dust that snake along its wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, arid landscape with a long, straight dirt road and a distant mountain range with purple peaks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09584314376115799 + ], + "opposite_logprob_tokens": [ + -2.7208430767059326 + ], + "logprob_reward": -0.09584314376115799, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.09584314376115799, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08955125510692596, + 0.11721004545688629, + 0.08152199536561966, + 0.06889957189559937, + 0.03895888477563858, + 0.05889531970024109, + 0.34525638818740845, + 0.02193388342857361, + 0.16387531161308289, + 0.013897295109927654 + ], + "credit_map": [ + 0.013377347029745579, + 0.009774222038686275, + 0.006160742603242397, + 0.007621877361088991, + 0.01724584959447384, + 0.015980495139956474, + 0.010844700038433075, + 0.008546026423573494, + 0.011098209768533707, + 0.008974604308605194, + 0.0069100577384233475, + 0.040768008679151535, + 0.005390915088355541, + 0.010858830064535141, + 0.005165317095816135, + 0.028044117614626884, + 0.01893681101500988, + 0.010966221801936626, + 0.016537124291062355, + 0.010662231594324112, + 0.007809937931597233, + 0.008856705389916897, + 0.004328742157667875, + 0.003424233989790082, + 0.01214651670306921, + 0.0111214779317379, + 0.009747976437211037, + 0.006992759183049202, + 0.006049801129847765, + 0.012466393411159515, + 0.004488855134695768, + 0.005885798018425703, + 0.007732612080872059, + 0.005565436091274023, + 0.009760168381035328, + 0.004492642357945442, + 0.0025822636671364307, + 0.004039706662297249, + 0.0021102502942085266, + 0.002675809897482395, + 0.011128541082143784, + 0.005429296754300594, + 0.018252704292535782, + 0.008475804701447487, + 0.0032179872505366802, + 0.00668643182143569, + 0.0029656169936060905, + 0.0027389379683882, + 0.020877612754702568, + 0.0181636493653059, + 0.05859930440783501, + 0.04126650094985962, + 0.02606566809117794, + 0.14933447539806366, + 0.013073704205453396, + 0.01787549816071987, + 0.0023873415775597095, + 0.0016553513705730438, + 0.002791422186419368, + 0.004903698805719614, + 0.0027586568612605333, + 0.0033796322531998158, + 0.002671665744856, + 0.0013861171901226044, + 0.018775638192892075, + 0.023002343252301216, + 0.028470173478126526, + 0.01797969453036785, + 0.01551330741494894, + 0.04641106352210045, + 0.007714509963989258, + 0.006008596625179052, + 0.0027724041137844324, + 0.0010819954331964254, + 0.0038439705967903137, + 0.0011007805587723851, + 0.0009719221852719784, + 0.002379537792876363, + 0.0009447017218917608, + 0.0008019849774427712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 1 + ], + "selected_reward": -0.09584314376115799 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vintage, faded cherry-red muscle car driving steadily forward down the center of the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's metal body coated in a thin layer of light-colored dust that dulls its sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's tires kick up thick, billowing clouds of beige dust that snake along its wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, arid landscape with a long, straight dirt road and a distant mountain range with purple peaks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009813752956688404 + ], + "opposite_logprob_tokens": [ + -4.6348137855529785 + ], + "logprob_reward": -0.009813752956688404, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009813752956688404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.172511488199234, + 0.16089558601379395, + 0.08634036034345627, + 0.027387840673327446, + 0.04757989943027496, + 0.0880548357963562, + 0.09635107964277267, + 0.08592429757118225, + 0.18475794792175293, + 0.050196655094623566 + ], + "credit_map": [ + 0.025706235319375992, + 0.020603151991963387, + 0.017849460244178772, + 0.0212198868393898, + 0.019274575635790825, + 0.040907785296440125, + 0.014764918014407158, + 0.012185473926365376, + 0.0282454751431942, + 0.014507969841361046, + 0.019927237182855606, + 0.05191374570131302, + 0.009480034932494164, + 0.018127810209989548, + 0.010733503848314285, + 0.007959817536175251, + 0.010025141760706902, + 0.010524071753025055, + 0.017680993303656578, + 0.016283003613352776, + 0.009900645352900028, + 0.012359538115561008, + 0.004955443553626537, + 0.004611527547240257, + 0.004198944661766291, + 0.003944516647607088, + 0.0024268035776913166, + 0.0034262374974787235, + 0.002966440049931407, + 0.00474619772285223, + 0.003066519508138299, + 0.002612181007862091, + 0.005279983393847942, + 0.006722555961459875, + 0.005027019418776035, + 0.006991251837462187, + 0.006696029100567102, + 0.006878744810819626, + 0.005630283150821924, + 0.004354030359536409, + 0.016133824363350868, + 0.012971287593245506, + 0.020839489996433258, + 0.010564848780632019, + 0.0055685690604150295, + 0.0118612851947546, + 0.006044033449143171, + 0.00407149875536561, + 0.018423791974782944, + 0.009682897478342056, + 0.016687458381056786, + 0.014955401420593262, + 0.01153829600661993, + 0.009465587325394154, + 0.008624281734228134, + 0.006973365321755409, + 0.010222205892205238, + 0.008614190854132175, + 0.007963309995830059, + 0.013112004846334457, + 0.014141250401735306, + 0.015537315979599953, + 0.010925808921456337, + 0.005408214405179024, + 0.04341140761971474, + 0.015225975774228573, + 0.0199968833476305, + 0.01766376569867134, + 0.013583765365183353, + 0.055045682936906815, + 0.009772908873856068, + 0.010057555511593819, + 0.0039613451808691025, + 0.0044244928285479546, + 0.0048894817009568214, + 0.004888934548944235, + 0.008205734193325043, + 0.015251106582581997, + 0.005259566940367222, + 0.003315993119031191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.009813752956688404 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vintage, faded cherry-red muscle car driving steadily forward down the center of the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's metal body coated in a thin layer of light-colored dust that dulls its sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's tires kick up thick, billowing clouds of beige dust that snake along its wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, arid landscape with a long, straight dirt road and a distant mountain range with purple peaks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866408929228783 + ], + "opposite_logprob_tokens": [ + -3.048664093017578 + ], + "logprob_reward": -0.04866408929228783, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866408929228783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16319549083709717, + 0.24906903505325317, + 0.0866822600364685, + 0.11187107115983963, + 0.05958821251988411, + 0.08639675378799438, + 0.051722146570682526, + 0.06219268962740898, + 0.0862104520201683, + 0.0430719293653965 + ], + "credit_map": [ + 0.02644789218902588, + 0.016525253653526306, + 0.017700161784887314, + 0.021914755925536156, + 0.011070801876485348, + 0.048363011330366135, + 0.009818022139370441, + 0.011355599388480186, + 0.02750728279352188, + 0.018461786210536957, + 0.034431226551532745, + 0.0911203920841217, + 0.015032723546028137, + 0.02774040773510933, + 0.011745964176952839, + 0.02302926778793335, + 0.01027181837707758, + 0.011389560997486115, + 0.011360486969351768, + 0.01748495176434517, + 0.008216372691094875, + 0.012123147957026958, + 0.006553218234330416, + 0.009282704442739487, + 0.016356086358428, + 0.019082086160779, + 0.020921042189002037, + 0.014612138271331787, + 0.012216421775519848, + 0.016827957704663277, + 0.005730717908591032, + 0.006124619394540787, + 0.006869391072541475, + 0.010390492156147957, + 0.005375893320888281, + 0.008164551109075546, + 0.005494320765137672, + 0.013895430602133274, + 0.004263342823833227, + 0.005134786479175091, + 0.012347790412604809, + 0.01134063396602869, + 0.0207111444324255, + 0.010771889239549637, + 0.006537436507642269, + 0.01487736590206623, + 0.005080681759864092, + 0.004729808308184147, + 0.005663043819367886, + 0.00413915142416954, + 0.010298297740519047, + 0.0049517154693603516, + 0.004254964645951986, + 0.01692088693380356, + 0.002817321103066206, + 0.0026767663657665253, + 0.007119627669453621, + 0.00750800920650363, + 0.0061136092990636826, + 0.006928815972059965, + 0.010208304971456528, + 0.010175677947700024, + 0.00852583535015583, + 0.0056128110736608505, + 0.018156135454773903, + 0.0062913112342357635, + 0.005461591761559248, + 0.0162950586527586, + 0.011940238997340202, + 0.017561152577400208, + 0.0064712692983448505, + 0.004033686593174934, + 0.0034509520046412945, + 0.002403881633654237, + 0.007572474889457226, + 0.004534048028290272, + 0.005142799112945795, + 0.01570012979209423, + 0.002455380978062749, + 0.001812263042666018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.04866408929228783 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vintage, faded cherry-red muscle car driving steadily forward down the center of the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's metal body coated in a thin layer of light-colored dust that dulls its sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's tires kick up thick, billowing clouds of beige dust that snake along its wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, arid landscape with a long, straight dirt road and a distant mountain range with purple peaks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020584525540471077 + ], + "opposite_logprob_tokens": [ + -3.8955845832824707 + ], + "logprob_reward": -0.020584525540471077, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020584525540471077, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08256042748689651, + 0.10460181534290314, + 0.17049740254878998, + 0.09302185475826263, + 0.11535695195198059, + 0.02947954833507538, + 0.26874840259552, + 0.058151956647634506, + 0.06414082646369934, + 0.013440795242786407 + ], + "credit_map": [ + 0.012596913613379002, + 0.010331188328564167, + 0.009247638285160065, + 0.01050163246691227, + 0.009502769447863102, + 0.019064178690314293, + 0.005443116649985313, + 0.005872989539057016, + 0.014284438453614712, + 0.013970548287034035, + 0.017818761989474297, + 0.029274601489305496, + 0.007419170346111059, + 0.013093662448227406, + 0.004999925848096609, + 0.003740714630112052, + 0.040440578013658524, + 0.014504225924611092, + 0.03778857737779617, + 0.036462053656578064, + 0.010891714133322239, + 0.016977205872535706, + 0.0067984736524522305, + 0.0066345748491585255, + 0.016094118356704712, + 0.016574915498495102, + 0.017945971339941025, + 0.013385391794145107, + 0.007650497369468212, + 0.012308488599956036, + 0.0041631232015788555, + 0.004899353254586458, + 0.01559994462877512, + 0.018045304343104362, + 0.012168331071734428, + 0.016168206930160522, + 0.017854953184723854, + 0.020117636770009995, + 0.007728693075478077, + 0.007673884276300669, + 0.003927936777472496, + 0.0025928483810275793, + 0.008408132009208202, + 0.004425807856023312, + 0.003063483629375696, + 0.0032500596717000008, + 0.0020891267340630293, + 0.0017221547896042466, + 0.016893496736884117, + 0.014446481131017208, + 0.04553825408220291, + 0.03190045431256294, + 0.02106163464486599, + 0.11497607827186584, + 0.009716070257127285, + 0.014215906150639057, + 0.005809864029288292, + 0.00904091913253069, + 0.005972872022539377, + 0.013390963897109032, + 0.0054853809997439384, + 0.011343799531459808, + 0.004785498604178429, + 0.0023226579651236534, + 0.007772247307002544, + 0.0033500215504318476, + 0.002678664866834879, + 0.012080267071723938, + 0.005144751630723476, + 0.028243035078048706, + 0.002538414904847741, + 0.002333429642021656, + 0.002317240461707115, + 0.0013710857601836324, + 0.0013728166231885552, + 0.0014239216689020395, + 0.001421405584551394, + 0.003739413106814027, + 0.0009367400780320168, + 0.0008581719594076276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 4 + ], + "selected_reward": -0.020584525540471077 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vintage, faded cherry-red muscle car driving steadily forward down the center of the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's metal body coated in a thin layer of light-colored dust that dulls its sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's tires kick up thick, billowing clouds of beige dust that snake along its wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, arid landscape with a long, straight dirt road and a distant mountain range with purple peaks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006730742286890745 + ], + "opposite_logprob_tokens": [ + -5.006730556488037 + ], + "logprob_reward": -0.006730742286890745, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006730742286890745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1097048968076706, + 0.06656713038682938, + 0.12132909893989563, + 0.05486738309264183, + 0.10296016186475754, + 0.34628981351852417, + 0.07896865159273148, + 0.03388652950525284, + 0.0636342316865921, + 0.021792132407426834 + ], + "credit_map": [ + 0.019337119534611702, + 0.020047636702656746, + 0.011522061191499233, + 0.009739959612488747, + 0.008238168433308601, + 0.021051065996289253, + 0.00889295618981123, + 0.01087592076510191, + 0.01393598597496748, + 0.006586740724742413, + 0.006124134175479412, + 0.006943999789655209, + 0.005553541239351034, + 0.007719271816313267, + 0.00422851974144578, + 0.015474931336939335, + 0.02793114446103573, + 0.013964680023491383, + 0.030847005546092987, + 0.014613887295126915, + 0.005681217182427645, + 0.010696757584810257, + 0.005663850344717503, + 0.011930548585951328, + 0.010034935548901558, + 0.009038622491061687, + 0.011320123448967934, + 0.0066760205663740635, + 0.004666789434850216, + 0.007221289910376072, + 0.002816558349877596, + 0.003093040082603693, + 0.019452296197414398, + 0.013130231760442257, + 0.02437111735343933, + 0.01228558924049139, + 0.008326584473252296, + 0.015315237455070019, + 0.004821110982447863, + 0.0052579776383936405, + 0.07892995327711105, + 0.02401287667453289, + 0.100381039083004, + 0.05112820863723755, + 0.022309431806206703, + 0.03389139100909233, + 0.01770918071269989, + 0.01792769879102707, + 0.0051796589978039265, + 0.0038300668820738792, + 0.011271795257925987, + 0.010640155524015427, + 0.006347193848341703, + 0.03485549986362457, + 0.0026785710360854864, + 0.004165705293416977, + 0.005480824504047632, + 0.0040701390244066715, + 0.007086757104843855, + 0.003184510627761483, + 0.0033217165619134903, + 0.005865191575139761, + 0.0023829203564673662, + 0.0024944639299064875, + 0.011173028498888016, + 0.00927337072789669, + 0.0137649429962039, + 0.0068551236763596535, + 0.008075377903878689, + 0.007484577130526304, + 0.004110096488147974, + 0.0028977140318602324, + 0.002304813824594021, + 0.0010725089814513922, + 0.0013542614178732038, + 0.0024651545099914074, + 0.003468813607469201, + 0.008462107740342617, + 0.0015519233420491219, + 0.0011125477030873299 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.006730742286890745 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file